* fix(anthropic): use dynamic max_tokens based on model
When users don't specify max_tokens in requests to Anthropic models,
LiteLLM now uses the correct max_output_tokens value from the model
pricing JSON instead of a hardcoded 4096.
This fixes truncated responses for Claude 3.5+ models which support
higher output limits (8192 for Claude 3.5, 128k for Claude 3.7, etc.)
Fixes#8835
* fix(anthropic): restore env var support for backwards compatibility
Keep DEFAULT_ANTHROPIC_CHAT_MAX_TOKENS as fallback when model is not
found in JSON, allowing users to configure via environment variable.
Add direct Stability AI REST API support for image generation endpoints.
This enables using Stability's SD3, SD3.5, and Stable Image models via
LiteLLM's OpenAI-compatible interface.
Changes:
- Add STABILITY provider to LlmProviders enum
- Create StabilityImageGenerationConfig with multipart/form-data support
- Add OpenAI size to Stability aspect_ratio mapping
- Register provider in ProviderConfigManager
- Add 9 Stability models to model_prices_and_context_window.json
- Add documentation at docs/providers/stability.md
- Add 25 unit tests
Supported models:
- stability/sd3, sd3-large, sd3-large-turbo, sd3-medium
- stability/sd3.5-large, sd3.5-large-turbo, sd3.5-medium
- stability/stable-image-ultra, stable-image-core
Fixes#15337
Perplexity API returns pre-calculated costs in `usage.cost.total_cost`
that include the `request_cost` (fixed per-request fee). LiteLLM was
ignoring this and calculating costs manually, resulting in ~27x
underreporting (e.g., $0.0002 vs actual $0.006).
Changes:
- Use `usage.cost.total_cost` from Perplexity response when available
- Fall back to manual calculation if cost object not present
- Add tests for both behaviors
- Filter skip_mcp_handler and other internal params in fallback_utils.py before calling acompletion
Fixes issue where internal parameters were being passed to provider APIs causing errors
- Remove deployment field from GCS bucket logger test metadata
Fixes model name mismatch where deployment field was overriding the model in logging
- Update Bedrock Titan test to use non-deprecated model (titan-text-express-v1)
Fixes test failure due to deprecated amazon.titan-text-lite-v1 model
- Add agent_id: null to expected JSON to match actual payload structure
- Fixes test_async_gcs_pub_sub_v1 test failure
- agent_id is an optional field in SpendLogsPayload that is always included (as null when not provided)
* fix(bedrock): filter non-serializable objects from request params
- Enhanced filter_exceptions_from_params() to filter callable objects (functions) and Logging objects
- Applied filtering in Bedrock's _prepare_request_params() before deepcopy
- Applied filtering to additional_request_params before JSON serialization
- Prevents TypeError during deepcopy (APIConnectionError objects) and JSON serialization (functions, Logging objects)
- Fixes test_bedrock_tool_calling test failures
Root cause: MCP-related functions (handle_chat_completion_with_mcp, completion_callable) and litellm_logging_obj were incorrectly added to optional_params via add_provider_specific_params_to_optional_params(), which then ended up in additional_request_params. These objects should be in litellm_params, not optional_params.
* fix(bedrock): filter internal MCP parameters from API requests
Filter out LiteLLM internal/MCP-related parameters (skip_mcp_handler,
mcp_handler_context, _skip_mcp_handler) from additional_request_params
before sending to Bedrock API to prevent 'extraneous key' errors.
- Added filter_internal_params() helper function in core_helpers.py
- Applied filtering in Bedrock's _prepare_request_params() method
- Fixes test_bedrock_completion.py::test_bedrock_tool_calling
* fix: mypy type error
* fix: add filter_exceptions_from_params to recursive function ignore list
- Add filter_exceptions_from_params to IGNORE_FUNCTIONS in recursive_detector.py
- Function is safe: has max_depth parameter (default 20) to prevent infinite recursion
* fix: correct Request headers format in JWT auth test
Fix test_jwt_non_admin_team_route_access by converting headers to bytes
format as required by Starlette's ASGI specification. Headers must be
bytes tuples with lowercase header names.
This allows dict(request.headers) to work correctly and enables the
authorization check to run, producing the expected error message.
* fix: ignore UUID trace_id from standard_logging_object, use litellm_call_id
The issue was that standard_logging_object.trace_id contains a UUID
(from litellm_trace_id default), which was being used instead of
falling back to litellm_call_id. This caused the test to fail because
it expected 'my-unique-call-id' but got a UUID.
Now we properly detect UUIDs (36 chars with 4 hyphens in specific positions)
and ignore them, allowing the fallback to litellm_call_id to work correctly.
This ensures we use litellm_call_id when no explicit trace_id is provided,
which gets stored in the cache and returned by _get_trace_id().
* fix: use existing_trace_id when provided instead of litellm_call_id
When existing_trace_id is provided in metadata, it should be used as the
trace_id to return (and store in cache), not litellm_call_id. This fixes
the test case where existing_trace_id is set and should be returned by
_get_trace_id().
Add support for the 'xhigh' reasoning effort level on all gpt-5.2 model
variants, not just gpt-5.2-pro. This enables deeper reasoning capabilities
for the base gpt-5.2 model.
Changes:
- Add is_model_gpt_5_2_model() method to detect gpt-5.2 variants
- Update xhigh validation to allow gpt-5.2 models
- Update documentation with gpt-5.2 reasoning_effort support
- Update tests to reflect new behavior
- Add database and Redis setup to litellm_mapped_tests_proxy job in CircleCI
- Create shared test helpers in tests/test_litellm/proxy/conftest.py for proxy test setup
- Refactor health endpoint tests to use shared helpers from conftest
- Support automatic Redis cache configuration when REDIS_HOST is set
- Ensure minimal config is created when Redis/database is needed
Fixes#17821
The `is_cached_message` function crashed with TypeError when message
content was a string instead of a list of content blocks.
Changes:
- Add explicit `isinstance(content, list)` check before iteration
- Add `isinstance(content_item, dict)` check inside loop to skip non-dict items
- Use `.get()` for safer nested dict access
- Follow same pattern as `extract_ttl_from_cached_messages` (same module)
Tests:
- Add TestIsCachedMessage class with 9 test cases covering:
- String content (the reported bug)
- None content
- Missing content key
- Empty list content
- List with/without cache_control
- Mixed content types (strings + dicts)
- Wrong cache_control type
Moved speechConfig from RequestBody to GenerationConfig TypedDict so that
TTS configuration survives the filtering in _transform_request_body().
This fixes the 400 INVALID_ARGUMENT error when using Gemini TTS models
(gemini-2.5-flash-tts, gemini-2.5-flash-preview-tts, etc.) with both
vertex_ai and gemini providers.
Fixes: speechConfig was being created correctly in map_openai_params()
but then filtered out because GenerationConfig.__annotations__.keys()
didn't include it.
Tested with both preview and non-preview TTS model names and both
vertex_ai and gemini providers.
* feat(langfuse): Add support for custom masking function
Allow users to pass a custom masking function via metadata to selectively
redact sensitive data (credit cards, emails, PII) before sending to Langfuse.
Usage:
```python
def mask_pii(data):
if isinstance(data, str):
data = re.sub(r'\b\d{4}[\s-]?\d{4}[\s-]?\d{4}[\s-]?\d{4}\b', '[CARD]', data)
return data
litellm.completion(
model="gpt-4",
messages=[...],
metadata={"langfuse_masking_function": mask_pii}
)
```
* fix(langfuse): Isolate masking function from other logging integrations
Extract langfuse_masking_function from metadata early in the flow and store
it in a dedicated key (_langfuse_masking_function) that only the Langfuse
logger knows to look for. This prevents the callable from leaking to other
logging integrations (Datadog, S3, etc.) which would serialize it as
"<function at 0x...>".
Changes:
- scrub_sensitive_keys_in_metadata() now extracts and stores the function
- Langfuse logger looks in the dedicated key first, falls back to metadata
- Added tests to verify isolation works correctly
* feat(deepseek): add native support for thinking and reasoning_effort params
Add proper parameter mapping for DeepSeek thinking mode, allowing users
to use the unified LiteLLM interface instead of extra_body workarounds.
Supported formats:
- thinking={"type": "enabled"}
- thinking={"type": "enabled", "budget_tokens": X} (budget_tokens ignored)
- reasoning_effort="low|medium|high" (maps to thinking enabled)
DeepSeek only supports {"type": "enabled"} without budget_tokens,
so any budget_tokens are stripped and all reasoning_effort values
(except "none") map to enabled.
Reference: https://api-docs.deepseek.com/guides/thinking_mode
* docs(deepseek): add thinking and reasoning_effort parameter documentation
Sameer Kankute