From 25faefc062de2e7e967e18ca886b0c5f27c1e50b Mon Sep 17 00:00:00 2001
From: Denys Moskalenko <denysmoskalenko11@gmail.com>
Date: Sat, 15 Nov 2025 16:30:40 +0100
Subject: [PATCH 1/3]  Bedrock: add prompt caching support and verification

  - Emit cache-point tool entries so Bedrock accepts cached tool definitions
  - Document and test prompt caching (writes + reads) with cassette-body checks
  - Refresh Bedrock cassettes and type annotations to align with the new flow
---
 docs/models/bedrock.md                        |  73 +++++++
 pydantic_ai_slim/pydantic_ai/messages.py      |   1 +
 .../pydantic_ai/models/bedrock.py             |  62 +++++-
 ...edrock_cache_point_adds_cache_control.yaml | 103 ++++++++++
 .../test_bedrock_cache_write_and_read.yaml    | 148 ++++++++++++++
 ...st_bedrock_model_usage_limit_exceeded.yaml |   8 +-
 tests/models/test_bedrock.py                  | 183 +++++++++++++++---
 7 files changed, 543 insertions(+), 35 deletions(-)
 create mode 100644 tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml
 create mode 100644 tests/models/cassettes/test_bedrock/test_bedrock_cache_write_and_read.yaml

diff --git a/docs/models/bedrock.md b/docs/models/bedrock.md
index 55cb47bf15..667ca487d4 100644
--- a/docs/models/bedrock.md
+++ b/docs/models/bedrock.md
@@ -74,6 +74,79 @@ model = BedrockConverseModel(model_name='us.amazon.nova-pro-v1:0')
 agent = Agent(model=model, model_settings=bedrock_model_settings)
 ```
 
+## Prompt Caching
+
+Bedrock supports [prompt caching](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html) on Anthropic models so you can reuse expensive context across requests. Pydantic AI exposes the same three strategies as Anthropic:
+
+1. **Cache User Messages with [`CachePoint`][pydantic_ai.messages.CachePoint]**: Insert a `CachePoint` marker to cache everything before it in the current user message.
+2. **Cache System Instructions**: Enable [`BedrockModelSettings.bedrock_cache_instructions`][pydantic_ai.models.bedrock.BedrockModelSettings.bedrock_cache_instructions] to append a cache point after the system prompt.
+3. **Cache Tool Definitions**: Enable [`BedrockModelSettings.bedrock_cache_tool_definitions`][pydantic_ai.models.bedrock.BedrockModelSettings.bedrock_cache_tool_definitions] to cache your tool schemas.
+
+> [!NOTE]
+> AWS only serves cached content once a segment crosses the provider-specific minimum token thresholds (see the [Bedrock prompt caching docs](https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html)). Short prompts or tool definitions below those limits will bypass the cache, so don't expect savings for tiny payloads.
+
+You can combine all of them:
+
+```python
+from pydantic_ai import Agent, CachePoint, RunContext
+from pydantic_ai.models.bedrock import BedrockConverseModel, BedrockModelSettings
+
+model = BedrockConverseModel('us.anthropic.claude-sonnet-4-5-20250929-v1:0')
+agent = Agent(
+    model,
+    system_prompt='Detailed instructions...',
+    model_settings=BedrockModelSettings(
+        bedrock_cache_instructions=True,
+        bedrock_cache_tool_definitions=True,
+    ),
+)
+
+
+@agent.tool
+async def search_docs(ctx: RunContext, query: str) -> str:
+    return f'Results for {query}'
+
+
+async def main():
+    result1 = await agent.run(
+        [
+            'Long cached context...',
+            CachePoint(),
+            'First question',
+        ]
+    )
+    result2 = await agent.run(
+        [
+            'Long cached context...',
+            CachePoint(),
+            'Second question',
+        ]
+    )
+    print(result1.output, result1.usage())
+    print(result2.output, result2.usage())
+```
+
+Access cache usage statistics via [`RequestUsage`][pydantic_ai.usage.RequestUsage]:
+
+```python
+from pydantic_ai import Agent, CachePoint
+
+agent = Agent('bedrock:us.anthropic.claude-sonnet-4-5-20250929-v1:0')
+
+
+async def main():
+    result = await agent.run(
+        [
+            'Reference material...',
+            CachePoint(),
+            'What changed since last time?',
+        ]
+    )
+    usage = result.usage()
+    print(f'Cache writes: {usage.cache_write_tokens}')
+    print(f'Cache reads: {usage.cache_read_tokens}')
+```
+
 ## `provider` argument
 
 You can provide a custom `BedrockProvider` via the `provider` argument. This is useful when you want to specify credentials directly or use a custom boto3 client:
diff --git a/pydantic_ai_slim/pydantic_ai/messages.py b/pydantic_ai_slim/pydantic_ai/messages.py
index 1f3b5cd6e5..3d37b32091 100644
--- a/pydantic_ai_slim/pydantic_ai/messages.py
+++ b/pydantic_ai_slim/pydantic_ai/messages.py
@@ -622,6 +622,7 @@ class CachePoint:
     Supported by:
 
     - Anthropic
+    - Amazon Bedrock (Converse API)
     """
 
     kind: Literal['cache-point'] = 'cache-point'
diff --git a/pydantic_ai_slim/pydantic_ai/models/bedrock.py b/pydantic_ai_slim/pydantic_ai/models/bedrock.py
index acb98e5ec0..b2fcecaffd 100644
--- a/pydantic_ai_slim/pydantic_ai/models/bedrock.py
+++ b/pydantic_ai_slim/pydantic_ai/models/bedrock.py
@@ -208,6 +208,21 @@ class BedrockModelSettings(ModelSettings, total=False):
     See more about it on <https://docs.aws.amazon.com/bedrock/latest/userguide/model-parameters.html>.
     """
 
+    bedrock_cache_tool_definitions: bool
+    """Whether to add a cache point after the last tool definition.
+
+    When enabled, the last tool in the `tools` array will include a `cachePoint`, allowing Bedrock to cache tool
+    definitions and reduce costs for compatible models.
+    See https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html for more information.
+    """
+
+    bedrock_cache_instructions: bool
+    """Whether to add a cache point after the system prompt blocks.
+
+    When enabled, an extra `cachePoint` is appended to the system prompt so Bedrock can cache system instructions.
+    See https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html for more information.
+    """
+
 
 @dataclass(init=False)
 class BedrockConverseModel(Model):
@@ -299,7 +314,8 @@ async def count_tokens(
         Check the actual supported models on <https://docs.aws.amazon.com/bedrock/latest/userguide/count-tokens.html>
         """
         model_settings, model_request_parameters = self.prepare_request(model_settings, model_request_parameters)
-        system_prompt, bedrock_messages = await self._map_messages(messages, model_request_parameters)
+        settings = cast(BedrockModelSettings, model_settings or {})
+        system_prompt, bedrock_messages = await self._map_messages(messages, model_request_parameters, settings)
         params: CountTokensRequestTypeDef = {
             'modelId': self._remove_inference_geo_prefix(self.model_name),
             'input': {
@@ -374,6 +390,8 @@ async def _process_response(self, response: ConverseResponseTypeDef) -> ModelRes
         u = usage.RequestUsage(
             input_tokens=response['usage']['inputTokens'],
             output_tokens=response['usage']['outputTokens'],
+            cache_read_tokens=response['usage'].get('cacheReadInputTokens', 0),
+            cache_write_tokens=response['usage'].get('cacheWriteInputTokens', 0),
         )
         response_id = response.get('ResponseMetadata', {}).get('RequestId', None)
         raw_finish_reason = response['stopReason']
@@ -417,8 +435,9 @@ async def _messages_create(
         model_settings: BedrockModelSettings | None,
         model_request_parameters: ModelRequestParameters,
     ) -> ConverseResponseTypeDef | ConverseStreamResponseTypeDef:
-        system_prompt, bedrock_messages = await self._map_messages(messages, model_request_parameters)
-        inference_config = self._map_inference_config(model_settings)
+        settings = model_settings or BedrockModelSettings()
+        system_prompt, bedrock_messages = await self._map_messages(messages, model_request_parameters, settings)
+        inference_config = self._map_inference_config(settings)
 
         params: ConverseRequestTypeDef = {
             'modelId': self.model_name,
@@ -427,7 +446,7 @@ async def _messages_create(
             'inferenceConfig': inference_config,
         }
 
-        tool_config = self._map_tool_config(model_request_parameters)
+        tool_config = self._map_tool_config(model_request_parameters, settings)
         if tool_config:
             params['toolConfig'] = tool_config
 
@@ -481,11 +500,18 @@ def _map_inference_config(
 
         return inference_config
 
-    def _map_tool_config(self, model_request_parameters: ModelRequestParameters) -> ToolConfigurationTypeDef | None:
+    def _map_tool_config(
+        self,
+        model_request_parameters: ModelRequestParameters,
+        model_settings: BedrockModelSettings | None,
+    ) -> ToolConfigurationTypeDef | None:
         tools = self._get_tools(model_request_parameters)
         if not tools:
             return None
 
+        if model_settings and model_settings.get('bedrock_cache_tool_definitions'):
+            tools.append({'cachePoint': {'type': 'default'}})
+
         tool_choice: ToolChoiceTypeDef
         if not model_request_parameters.allow_text_output:
             tool_choice = {'any': {}}
@@ -499,12 +525,16 @@ def _map_tool_config(self, model_request_parameters: ModelRequestParameters) ->
         return tool_config
 
     async def _map_messages(  # noqa: C901
-        self, messages: list[ModelMessage], model_request_parameters: ModelRequestParameters
+        self,
+        messages: list[ModelMessage],
+        model_request_parameters: ModelRequestParameters,
+        model_settings: BedrockModelSettings | None,
     ) -> tuple[list[SystemContentBlockTypeDef], list[MessageUnionTypeDef]]:
         """Maps a `pydantic_ai.Message` to the Bedrock `MessageUnionTypeDef`.
 
         Groups consecutive ToolReturnPart objects into a single user message as required by Bedrock Claude/Nova models.
         """
+        settings = model_settings or BedrockModelSettings()
         profile = BedrockModelProfile.from_profile(self.profile)
         system_prompt: list[SystemContentBlockTypeDef] = []
         bedrock_messages: list[MessageUnionTypeDef] = []
@@ -613,10 +643,13 @@ async def _map_messages(  # noqa: C901
         if instructions := self._get_instructions(messages, model_request_parameters):
             system_prompt.insert(0, {'text': instructions})
 
+        if system_prompt and settings.get('bedrock_cache_instructions'):
+            system_prompt.append({'cachePoint': {'type': 'default'}})
+
         return system_prompt, processed_messages
 
     @staticmethod
-    async def _map_user_prompt(part: UserPromptPart, document_count: Iterator[int]) -> list[MessageUnionTypeDef]:
+    async def _map_user_prompt(part: UserPromptPart, document_count: Iterator[int]) -> list[MessageUnionTypeDef]:  # noqa: C901
         content: list[ContentBlockUnionTypeDef] = []
         if isinstance(part.content, str):
             content.append({'text': part.content})
@@ -674,8 +707,17 @@ async def _map_user_prompt(part: UserPromptPart, document_count: Iterator[int])
                 elif isinstance(item, AudioUrl):  # pragma: no cover
                     raise NotImplementedError('Audio is not supported yet.')
                 elif isinstance(item, CachePoint):
-                    # Bedrock support has not been implemented yet: https://github.com/pydantic/pydantic-ai/issues/3418
-                    pass
+                    if not content or 'cachePoint' in content[-1]:
+                        raise UserError(
+                            'CachePoint cannot be the first content in a user message - there must be previous content to cache when using Bedrock. '
+                            'To cache system instructions or tool definitions, use the `bedrock_cache_instructions` or `bedrock_cache_tool_definitions` settings instead.'
+                        )
+                    if 'text' not in content[-1]:
+                        # AWS currently rejects cache points that directly follow non-text content.
+                        # Insert an empty text block as a workaround (see https://github.com/pydantic/pydantic-ai/issues/3418
+                        # and https://github.com/pydantic/pydantic-ai/pull/2560#discussion_r2349209916).
+                        content.append({'text': '\n'})
+                    content.append({'cachePoint': {'type': 'default'}})
                 else:
                     assert_never(item)
         return [{'role': 'user', 'content': content}]
@@ -796,6 +838,8 @@ def _map_usage(self, metadata: ConverseStreamMetadataEventTypeDef) -> usage.Requ
         return usage.RequestUsage(
             input_tokens=metadata['usage']['inputTokens'],
             output_tokens=metadata['usage']['outputTokens'],
+            cache_read_tokens=metadata['usage'].get('cacheReadInputTokens', 0),
+            cache_write_tokens=metadata['usage'].get('cacheWriteInputTokens', 0),
         )
 
 
diff --git a/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml b/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml
new file mode 100644
index 0000000000..4bd60afe27
--- /dev/null
+++ b/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml
@@ -0,0 +1,103 @@
+interactions:
+- request:
+    body: '{"messages": [{"role": "user", "content": [{"text": "long promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
+      prompt"}, {"cachePoint": {"type": "default"}}, {"text": "Response only number What is 2 + 3"}]}], "system": [], "inferenceConfig":
+      {}}'
+    headers:
+      amz-sdk-invocation-id:
+      - !!binary |
+        MWQ3YjUzZDItNTI1NS00NDJhLWE5ZjAtZDM0YTMzMzcxOTI5
+      amz-sdk-request:
+      - !!binary |
+        YXR0ZW1wdD0x
+      content-length:
+      - '6781'
+      content-type:
+      - !!binary |
+        YXBwbGljYXRpb24vanNvbg==
+    method: POST
+    uri: https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-sonnet-4-5-20250929-v1%3A0/converse
+  response:
+    headers:
+      connection:
+      - keep-alive
+      content-length:
+      - '321'
+      content-type:
+      - application/json
+    parsed_body:
+      metrics:
+        latencyMs: 1867
+      output:
+        message:
+          content:
+          - text: '5'
+          role: assistant
+      stopReason: end_turn
+      usage:
+        cacheReadInputTokenCount: 0
+        cacheReadInputTokens: 0
+        cacheWriteInputTokenCount: 1203
+        cacheWriteInputTokens: 1203
+        inputTokens: 15
+        outputTokens: 5
+        serverToolUsage: {}
+        totalTokens: 1223
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/tests/models/cassettes/test_bedrock/test_bedrock_cache_write_and_read.yaml b/tests/models/cassettes/test_bedrock/test_bedrock_cache_write_and_read.yaml
new file mode 100644
index 0000000000..c1120f0e55
--- /dev/null
+++ b/tests/models/cassettes/test_bedrock/test_bedrock_cache_write_and_read.yaml
@@ -0,0 +1,148 @@
+interactions:
+- request:
+    body: '{"messages": [{"role": "user", "content": [{"text": "Newer response with something except single number\nNewer
+      response with something except single number\nNewer response with something except single number\nNewer response with
+      something except single number\nNewer response with something except single number\nNewer response with something except
+      single number\nNewer response with something except single number\nNewer response with something except single number\nNewer
+      response with something except single number\nNewer response with something except single number\n"}, {"cachePoint":
+      {"type": "default"}}, {"document": {"name": "Document 1", "format": "txt", "source": {"bytes": "WW91IGFyZSBhIGdyZWF0IG1hdGhlbWF0aWNpYW4="}}},
+      {"text": "\n"}, {"cachePoint": {"type": "default"}}, {"text": "What is 10 + 11?"}]}], "system": [{"text": "YOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\n"}, {"cachePoint": {"type":
+      "default"}}], "inferenceConfig": {}, "toolConfig": {"tools": [{"toolSpec": {"name": "catalog_lookup", "inputSchema":
+      {"json": {"additionalProperties": false, "properties": {}, "type": "object"}}}}, {"toolSpec": {"name": "diagnostics",
+      "inputSchema": {"json": {"additionalProperties": false, "properties": {}, "type": "object"}}}}, {"cachePoint": {"type":
+      "default"}}], "toolChoice": {"auto": {}}}}'
+    headers:
+      amz-sdk-invocation-id:
+      - !!binary |
+        MTNhZTU1NTctNmMwNC00Y2Q3LTg5YzQtZGJiMmQ4OWQ2NzBl
+      amz-sdk-request:
+      - !!binary |
+        YXR0ZW1wdD0x
+      content-length:
+      - '3416'
+      content-type:
+      - !!binary |
+        YXBwbGljYXRpb24vanNvbg==
+    method: POST
+    uri: https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-sonnet-4-5-20250929-v1%3A0/converse
+  response:
+    headers:
+      connection:
+      - keep-alive
+      content-length:
+      - '322'
+      content-type:
+      - application/json
+    parsed_body:
+      metrics:
+        latencyMs: 2053
+      output:
+        message:
+          content:
+          - text: '21'
+          role: assistant
+      stopReason: end_turn
+      usage:
+        cacheReadInputTokenCount: 0
+        cacheReadInputTokens: 0
+        cacheWriteInputTokenCount: 1313
+        cacheWriteInputTokens: 1313
+        inputTokens: 11
+        outputTokens: 5
+        serverToolUsage: {}
+        totalTokens: 1329
+    status:
+      code: 200
+      message: OK
+- request:
+    body: '{"messages": [{"role": "user", "content": [{"text": "Newer response with something except single number\nNewer
+      response with something except single number\nNewer response with something except single number\nNewer response with
+      something except single number\nNewer response with something except single number\nNewer response with something except
+      single number\nNewer response with something except single number\nNewer response with something except single number\nNewer
+      response with something except single number\nNewer response with something except single number\n"}, {"cachePoint":
+      {"type": "default"}}, {"document": {"name": "Document 1", "format": "txt", "source": {"bytes": "WW91IGFyZSBhIGdyZWF0IG1hdGhlbWF0aWNpYW4="}}},
+      {"text": "\n"}, {"cachePoint": {"type": "default"}}, {"text": "What is 10 + 11?"}]}], "system": [{"text": "YOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\n"}, {"cachePoint": {"type":
+      "default"}}], "inferenceConfig": {}, "toolConfig": {"tools": [{"toolSpec": {"name": "catalog_lookup", "inputSchema":
+      {"json": {"additionalProperties": false, "properties": {}, "type": "object"}}}}, {"toolSpec": {"name": "diagnostics",
+      "inputSchema": {"json": {"additionalProperties": false, "properties": {}, "type": "object"}}}}, {"cachePoint": {"type":
+      "default"}}], "toolChoice": {"auto": {}}}}'
+    headers:
+      amz-sdk-invocation-id:
+      - !!binary |
+        MTVjNDBlYzYtZjE1OC00MWRkLWJjNzItYWJkZjExZWVmZTBk
+      amz-sdk-request:
+      - !!binary |
+        YXR0ZW1wdD0x
+      content-length:
+      - '3416'
+      content-type:
+      - !!binary |
+        YXBwbGljYXRpb24vanNvbg==
+    method: POST
+    uri: https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-sonnet-4-5-20250929-v1%3A0/converse
+  response:
+    headers:
+      connection:
+      - keep-alive
+      content-length:
+      - '322'
+      content-type:
+      - application/json
+    parsed_body:
+      metrics:
+        latencyMs: 2030
+      output:
+        message:
+          content:
+          - text: '21'
+          role: assistant
+      stopReason: end_turn
+      usage:
+        cacheReadInputTokenCount: 1313
+        cacheReadInputTokens: 1313
+        cacheWriteInputTokenCount: 0
+        cacheWriteInputTokens: 0
+        inputTokens: 11
+        outputTokens: 5
+        serverToolUsage: {}
+        totalTokens: 1329
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/tests/models/cassettes/test_bedrock/test_bedrock_model_usage_limit_exceeded.yaml b/tests/models/cassettes/test_bedrock/test_bedrock_model_usage_limit_exceeded.yaml
index 92cefc9519..09e42ffceb 100644
--- a/tests/models/cassettes/test_bedrock/test_bedrock_model_usage_limit_exceeded.yaml
+++ b/tests/models/cassettes/test_bedrock/test_bedrock_model_usage_limit_exceeded.yaml
@@ -1,16 +1,16 @@
 interactions:
 - request:
     body: '{"input": {"converse": {"messages": [{"role": "user", "content": [{"text": "The quick brown fox jumps over the
-      lazydog."}]}], "system": []}}}'
+      lazydog."}, {"cachePoint": {"type": "default"}}, {"text": "What was next?"}]}], "system": []}}}'
     headers:
       amz-sdk-invocation-id:
       - !!binary |
-        ZDYxNmVkOTktYzgwMi00MDE0LTljZGUtYWFjMjk5N2I2MDFj
+        Y2FhYzlhNjAtNWYyOC00NjhmLWJkZTgtODM0OTdhMjc1ZTI3
       amz-sdk-request:
       - !!binary |
         YXR0ZW1wdD0x
       content-length:
-      - '141'
+      - '206'
       content-type:
       - !!binary |
         YXBwbGljYXRpb24vanNvbg==
@@ -25,7 +25,7 @@ interactions:
       content-type:
       - application/json
     parsed_body:
-      inputTokens: 19
+      inputTokens: 23
     status:
       code: 200
       message: OK
diff --git a/tests/models/test_bedrock.py b/tests/models/test_bedrock.py
index cce18a9227..e36852e9e0 100644
--- a/tests/models/test_bedrock.py
+++ b/tests/models/test_bedrock.py
@@ -15,6 +15,7 @@
     FunctionToolCallEvent,
     FunctionToolResultEvent,
     ImageUrl,
+    ModelMessage,
     ModelRequest,
     ModelResponse,
     PartDeltaEvent,
@@ -33,13 +34,9 @@
     VideoUrl,
 )
 from pydantic_ai.agent import Agent
-from pydantic_ai.exceptions import ModelHTTPError, ModelRetry, UsageLimitExceeded
+from pydantic_ai.exceptions import ModelHTTPError, ModelRetry, UsageLimitExceeded, UserError
 from pydantic_ai.messages import AgentStreamEvent
 from pydantic_ai.models import ModelRequestParameters
-from pydantic_ai.models.bedrock import BedrockConverseModel, BedrockModelSettings
-from pydantic_ai.models.openai import OpenAIResponsesModel, OpenAIResponsesModelSettings
-from pydantic_ai.providers.bedrock import BedrockProvider
-from pydantic_ai.providers.openai import OpenAIProvider
 from pydantic_ai.run import AgentRunResult, AgentRunResultEvent
 from pydantic_ai.tools import ToolDefinition
 from pydantic_ai.usage import RequestUsage, RunUsage, UsageLimits
@@ -47,7 +44,10 @@
 from ..conftest import IsDatetime, IsInstance, IsStr, try_import
 
 with try_import() as imports_successful:
-    pass
+    from pydantic_ai.models.bedrock import BedrockConverseModel, BedrockModelSettings
+    from pydantic_ai.models.openai import OpenAIResponsesModel, OpenAIResponsesModelSettings
+    from pydantic_ai.providers.bedrock import BedrockProvider
+    from pydantic_ai.providers.openai import OpenAIProvider
 
 pytestmark = [
     pytest.mark.skipif(not imports_successful(), reason='bedrock not installed'),
@@ -99,6 +99,7 @@ async def test_bedrock_model(allow_model_requests: None, bedrock_provider: Bedro
     )
 
 
+@pytest.mark.vcr()
 async def test_bedrock_model_usage_limit_exceeded(
     allow_model_requests: None,
     bedrock_provider: BedrockProvider,
@@ -108,14 +109,15 @@ async def test_bedrock_model_usage_limit_exceeded(
 
     with pytest.raises(
         UsageLimitExceeded,
-        match='The next request would exceed the input_tokens_limit of 18 \\(input_tokens=19\\)',
+        match='The next request would exceed the input_tokens_limit of 18 \\(input_tokens=23\\)',
     ):
         await agent.run(
-            'The quick brown fox jumps over the lazydog.',
+            ['The quick brown fox jumps over the lazydog.', CachePoint(), 'What was next?'],
             usage_limits=UsageLimits(input_tokens_limit=18, count_tokens_before_request=True),
         )
 
 
+@pytest.mark.vcr()
 async def test_bedrock_model_usage_limit_not_exceeded(
     allow_model_requests: None,
     bedrock_provider: BedrockProvider,
@@ -1224,7 +1226,7 @@ async def test_bedrock_group_consecutive_tool_return_parts(bedrock_provider: Bed
     ]
 
     # Call the mapping function directly
-    _, bedrock_messages = await model._map_messages(req, ModelRequestParameters())  # type: ignore[reportPrivateUsage]
+    _, bedrock_messages = await model._map_messages(req, ModelRequestParameters(), BedrockModelSettings())  # type: ignore[reportPrivateUsage]
 
     assert bedrock_messages == snapshot(
         [
@@ -1345,7 +1347,7 @@ async def test_bedrock_mistral_tool_result_format(bedrock_provider: BedrockProvi
     # Models other than Mistral support toolResult.content with text, not json
     model = BedrockConverseModel('us.amazon.nova-micro-v1:0', provider=bedrock_provider)
     # Call the mapping function directly
-    _, bedrock_messages = await model._map_messages(req, ModelRequestParameters())  # type: ignore[reportPrivateUsage]
+    _, bedrock_messages = await model._map_messages(req, ModelRequestParameters(), BedrockModelSettings())  # type: ignore[reportPrivateUsage]
 
     assert bedrock_messages == snapshot(
         [
@@ -1361,7 +1363,7 @@ async def test_bedrock_mistral_tool_result_format(bedrock_provider: BedrockProvi
     # Mistral requires toolResult.content to hold json, not text
     model = BedrockConverseModel('mistral.mistral-7b-instruct-v0:2', provider=bedrock_provider)
     # Call the mapping function directly
-    _, bedrock_messages = await model._map_messages(req, ModelRequestParameters())  # type: ignore[reportPrivateUsage]
+    _, bedrock_messages = await model._map_messages(req, ModelRequestParameters(), BedrockModelSettings())  # type: ignore[reportPrivateUsage]
 
     assert bedrock_messages == snapshot(
         [
@@ -1385,7 +1387,7 @@ async def test_bedrock_no_tool_choice(bedrock_provider: BedrockProvider):
 
     # Amazon Nova supports tool_choice
     model = BedrockConverseModel('us.amazon.nova-micro-v1:0', provider=bedrock_provider)
-    tool_config = model._map_tool_config(mrp)  # type: ignore[reportPrivateUsage]
+    tool_config = model._map_tool_config(mrp, BedrockModelSettings())  # type: ignore[reportPrivateUsage]
 
     assert tool_config == snapshot(
         {
@@ -1406,7 +1408,7 @@ async def test_bedrock_no_tool_choice(bedrock_provider: BedrockProvider):
 
     # Anthropic supports tool_choice
     model = BedrockConverseModel('us.anthropic.claude-3-7-sonnet-20250219-v1:0', provider=bedrock_provider)
-    tool_config = model._map_tool_config(mrp)  # type: ignore[reportPrivateUsage]
+    tool_config = model._map_tool_config(mrp, BedrockModelSettings())  # type: ignore[reportPrivateUsage]
 
     assert tool_config == snapshot(
         {
@@ -1427,7 +1429,7 @@ async def test_bedrock_no_tool_choice(bedrock_provider: BedrockProvider):
 
     # Other models don't support tool_choice
     model = BedrockConverseModel('us.meta.llama4-maverick-17b-instruct-v1:0', provider=bedrock_provider)
-    tool_config = model._map_tool_config(mrp)  # type: ignore[reportPrivateUsage]
+    tool_config = model._map_tool_config(mrp, BedrockModelSettings())  # type: ignore[reportPrivateUsage]
 
     assert tool_config == snapshot(
         {
@@ -1515,12 +1517,149 @@ async def test_bedrock_streaming_error(allow_model_requests: None, bedrock_provi
     assert exc_info.value.body.get('Error', {}).get('Message') == 'The provided model identifier is invalid.'  # type: ignore[union-attr]
 
 
-async def test_cache_point_filtering():
-    """Test that CachePoint is filtered out in Bedrock message mapping."""
-    from itertools import count
+@pytest.mark.vcr()
+async def test_bedrock_cache_point_adds_cache_control(allow_model_requests: None, bedrock_provider: BedrockProvider):
+    """Record a real Bedrock call to confirm cache points reach AWS (requires ~1k tokens)."""
+    model = BedrockConverseModel('us.anthropic.claude-sonnet-4-5-20250929-v1:0', provider=bedrock_provider)
+    agent = Agent(model=model)
+    long_context = 'long prompt' * 600  # More tokens to activate a cache
+
+    result = await agent.run([long_context, CachePoint(), 'Response only number What is 2 + 3'])
+    assert result.output == snapshot('5')
+    assert result.usage() == snapshot(RunUsage(input_tokens=15, output_tokens=5, cache_write_tokens=1203, requests=1))
+
+
+@pytest.mark.vcr()
+async def test_bedrock_cache_write_and_read(allow_model_requests: None, bedrock_provider: BedrockProvider):
+    """Integration test covering tool and instruction caching using a recorded cassette."""
+    model = BedrockConverseModel('us.anthropic.claude-sonnet-4-5-20250929-v1:0', provider=bedrock_provider)
+    system_prompt = 'YOU MUST RESPONSE ONLY WITH SINGLE NUMBER\n' * 50  # More tokens to activate a cache
+    agent = Agent(
+        model,
+        system_prompt=system_prompt,
+        model_settings=BedrockModelSettings(
+            bedrock_cache_tool_definitions=True,
+            bedrock_cache_instructions=True,
+        ),
+    )
+
+    @agent.tool_plain
+    def catalog_lookup() -> str:  # pragma: no cover - exercised via agent call
+        return 'catalog-ok'
+
+    @agent.tool_plain
+    def diagnostics() -> str:  # pragma: no cover - exercised via agent call
+        return 'diagnostics-ok'
 
-    # Test the static method directly
-    messages = await BedrockConverseModel._map_user_prompt(UserPromptPart(content=['text', CachePoint()]), count())  # pyright: ignore[reportPrivateUsage]
-    # CachePoint should be filtered out, message should still be valid
-    assert len(messages) == 1
-    assert messages[0]['role'] == 'user'
+    long_context = 'Newer response with something except single number\n' * 10
+    document = BinaryContent(data=b'You are a great mathematician', media_type='text/plain')
+    run_args = [long_context, CachePoint(), document, CachePoint(), 'What is 10 + 11?']
+
+    first = await agent.run(run_args)
+    assert first.output == snapshot('21')
+    first_usage = first.usage()
+    assert first_usage == snapshot(RunUsage(input_tokens=11, output_tokens=5, cache_write_tokens=1313, requests=1))
+
+    second = await agent.run(run_args)
+    assert second.output == snapshot('21')
+    second_usage = second.usage()
+    assert second_usage == snapshot(RunUsage(input_tokens=11, output_tokens=5, cache_read_tokens=1313, requests=1))
+
+
+async def test_bedrock_cache_point_as_first_content_raises_error(
+    allow_model_requests: None, bedrock_provider: BedrockProvider
+):
+    """CachePoint should raise a UserError if it appears before any other content."""
+    model = BedrockConverseModel('anthropic.claude-3-7-sonnet-20250219-v1:0', provider=bedrock_provider)
+    messages: list[ModelMessage] = [ModelRequest(parts=[UserPromptPart(content=[CachePoint(), 'This should fail'])])]
+    with pytest.raises(UserError, match='CachePoint cannot be the first content in a user message'):
+        await model._map_messages(messages, ModelRequestParameters(), BedrockModelSettings())  # pyright: ignore[reportPrivateUsage]
+
+
+# Bedrock currently errors if a cache point immediately follows non-text content, so we inject a newline block.
+async def test_bedrock_cache_point_after_binary_content_workaround(
+    allow_model_requests: None, bedrock_provider: BedrockProvider
+):
+    model = BedrockConverseModel('us.anthropic.claude-3-5-sonnet-20240620-v1:0', provider=bedrock_provider)
+    messages: list[ModelMessage] = [
+        ModelRequest(
+            parts=[
+                UserPromptPart(
+                    content=[
+                        'Process the attached text file. Return the answer only.',
+                        BinaryContent(data=b'What is 2+2? Provide the answer only.', media_type='text/plain'),
+                        CachePoint(),
+                    ]
+                )
+            ]
+        )
+    ]
+    _, bedrock_messages = await model._map_messages(messages, ModelRequestParameters(), BedrockModelSettings())  # pyright: ignore[reportPrivateUsage]
+    assert bedrock_messages[0]['content'] == snapshot(
+        [
+            {'text': 'Process the attached text file. Return the answer only.'},
+            {
+                'document': {
+                    'name': 'Document 1',
+                    'format': 'txt',
+                    'source': {'bytes': b'What is 2+2? Provide the answer only.'},
+                }
+            },
+            {'text': '\n'},  # Empty line after BinaryContent as temp workaround unless bedrock will fix the bug
+            {'cachePoint': {'type': 'default'}},
+        ]
+    )
+
+
+async def test_bedrock_cache_point_multiple_markers(allow_model_requests: None, bedrock_provider: BedrockProvider):
+    model = BedrockConverseModel('us.anthropic.claude-3-5-haiku-20241022-v1:0', provider=bedrock_provider)
+    messages: list[ModelMessage] = [
+        ModelRequest(
+            parts=[UserPromptPart(content=['First chunk', CachePoint(), 'Second chunk', CachePoint(), 'Question'])]
+        )
+    ]
+    _, bedrock_messages = await model._map_messages(messages, ModelRequestParameters(), BedrockModelSettings())  # pyright: ignore[reportPrivateUsage]
+    assert bedrock_messages[0]['content'] == snapshot(
+        [
+            {'text': 'First chunk'},
+            {'cachePoint': {'type': 'default'}},
+            {'text': 'Second chunk'},
+            {'cachePoint': {'type': 'default'}},
+            {'text': 'Question'},
+        ]
+    )
+
+
+async def test_bedrock_cache_tool_definitions(allow_model_requests: None, bedrock_provider: BedrockProvider):
+    model = BedrockConverseModel('anthropic.claude-3-5-sonnet-20241022-v2:0', provider=bedrock_provider)
+    params = ModelRequestParameters(
+        function_tools=[
+            ToolDefinition(name='tool_one'),
+            ToolDefinition(name='tool_two'),
+        ]
+    )
+    params = model.customize_request_parameters(params)
+    tool_config = model._map_tool_config(  # pyright: ignore[reportPrivateUsage]
+        params,
+        BedrockModelSettings(bedrock_cache_tool_definitions=True),
+    )
+    assert tool_config and len(tool_config['tools']) == 3
+    assert tool_config['tools'][-1] == {'cachePoint': {'type': 'default'}}
+
+
+async def test_bedrock_cache_instructions(allow_model_requests: None, bedrock_provider: BedrockProvider):
+    model = BedrockConverseModel('us.anthropic.claude-3-5-sonnet-20240620-v1:0', provider=bedrock_provider)
+    messages: list[ModelMessage] = [
+        ModelRequest(parts=[SystemPromptPart(content='System instructions to cache.'), UserPromptPart(content='Hi!')])
+    ]
+    system_prompt, _ = await model._map_messages(  # pyright: ignore[reportPrivateUsage]
+        messages,
+        ModelRequestParameters(),
+        BedrockModelSettings(bedrock_cache_instructions=True),
+    )
+    assert system_prompt == snapshot(
+        [
+            {'text': 'System instructions to cache.'},
+            {'cachePoint': {'type': 'default'}},
+        ]
+    )

From 59c6868225b3d36045f6cc7c540b8a5a221240e8 Mon Sep 17 00:00:00 2001
From: Denys Moskalenko <denysmoskalenko11@gmail.com>
Date: Sat, 15 Nov 2025 17:12:05 +0100
Subject: [PATCH 2/3] Skip documentation examples from tests

---
 docs/models/bedrock.md | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/docs/models/bedrock.md b/docs/models/bedrock.md
index 667ca487d4..9cd42bcdc3 100644
--- a/docs/models/bedrock.md
+++ b/docs/models/bedrock.md
@@ -87,7 +87,7 @@ Bedrock supports [prompt caching](https://docs.aws.amazon.com/bedrock/latest/use
 
 You can combine all of them:
 
-```python
+```python {test="skip"}
 from pydantic_ai import Agent, CachePoint, RunContext
 from pydantic_ai.models.bedrock import BedrockConverseModel, BedrockModelSettings
 
@@ -128,7 +128,7 @@ async def main():
 
 Access cache usage statistics via [`RequestUsage`][pydantic_ai.usage.RequestUsage]:
 
-```python
+```python {test="skip"}
 from pydantic_ai import Agent, CachePoint
 
 agent = Agent('bedrock:us.anthropic.claude-sonnet-4-5-20250929-v1:0')

From 900d5423e6393f8d81f6721d65b0f199b24a815b Mon Sep 17 00:00:00 2001
From: Denys Moskalenko <denysmoskalenko11@gmail.com>
Date: Wed, 19 Nov 2025 11:59:07 +0100
Subject: [PATCH 3/3] Parametrize `test_bedrock_cache_point_adds_cache_control`
 to support multiple models

---
 ...edrock_cache_point_adds_cache_control.yaml | 103 ------------------
 ...adds_cache_control[claude-sonnet-4-5].yaml |  91 ++++++++++++++++
 ...e_point_adds_cache_control[nova-lite].yaml |  91 ++++++++++++++++
 tests/models/test_bedrock.py                  |  34 ++++--
 4 files changed, 206 insertions(+), 113 deletions(-)
 delete mode 100644 tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml
 create mode 100644 tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[claude-sonnet-4-5].yaml
 create mode 100644 tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[nova-lite].yaml

diff --git a/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml b/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml
deleted file mode 100644
index 4bd60afe27..0000000000
--- a/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control.yaml
+++ /dev/null
@@ -1,103 +0,0 @@
-interactions:
-- request:
-    body: '{"messages": [{"role": "user", "content": [{"text": "long promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong promptlong
-      prompt"}, {"cachePoint": {"type": "default"}}, {"text": "Response only number What is 2 + 3"}]}], "system": [], "inferenceConfig":
-      {}}'
-    headers:
-      amz-sdk-invocation-id:
-      - !!binary |
-        MWQ3YjUzZDItNTI1NS00NDJhLWE5ZjAtZDM0YTMzMzcxOTI5
-      amz-sdk-request:
-      - !!binary |
-        YXR0ZW1wdD0x
-      content-length:
-      - '6781'
-      content-type:
-      - !!binary |
-        YXBwbGljYXRpb24vanNvbg==
-    method: POST
-    uri: https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-sonnet-4-5-20250929-v1%3A0/converse
-  response:
-    headers:
-      connection:
-      - keep-alive
-      content-length:
-      - '321'
-      content-type:
-      - application/json
-    parsed_body:
-      metrics:
-        latencyMs: 1867
-      output:
-        message:
-          content:
-          - text: '5'
-          role: assistant
-      stopReason: end_turn
-      usage:
-        cacheReadInputTokenCount: 0
-        cacheReadInputTokens: 0
-        cacheWriteInputTokenCount: 1203
-        cacheWriteInputTokens: 1203
-        inputTokens: 15
-        outputTokens: 5
-        serverToolUsage: {}
-        totalTokens: 1223
-    status:
-      code: 200
-      message: OK
-version: 1
diff --git a/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[claude-sonnet-4-5].yaml b/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[claude-sonnet-4-5].yaml
new file mode 100644
index 0000000000..519eebbf3d
--- /dev/null
+++ b/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[claude-sonnet-4-5].yaml
@@ -0,0 +1,91 @@
+interactions:
+- request:
+    body: '{"messages": [{"role": "user", "content": [{"text": "ONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\n"}, {"cachePoint": {"type": "default"}}, {"text": "Response
+      only number What is 2 + 3"}]}], "system": [{"text": "YOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\n"}, {"cachePoint": {"type": "default"}}], "inferenceConfig": {}}'
+    headers:
+      amz-sdk-invocation-id:
+      - !!binary |
+        Y2RmYWJiOGYtYjM0MC00NzY4LTgwZTEtMDI5NzZiZDdiZjVm
+      amz-sdk-request:
+      - !!binary |
+        YXR0ZW1wdD0x
+      content-length:
+      - '5580'
+      content-type:
+      - !!binary |
+        YXBwbGljYXRpb24vanNvbg==
+    method: POST
+    uri: https://bedrock-runtime.us-east-1.amazonaws.com/model/us.anthropic.claude-sonnet-4-5-20250929-v1%3A0/converse
+  response:
+    headers:
+      connection:
+      - keep-alive
+      content-length:
+      - '321'
+      content-type:
+      - application/json
+    parsed_body:
+      metrics:
+        latencyMs: 2015
+      output:
+        message:
+          content:
+          - text: '5'
+          role: assistant
+      stopReason: end_turn
+      usage:
+        cacheReadInputTokenCount: 0
+        cacheReadInputTokens: 0
+        cacheWriteInputTokenCount: 1503
+        cacheWriteInputTokens: 1503
+        inputTokens: 14
+        outputTokens: 5
+        serverToolUsage: {}
+        totalTokens: 1522
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[nova-lite].yaml b/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[nova-lite].yaml
new file mode 100644
index 0000000000..8cfeda778e
--- /dev/null
+++ b/tests/models/cassettes/test_bedrock/test_bedrock_cache_point_adds_cache_control[nova-lite].yaml
@@ -0,0 +1,91 @@
+interactions:
+- request:
+    body: '{"messages": [{"role": "user", "content": [{"text": "ONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\nONLY
+      SINGLE NUMBER IN RESPONSE\nONLY SINGLE NUMBER IN RESPONSE\n"}, {"cachePoint": {"type": "default"}}, {"text": "Response
+      only number What is 2 + 3"}]}], "system": [{"text": "YOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU
+      MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE
+      NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY
+      WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE
+      ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST RESPONSE ONLY WITH SINGLE NUMBER\nYOU MUST
+      RESPONSE ONLY WITH SINGLE NUMBER\n"}, {"cachePoint": {"type": "default"}}], "inferenceConfig": {}}'
+    headers:
+      amz-sdk-invocation-id:
+      - !!binary |
+        MWRhOGY1NGItNThhMC00YTc3LTg0YjYtNzIyZGEwOGVmMjI0
+      amz-sdk-request:
+      - !!binary |
+        YXR0ZW1wdD0x
+      content-length:
+      - '5580'
+      content-type:
+      - !!binary |
+        YXBwbGljYXRpb24vanNvbg==
+    method: POST
+    uri: https://bedrock-runtime.us-east-1.amazonaws.com/model/us.amazon.nova-lite-v1%3A0/converse
+  response:
+    headers:
+      connection:
+      - keep-alive
+      content-length:
+      - '320'
+      content-type:
+      - application/json
+    parsed_body:
+      metrics:
+        latencyMs: 797
+      output:
+        message:
+          content:
+          - text: '5'
+          role: assistant
+      stopReason: end_turn
+      usage:
+        cacheReadInputTokenCount: 0
+        cacheReadInputTokens: 0
+        cacheWriteInputTokenCount: 1298
+        cacheWriteInputTokens: 1298
+        inputTokens: 10
+        outputTokens: 2
+        serverToolUsage: {}
+        totalTokens: 1310
+    status:
+      code: 200
+      message: OK
+version: 1
diff --git a/tests/models/test_bedrock.py b/tests/models/test_bedrock.py
index e36852e9e0..71f2ee6abc 100644
--- a/tests/models/test_bedrock.py
+++ b/tests/models/test_bedrock.py
@@ -44,7 +44,7 @@
 from ..conftest import IsDatetime, IsInstance, IsStr, try_import
 
 with try_import() as imports_successful:
-    from pydantic_ai.models.bedrock import BedrockConverseModel, BedrockModelSettings
+    from pydantic_ai.models.bedrock import BedrockConverseModel, BedrockModelName, BedrockModelSettings
     from pydantic_ai.models.openai import OpenAIResponsesModel, OpenAIResponsesModelSettings
     from pydantic_ai.providers.bedrock import BedrockProvider
     from pydantic_ai.providers.openai import OpenAIProvider
@@ -1518,25 +1518,39 @@ async def test_bedrock_streaming_error(allow_model_requests: None, bedrock_provi
 
 
 @pytest.mark.vcr()
-async def test_bedrock_cache_point_adds_cache_control(allow_model_requests: None, bedrock_provider: BedrockProvider):
+@pytest.mark.parametrize(
+    'model_name',
+    [
+        pytest.param('us.anthropic.claude-sonnet-4-5-20250929-v1:0', id='claude-sonnet-4-5'),
+        pytest.param('us.amazon.nova-lite-v1:0', id='nova-lite'),
+    ],
+)
+async def test_bedrock_cache_point_adds_cache_control(
+    allow_model_requests: None, bedrock_provider: BedrockProvider, model_name: BedrockModelName
+):
     """Record a real Bedrock call to confirm cache points reach AWS (requires ~1k tokens)."""
-    model = BedrockConverseModel('us.anthropic.claude-sonnet-4-5-20250929-v1:0', provider=bedrock_provider)
-    agent = Agent(model=model)
-    long_context = 'long prompt' * 600  # More tokens to activate a cache
+    model = BedrockConverseModel(model_name, provider=bedrock_provider)
+    agent = Agent(
+        model,
+        system_prompt='YOU MUST RESPONSE ONLY WITH SINGLE NUMBER\n' * 50,  # More tokens to activate a cache
+        model_settings=BedrockModelSettings(bedrock_cache_instructions=True),
+    )
+    long_context = 'ONLY SINGLE NUMBER IN RESPONSE\n' * 100  # More tokens to activate a cache
 
     result = await agent.run([long_context, CachePoint(), 'Response only number What is 2 + 3'])
     assert result.output == snapshot('5')
-    assert result.usage() == snapshot(RunUsage(input_tokens=15, output_tokens=5, cache_write_tokens=1203, requests=1))
+    # Different tokens usage depending on a model
+    assert result.usage().cache_write_tokens >= 1000
+    assert result.usage().input_tokens <= 20
 
 
 @pytest.mark.vcr()
 async def test_bedrock_cache_write_and_read(allow_model_requests: None, bedrock_provider: BedrockProvider):
     """Integration test covering tool and instruction caching using a recorded cassette."""
     model = BedrockConverseModel('us.anthropic.claude-sonnet-4-5-20250929-v1:0', provider=bedrock_provider)
-    system_prompt = 'YOU MUST RESPONSE ONLY WITH SINGLE NUMBER\n' * 50  # More tokens to activate a cache
     agent = Agent(
         model,
-        system_prompt=system_prompt,
+        system_prompt='YOU MUST RESPONSE ONLY WITH SINGLE NUMBER\n' * 50,  # More tokens to activate a cache
         model_settings=BedrockModelSettings(
             bedrock_cache_tool_definitions=True,
             bedrock_cache_instructions=True,
@@ -1580,7 +1594,7 @@ async def test_bedrock_cache_point_as_first_content_raises_error(
 async def test_bedrock_cache_point_after_binary_content_workaround(
     allow_model_requests: None, bedrock_provider: BedrockProvider
 ):
-    model = BedrockConverseModel('us.anthropic.claude-3-5-sonnet-20240620-v1:0', provider=bedrock_provider)
+    model = BedrockConverseModel('us.anthropic.claude-haiku-4-5-20251001-v1:0', provider=bedrock_provider)
     messages: list[ModelMessage] = [
         ModelRequest(
             parts=[
@@ -1631,7 +1645,7 @@ async def test_bedrock_cache_point_multiple_markers(allow_model_requests: None,
 
 
 async def test_bedrock_cache_tool_definitions(allow_model_requests: None, bedrock_provider: BedrockProvider):
-    model = BedrockConverseModel('anthropic.claude-3-5-sonnet-20241022-v2:0', provider=bedrock_provider)
+    model = BedrockConverseModel('anthropic.claude-sonnet-4-20250514-v1:0', provider=bedrock_provider)
     params = ModelRequestParameters(
         function_tools=[
             ToolDefinition(name='tool_one'),