From 47b2479c94c5b00270b74fe4d22aff6be0add6cf Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Wed, 16 Sep 2026 18:23:14 -0700 Subject: [PATCH 1/2] fix(bedrock): gate Invoke tool search on the model map's supports_tool_search flag The Bedrock InvokeModel transformations decided whether to send the tool-search-tool-2025-10-19 beta from hardcoded model name lists (a pattern list on the messages path, an "opus-4" substring on the chat path), so Opus 4.8, Opus 5 and Sonnet 5 never got the beta on the messages path, Opus 5 and Sonnet 5 never got it on the chat path, Opus 4.1 got it without support, and /v1/model/info reported supports_tool_search as unset for all three. Both paths now read the model map through one shared helper: the Bedrock entries for Opus 4.8, Opus 5 and Sonnet 5 carry supports_tool_search explicitly, and a claude-tool-search fallback rule flags Claude 4.5 and newer for unmapped ids, inference-profile ARNs and mapped entries with no opinion, so the next Claude gets the beta with no code change. An explicit false on a resolved entry still wins. --- .../anthropic_claude3_transformation.py | 3 +- litellm/llms/bedrock/common_utils.py | 14 ++++ .../anthropic_claude3_transformation.py | 52 +++------------ ...odel_prices_and_context_window_backup.json | 36 ++++++++++ model_prices_and_context_window.json | 36 ++++++++++ tests/test_litellm/conftest.py | 15 +++++ .../test_fallback_generalizations.py | 40 ++++++++++++ ...ations_anthropic_claude3_transformation.py | 45 +++++++++++++ .../test_anthropic_claude3_transformation.py | 65 ++++++++++++------- 9 files changed, 239 insertions(+), 67 deletions(-) diff --git a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py index 38f280eef03..72bc43ba938 100644 --- a/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py +++ b/litellm/llms/bedrock/chat/invoke_transformations/anthropic_claude3_transformation.py @@ -18,6 +18,7 @@ from litellm.llms.bedrock.chat.invoke_transformations.base_invoke_transformation ) from litellm.llms.bedrock.common_utils import ( apply_bedrock_invoke_structured_output, + bedrock_supports_tool_search, get_anthropic_beta_from_headers, normalize_bedrock_opus_output_config_effort, normalize_custom_field_on_tools, @@ -265,7 +266,7 @@ class AmazonAnthropicClaudeConfig(AmazonInvokeConfig, AnthropicConfig): if tool_search_used and not (programmatic_tool_calling_used or input_examples_used): beta_set.discard(ANTHROPIC_TOOL_SEARCH_BETA_HEADER) - if "opus-4" in model.lower() or "opus_4" in model.lower(): + if bedrock_supports_tool_search(model): beta_set.add("tool-search-tool-2025-10-19") auto_beta_list: Final = filter_and_transform_beta_headers( diff --git a/litellm/llms/bedrock/common_utils.py b/litellm/llms/bedrock/common_utils.py index cb2c70e74c8..50a569a76c0 100644 --- a/litellm/llms/bedrock/common_utils.py +++ b/litellm/llms/bedrock/common_utils.py @@ -898,6 +898,20 @@ def bedrock_model_accepts_cache_points(model: str | None) -> bool: return any(entry.get("supports_prompt_caching") is True for entry in entries) +def bedrock_supports_tool_search(model: str) -> bool: + """ + Whether Bedrock InvokeModel admits the ``tool_search_tool_*`` tool types on ``model``. + + Backed by the ``supports_tool_search`` flag in ``model_prices_and_context_window.json``, + an exact entry or the ``claude-tool-search`` fallback rule for Claude 4.5 and newer, so a + newly released Claude carries the flag with no code change. An explicit ``false`` on the + resolved entry wins over the rule. + """ + from litellm.llms.anthropic.common_utils import AnthropicModelInfo + + return AnthropicModelInfo._supports_model_capability(model, "supports_tool_search", "bedrock") + + def is_claude_4_5_on_bedrock(model: str) -> bool: """ Check if the model supports Bedrock prompt caching with an extended '1h' TTL diff --git a/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py b/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py index a715d150b4c..4aa2afdbc78 100644 --- a/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py +++ b/litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude3_transformation.py @@ -31,6 +31,7 @@ from litellm.llms.bedrock.chat.invoke_transformations.base_invoke_transformation from litellm.llms.bedrock.common_utils import ( BedrockError, apply_bedrock_invoke_structured_output, + bedrock_supports_tool_search, ensure_bedrock_anthropic_messages_tool_names, get_anthropic_beta_from_headers, is_claude_4_5_on_bedrock, @@ -386,9 +387,10 @@ class AmazonAnthropicClaudeMessagesConfig( """ Check if the model supports tool search on Bedrock. - The model map's ``supports_tool_search`` flag is authoritative when - ``model`` resolves to an entry that sets it; the name patterns below - cover ids the map cannot resolve (ARNs, unlisted regional variants). + The model map's ``supports_tool_search`` flag is authoritative: an exact + entry, or the ``claude-tool-search`` fallback rule (Claude 4.5 and newer) + for ids the map cannot resolve (ARNs, unlisted regional variants) and for + mapped entries that carry no opinion. Ref: https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool @@ -398,46 +400,7 @@ class AmazonAnthropicClaudeMessagesConfig( Returns: True if the model supports tool search on Bedrock """ - catalog: Final = AnthropicModelInfo._get_provider_resolved_capability(model, "supports_tool_search", "bedrock") - if catalog is not None: - return catalog - - model_lower: Final = model.lower() - - supported_patterns: Final = [ - # Opus 4.5 - "opus-4.5", - "opus_4.5", - "opus-4-5", - "opus_4_5", - # Sonnet 4.5 - "sonnet-4.5", - "sonnet_4.5", - "sonnet-4-5", - "sonnet_4_5", - # Opus 4.6 - "opus-4.6", - "opus_4.6", - "opus-4-6", - "opus_4_6", - # sonnet 4.6 - "sonnet-4.6", - "sonnet_4.6", - "sonnet-4-6", - "sonnet_4_6", - # Opus 4.7 - "opus-4.7", - "opus_4.7", - "opus-4-7", - "opus_4_7", - # Haiku 4.5 - "haiku-4.5", - "haiku_4.5", - "haiku-4-5", - "haiku_4_5", - ] - - return any(pattern in model_lower for pattern in supported_patterns) + return bedrock_supports_tool_search(model) def _get_tool_search_beta_header_for_bedrock( self, @@ -453,7 +416,8 @@ class AmazonAnthropicClaudeMessagesConfig( Bedrock requires a different beta header for tool search than the Anthropic API when tool search is used without programmatic tool calling or input examples: `tool-search-tool-2025-10-19`, and only on - the models listed in `_supports_tool_search_on_bedrock`. + the models the model map flags as `supports_tool_search` + (`_supports_tool_search_on_bedrock`). Ref: https://platform.claude.com/docs/en/agents-and-tools/tool-use/tool-search-tool diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index e87a3fec99b..49113c994e3 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -1810,6 +1810,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1847,6 +1848,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1884,6 +1886,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1921,6 +1924,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1957,6 +1961,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1993,6 +1998,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2029,6 +2035,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2067,6 +2074,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2105,6 +2113,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2143,6 +2152,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2180,6 +2190,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2217,6 +2228,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2287,6 +2299,7 @@ "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2325,6 +2338,7 @@ "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2363,6 +2377,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2401,6 +2416,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2438,6 +2454,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2475,6 +2492,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -46114,6 +46132,7 @@ "cache_read_input_token_cost": 2.4e-07, "input_cost_per_token": 2.4e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -46147,6 +46166,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -46179,6 +46199,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -59477,6 +59498,15 @@ "supports_mid_conversation_system": true } }, + { + "name": "claude-tool-search", + "pattern": "claude-[a-z]+-(?:4[-._](?:[5-9]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", + "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"], + "description": "Claude at version 4.5 or higher, in any id shape that contains claude--: minors 4.5 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic shipped the tool search tool (tool_search_tool_regex, tool_search_tool_bm25) with Opus 4.5, Sonnet 4.5 and Haiku 4.5, and every newer Claude keeps it, so the flag follows the version instead of a per-model list. The Bedrock InvokeModel transformations read this flag to send the tool-search-tool-2025-10-19 beta, so a new Claude gets the beta with no code change.", + "model_info": { + "supports_tool_search": true + } + }, { "name": "wandb-reasoning-baseline", "pattern": "^wandb/", @@ -63214,6 +63244,7 @@ "cache_read_input_token_cost": 2.4e-07, "input_cost_per_token": 2.4e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63246,6 +63277,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63277,6 +63309,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63417,6 +63450,7 @@ "cache_read_input_token_cost": 2.4e-07, "input_cost_per_token": 2.4e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63449,6 +63483,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63480,6 +63515,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index e87a3fec99b..49113c994e3 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -1810,6 +1810,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1847,6 +1848,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1884,6 +1886,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1921,6 +1924,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1957,6 +1961,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -1993,6 +1998,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2029,6 +2035,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2067,6 +2074,7 @@ "cache_read_input_token_cost": 5e-07, "input_cost_per_token": 5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2105,6 +2113,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2143,6 +2152,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2180,6 +2190,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2217,6 +2228,7 @@ "cache_read_input_token_cost": 5.5e-07, "input_cost_per_token": 5.5e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2287,6 +2299,7 @@ "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2325,6 +2338,7 @@ "cache_read_input_token_cost": 2e-07, "input_cost_per_token": 2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2363,6 +2377,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2401,6 +2416,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2438,6 +2454,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -2475,6 +2492,7 @@ "cache_read_input_token_cost": 2.2e-07, "input_cost_per_token": 2.2e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -46114,6 +46132,7 @@ "cache_read_input_token_cost": 2.4e-07, "input_cost_per_token": 2.4e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -46147,6 +46166,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -46179,6 +46199,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock_converse", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -59477,6 +59498,15 @@ "supports_mid_conversation_system": true } }, + { + "name": "claude-tool-search", + "pattern": "claude-[a-z]+-(?:4[-._](?:[5-9]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", + "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"], + "description": "Claude at version 4.5 or higher, in any id shape that contains claude--: minors 4.5 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic shipped the tool search tool (tool_search_tool_regex, tool_search_tool_bm25) with Opus 4.5, Sonnet 4.5 and Haiku 4.5, and every newer Claude keeps it, so the flag follows the version instead of a per-model list. The Bedrock InvokeModel transformations read this flag to send the tool-search-tool-2025-10-19 beta, so a new Claude gets the beta with no code change.", + "model_info": { + "supports_tool_search": true + } + }, { "name": "wandb-reasoning-baseline", "pattern": "^wandb/", @@ -63214,6 +63244,7 @@ "cache_read_input_token_cost": 2.4e-07, "input_cost_per_token": 2.4e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63246,6 +63277,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63277,6 +63309,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63417,6 +63450,7 @@ "cache_read_input_token_cost": 2.4e-07, "input_cost_per_token": 2.4e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63449,6 +63483,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, @@ -63480,6 +63515,7 @@ "cache_read_input_token_cost": 6e-07, "input_cost_per_token": 6e-06, "litellm_provider": "bedrock", + "supports_tool_search": true, "max_input_tokens": 1000000, "max_output_tokens": 128000, "max_tokens": 128000, diff --git a/tests/test_litellm/conftest.py b/tests/test_litellm/conftest.py index a4f32df46ae..beca10d5555 100644 --- a/tests/test_litellm/conftest.py +++ b/tests/test_litellm/conftest.py @@ -206,6 +206,21 @@ def local_model_cost_map(monkeypatch): litellm.get_model_info.cache_clear() +@pytest.fixture +def local_beta_headers_config(monkeypatch): + """Pin the bundled ``anthropic_beta_headers_config.json`` so beta header assertions + do not depend on the network-fetched copy or on what earlier tests left cached.""" + from litellm.anthropic_beta_headers_manager import reload_beta_headers_config + + monkeypatch.setenv("LITELLM_LOCAL_ANTHROPIC_BETA_HEADERS", "True") + reload_beta_headers_config() + try: + yield + finally: + monkeypatch.delenv("LITELLM_LOCAL_ANTHROPIC_BETA_HEADERS", raising=False) + reload_beta_headers_config() + + def _run_coroutine_if_needed(result): if not asyncio.iscoroutine(result): return diff --git a/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py b/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py index 71e6e20b1a4..37a44867857 100644 --- a/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py +++ b/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py @@ -997,5 +997,45 @@ def test_shipped_claude_thinking_rules_backfill_only_anthropic(shipped_cost_map) assert match_fill_missing_generalizations("claude-sonnet-4-6", "anthropic") == { "supports_adaptive_thinking": True, "supports_legacy_thinking": True, + "supports_tool_search": True, } assert match_fill_missing_generalizations("claude-sonnet-4-6", "perplexity") is None + + +@pytest.mark.parametrize( + "model,provider,tool_search", + [ + ("us.anthropic.claude-opus-4-5", "bedrock", True), + ("claude-haiku-4-4", "anthropic", None), + ("claude-haiku-4-6", "anthropic", True), + ("claude-haiku-4-10", "anthropic", True), + ("claude-haiku-5-0", "anthropic", True), + ("claude-sonnet-5-1", "anthropic", True), + ("claude-newfam-6", "anthropic", True), + ("claude-haiku-4-20250514", "anthropic", None), + ], +) +def test_shipped_tool_search_rule_version_boundaries(shipped_cost_map, model, provider, tool_search): + """The claude-tool-search rule flags Claude 4.5 and newer in any family, bare major + or major-minor, and leaves 4.4 and date-suffixed 4.x ids without an opinion.""" + assert model not in litellm.model_cost + info = litellm.get_model_info(model, custom_llm_provider=provider) + assert info.get("supports_tool_search") is tool_search, model + + +def test_shipped_tool_search_rule_fills_mapped_claude_entries_without_flag(shipped_cost_map): + """A mapped Claude 4.5+ entry with no supports_tool_search key gets it from the rule + on the Claude providers, a mapped pre-4.5 entry stays without one, and a reseller + copy of the same model is not touched.""" + for key, model, provider in ( + ("claude-opus-4-7", "claude-opus-4-7", "anthropic"), + ("azure_ai/claude-opus-5", "claude-opus-5", "azure_ai"), + ): + assert "supports_tool_search" not in litellm.model_cost[key] + assert litellm.get_model_info(model, custom_llm_provider=provider)["supports_tool_search"] is True + + assert "supports_tool_search" not in litellm.model_cost["claude-opus-4-1"] + assert litellm.get_model_info("claude-opus-4-1", custom_llm_provider="anthropic").get("supports_tool_search") is None + + assert match_fill_missing_generalizations("claude-opus-5", "bedrock")["supports_tool_search"] is True + assert match_fill_missing_generalizations("claude-opus-5", "perplexity") is None diff --git a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py index bcba4bf7711..ec0bf6b842a 100644 --- a/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py +++ b/tests/test_litellm/llms/bedrock/chat/invoke_transformations/test_bedrock_chat_invoke_transformations_anthropic_claude3_transformation.py @@ -814,3 +814,48 @@ async def test_bedrock_invoke_claude_async_completion_inlines_document_url_sourc "type": "document", "source": {"type": "base64", "media_type": "application/pdf", "data": async_only_image_fetch.base64_png}, } in captured["body"]["messages"][0]["content"] + + +@pytest.mark.parametrize( + "model, expected_betas", + [ + pytest.param("us.anthropic.claude-opus-4-8", ["tool-search-tool-2025-10-19"], id="opus_4_8"), + pytest.param("us.anthropic.claude-opus-5", ["tool-search-tool-2025-10-19"], id="opus_5"), + pytest.param("us.anthropic.claude-sonnet-5", ["tool-search-tool-2025-10-19"], id="sonnet_5"), + pytest.param("us.anthropic.claude-haiku-4-5-20251001-v1:0", ["tool-search-tool-2025-10-19"], id="haiku_4_5"), + pytest.param("us.anthropic.claude-opus-4-1-20250805-v1:0", None, id="opus_4_1_unsupported"), + ], +) +def test_bedrock_chat_invoke_tool_search_beta_follows_model_map( + local_model_cost_map, local_beta_headers_config, model, expected_betas +): + """LIT-5851: the chat Invoke path used to add the ``tool-search-tool-2025-10-19`` + beta whenever the id contained ``opus-4``, so Opus 5 and Sonnet 5 lost it, Haiku + 4.5 never had it, and Opus 4.1 got it without support. The gate now follows the + model map's ``supports_tool_search`` flag, shared with the messages path.""" + result = AmazonAnthropicClaudeConfig().transform_request( + model=model, + messages=[{"role": "user", "content": "Add 2 and 3"}], + optional_params={ + "max_tokens": 64, + "tools": [ + {"type": "tool_search_tool_regex_20251119", "name": "tool_search_tool_regex"}, + { + "type": "function", + "function": { + "name": "add_numbers", + "description": "Add two integers", + "parameters": { + "type": "object", + "properties": {"a": {"type": "integer"}, "b": {"type": "integer"}}, + "required": ["a", "b"], + }, + }, + }, + ], + }, + litellm_params={}, + headers={}, + ) + + assert result.get("anthropic_beta") == expected_betas diff --git a/tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py b/tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py index 09ebc1a3c95..c503bf66ced 100644 --- a/tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py +++ b/tests/test_litellm/llms/bedrock/messages/invoke_transformations/test_anthropic_claude3_transformation.py @@ -2650,17 +2650,6 @@ def test_bedrock_clear_thinking_leaves_enabled_thinking_on_non_adaptive_model(): assert "output_config" not in request -@pytest.fixture -def local_beta_headers_config(monkeypatch): - from litellm.anthropic_beta_headers_manager import reload_beta_headers_config - - monkeypatch.setenv("LITELLM_LOCAL_ANTHROPIC_BETA_HEADERS", "True") - reload_beta_headers_config() - yield - monkeypatch.delenv("LITELLM_LOCAL_ANTHROPIC_BETA_HEADERS", raising=False) - reload_beta_headers_config() - - def test_bedrock_messages_preserves_clear_tool_uses_context_management_and_adds_beta( local_beta_headers_config, ): @@ -2826,9 +2815,12 @@ def test_filter_and_transform_beta_headers_passes_context_management_for_bedrock "us.anthropic.claude-haiku-4-5-20251001-v1:0", "us.anthropic.claude-sonnet-4-5-20250929-v1:0", "us.anthropic.claude-opus-4-7", + "us.anthropic.claude-opus-4-8", + "us.anthropic.claude-opus-5", + "us.anthropic.claude-sonnet-5", ], ) -def test_bedrock_messages_tool_search_adds_beta_header(local_beta_headers_config, model): +def test_bedrock_messages_tool_search_adds_beta_header(local_model_cost_map, local_beta_headers_config, model): """ LIT-4522: Bedrock InvokeModel only admits ``tool_search_tool_*`` tool types when the request body carries the ``tool-search-tool-2025-10-19`` beta; @@ -2838,6 +2830,11 @@ def test_bedrock_messages_tool_search_adds_beta_header(local_beta_headers_config Opus 4.7, so the beta was silently dropped for those models and every tool-search request failed. Verified live 2026-08-11: Bedrock returns 200 with ``server_tool_use`` for all three models once the beta is sent. + + LIT-5851: the same allowlist then missed Opus 4.8, Opus 5 and Sonnet 5, so + the gate now reads the model map's ``supports_tool_search`` flag (explicit + on the Bedrock entries, and the ``claude-tool-search`` rule for Claude 4.5 + and newer) instead of a per-model name list. """ from litellm.types.router import GenericLiteLLMParams @@ -2871,10 +2868,10 @@ def test_bedrock_messages_tool_search_adds_beta_header(local_beta_headers_config def test_bedrock_messages_tool_search_model_map_flag_is_authoritative(local_model_cost_map, monkeypatch): - """``supports_tool_search`` lives in the model map; the name patterns in - ``_supports_tool_search_on_bedrock`` are only a fallback for ids the map - cannot resolve. Flipping the mapped entry's flag to ``False`` must win even - though the model name still matches the ``haiku-4-5`` pattern.""" + """``supports_tool_search`` lives in the model map; the ``claude-tool-search`` + rule only fills entries that carry no opinion. Flipping the mapped entry's + flag to ``False`` must win even though the id is a Claude 4.5 the rule + would flag.""" import litellm from litellm.llms.anthropic.common_utils import AnthropicModelInfo @@ -2893,19 +2890,43 @@ def test_bedrock_messages_tool_search_model_map_flag_is_authoritative(local_mode @pytest.mark.parametrize( "model, expected", [ - pytest.param("us.anthropic.claude-opus-4-6-v99:9", True, id="unmapped_id_falls_back_to_patterns"), - pytest.param("anthropic.claude-3-5-sonnet-20240620-v1:0", False, id="mapped_entry_without_flag_no_pattern"), + pytest.param("us.anthropic.claude-opus-4-6-v99:9", True, id="unmapped_4_6_variant"), + pytest.param("us.anthropic.claude-haiku-5-2", True, id="unmapped_future_minor"), + pytest.param( + "arn:aws:bedrock:us-east-1:123456789012:inference-profile/us.anthropic.claude-opus-5", + True, + id="inference_profile_arn", + ), + pytest.param("anthropic.claude-3-5-sonnet-20240620-v1:0", False, id="mapped_claude_3_5_without_flag"), + pytest.param("us.anthropic.claude-opus-4-1-20250805-v1:0", False, id="mapped_opus_4_1_without_flag"), + pytest.param("us.anthropic.claude-sonnet-4-20250514-v1:0", False, id="mapped_dated_sonnet_4_without_flag"), ], ) -def test_bedrock_messages_tool_search_pattern_fallback(local_model_cost_map, model, expected): - """Ids the model map cannot resolve (or resolves without a - ``supports_tool_search`` opinion) fall through to the name patterns, so - ARNs and unlisted regional variants of supported families keep working.""" +def test_bedrock_messages_tool_search_follows_claude_tool_search_rule(local_model_cost_map, model, expected): + """Ids the model map cannot resolve, or resolves without a ``supports_tool_search`` + opinion, take the ``claude-tool-search`` fallback rule: Claude 4.5 and newer get + the beta, ARNs and unlisted regional variants included, and older Claudes do not.""" cfg = AmazonAnthropicClaudeMessagesConfig() assert cfg._supports_tool_search_on_bedrock(model) is expected +def test_bedrock_messages_tool_search_rule_fills_mapped_entry_without_flag(local_model_cost_map, monkeypatch): + """LIT-5851: a Bedrock entry that is in the map but carries no ``supports_tool_search`` + key, the state Opus 4.8, Opus 5 and Sonnet 5 shipped in, is filled by the + ``claude-tool-search`` rule instead of resolving to ``None`` and losing the beta.""" + import litellm + + model = "us.anthropic.claude-opus-5" + cfg = AmazonAnthropicClaudeMessagesConfig() + + monkeypatch.delitem(litellm.model_cost[model], "supports_tool_search") + litellm.get_model_info.cache_clear() + + assert litellm.get_model_info(model, custom_llm_provider="bedrock")["supports_tool_search"] is True + assert cfg._supports_tool_search_on_bedrock(model) is True + + def test_bedrock_messages_thinking_shape_follows_exact_bedrock_entry_flag( local_model_cost_map, monkeypatch ): From 064e49810a27985e0c6944017d215bb35d74495f Mon Sep 17 00:00:00 2001 From: mateo-berri <277851410+mateo-berri@users.noreply.github.com> Date: Thu, 17 Sep 2026 16:48:15 -0700 Subject: [PATCH 2/2] fix(bedrock): keep the tool search rule off azure_ai and pin dotted ids and Vertex fills --- ...odel_prices_and_context_window_backup.json | 4 ++-- model_prices_and_context_window.json | 4 ++-- .../test_fallback_generalizations.py | 19 ++++++++++++++----- 3 files changed, 18 insertions(+), 9 deletions(-) diff --git a/litellm/model_prices_and_context_window_backup.json b/litellm/model_prices_and_context_window_backup.json index 49113c994e3..fc6e69a50f5 100644 --- a/litellm/model_prices_and_context_window_backup.json +++ b/litellm/model_prices_and_context_window_backup.json @@ -59501,8 +59501,8 @@ { "name": "claude-tool-search", "pattern": "claude-[a-z]+-(?:4[-._](?:[5-9]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", - "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"], - "description": "Claude at version 4.5 or higher, in any id shape that contains claude--: minors 4.5 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic shipped the tool search tool (tool_search_tool_regex, tool_search_tool_bm25) with Opus 4.5, Sonnet 4.5 and Haiku 4.5, and every newer Claude keeps it, so the flag follows the version instead of a per-model list. The Bedrock InvokeModel transformations read this flag to send the tool-search-tool-2025-10-19 beta, so a new Claude gets the beta with no code change.", + "fill_missing_for_providers": ["anthropic", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"], + "description": "Claude at version 4.5 or higher, in any id shape that contains claude--: minors 4.5 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic's tool search docs list every Claude 4.5 and newer model as supported and Opus 4.1 and earlier as unsupported, so the flag follows the version instead of a per-model list. azure_ai is left out on purpose: Anthropic documents tool search as unavailable on Azure-hosted Foundry deployments, and the azure_ai/ key cannot tell those from Anthropic-hosted ones.", "model_info": { "supports_tool_search": true } diff --git a/model_prices_and_context_window.json b/model_prices_and_context_window.json index 49113c994e3..fc6e69a50f5 100644 --- a/model_prices_and_context_window.json +++ b/model_prices_and_context_window.json @@ -59501,8 +59501,8 @@ { "name": "claude-tool-search", "pattern": "claude-[a-z]+-(?:4[-._](?:[5-9]|[1-9]\\d)(?!\\d)|[5-9](?!\\d)(?:[-._]\\d{1,2}(?!\\d))?)", - "fill_missing_for_providers": ["anthropic", "azure_ai", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"], - "description": "Claude at version 4.5 or higher, in any id shape that contains claude--: minors 4.5 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic shipped the tool search tool (tool_search_tool_regex, tool_search_tool_bm25) with Opus 4.5, Sonnet 4.5 and Haiku 4.5, and every newer Claude keeps it, so the flag follows the version instead of a per-model list. The Bedrock InvokeModel transformations read this flag to send the tool-search-tool-2025-10-19 beta, so a new Claude gets the beta with no code change.", + "fill_missing_for_providers": ["anthropic", "bedrock", "bedrock_converse", "vertex_ai-anthropic_models"], + "description": "Claude at version 4.5 or higher, in any id shape that contains claude--: minors 4.5 through 4.99, any later major-minor, and bare 5+ majors so a new family like claude-fable-5 matches. Two-digit majors are deliberately not matched so ids like claude-opus-41 (4.1) are not read as major 41. Anthropic's tool search docs list every Claude 4.5 and newer model as supported and Opus 4.1 and earlier as unsupported, so the flag follows the version instead of a per-model list. azure_ai is left out on purpose: Anthropic documents tool search as unavailable on Azure-hosted Foundry deployments, and the azure_ai/ key cannot tell those from Anthropic-hosted ones.", "model_info": { "supports_tool_search": true } diff --git a/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py b/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py index 37a44867857..f7793286c7a 100644 --- a/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py +++ b/tests/test_litellm/litellm_core_utils/test_fallback_generalizations.py @@ -1008,6 +1008,8 @@ def test_shipped_claude_thinking_rules_backfill_only_anthropic(shipped_cost_map) ("us.anthropic.claude-opus-4-5", "bedrock", True), ("claude-haiku-4-4", "anthropic", None), ("claude-haiku-4-6", "anthropic", True), + ("claude-opus-4.5", "anthropic", True), + ("claude-opus-4_5", "anthropic", True), ("claude-haiku-4-10", "anthropic", True), ("claude-haiku-5-0", "anthropic", True), ("claude-sonnet-5-1", "anthropic", True), @@ -1017,7 +1019,8 @@ def test_shipped_claude_thinking_rules_backfill_only_anthropic(shipped_cost_map) ) def test_shipped_tool_search_rule_version_boundaries(shipped_cost_map, model, provider, tool_search): """The claude-tool-search rule flags Claude 4.5 and newer in any family, bare major - or major-minor, and leaves 4.4 and date-suffixed 4.x ids without an opinion.""" + or major-minor with a dash, dot or underscore delimiter, and leaves 4.4 and + date-suffixed 4.x ids without an opinion.""" assert model not in litellm.model_cost info = litellm.get_model_info(model, custom_llm_provider=provider) assert info.get("supports_tool_search") is tool_search, model @@ -1025,17 +1028,23 @@ def test_shipped_tool_search_rule_version_boundaries(shipped_cost_map, model, pr def test_shipped_tool_search_rule_fills_mapped_claude_entries_without_flag(shipped_cost_map): """A mapped Claude 4.5+ entry with no supports_tool_search key gets it from the rule - on the Claude providers, a mapped pre-4.5 entry stays without one, and a reseller - copy of the same model is not touched.""" + on Anthropic direct, Vertex and Bedrock, a mapped pre-4.5 entry stays without one, + and Azure Foundry and reseller copies of the same model are not touched.""" for key, model, provider in ( ("claude-opus-4-7", "claude-opus-4-7", "anthropic"), - ("azure_ai/claude-opus-5", "claude-opus-5", "azure_ai"), + ("vertex_ai/claude-opus-5", "claude-opus-5", "vertex_ai"), ): assert "supports_tool_search" not in litellm.model_cost[key] assert litellm.get_model_info(model, custom_llm_provider=provider)["supports_tool_search"] is True assert "supports_tool_search" not in litellm.model_cost["claude-opus-4-1"] - assert litellm.get_model_info("claude-opus-4-1", custom_llm_provider="anthropic").get("supports_tool_search") is None + opus_4_1_info = litellm.get_model_info("claude-opus-4-1", custom_llm_provider="anthropic") + assert opus_4_1_info.get("supports_tool_search") is None + + assert "supports_tool_search" not in litellm.model_cost["azure_ai/claude-opus-5"] + azure_opus_5_info = litellm.get_model_info("claude-opus-5", custom_llm_provider="azure_ai") + assert azure_opus_5_info.get("supports_tool_search") is None assert match_fill_missing_generalizations("claude-opus-5", "bedrock")["supports_tool_search"] is True + assert "supports_tool_search" not in match_fill_missing_generalizations("claude-opus-5", "azure_ai") assert match_fill_missing_generalizations("claude-opus-5", "perplexity") is None