fix(anthropic): translate adaptive thinking/effort to pre-4.6 model support

litellm/llms/anthropic/experimental_pass_through/messages/transformation.py

@@ -3,6 +3,7 @@
 import httpx
 
 from litellm.constants import (
+    ANTHROPIC_MIN_THINKING_BUDGET_TOKENS,
     DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
     DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
     DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET,
@@ -32,6 +33,12 @@
 
 DEFAULT_ANTHROPIC_API_VERSION = "2023-06-01"
 
+DROP_UNSUPPORTED_ADAPTIVE_EFFORT_WARNING = (
+    "Dropping adaptive `thinking`/`output_config.effort` for model=%s: the model "
+    "does not support extended thinking, or max_tokens is too small to fit the "
+    "minimum thinking budget."
+)
+
 
 class AnthropicMessagesConfig(BaseAnthropicMessagesConfig):
     def get_supported_anthropic_messages_params(self, model: str) -> list:
@@ -253,6 +260,111 @@ def _translate_legacy_thinking_for_adaptive_model(model: str, optional_params: D
         existing_output_config.setdefault("effort", effort)
         optional_params["output_config"] = existing_output_config
 
+    @staticmethod
+    def _translate_adaptive_effort_for_non_adaptive_model(
+        model: str, optional_params: Dict, max_tokens: Optional[int]
+    ) -> None:
+        """Translate the 4.6+ adaptive-thinking interface (``thinking.type=adaptive``
+        and/or ``output_config.effort``) down to what an older Anthropic model
+        supports. Clients like Claude Code send this interface unconditionally, so
+        without translation it reaches a pre-4.6 model and Anthropic rejects it with
+        "This model does not support the effort parameter".
+
+        The reshape is silent, matching how the messages path already strips
+        unsupported ``output_config`` for older models (bedrock invoke, issue
+        #22797): the goal is to keep the request working, not to fail it.
+
+        ``thinking.type=adaptive`` and ``output_config.effort`` are independent
+        capabilities. Adaptive thinking needs ``supports_adaptive_thinking`` (4.6+);
+        ``output_config.effort`` needs ``supports_output_config``, which some
+        non-adaptive models (e.g. Claude Opus 4.5) advertise on its own. So the two
+        are handled separately:
+
+        - Adaptive-thinking models (4.6+): both are native, left untouched.
+        - ``supports_output_config`` but non-adaptive (Opus 4.5): keep
+          ``output_config.effort`` (native), only drop the unsupported adaptive
+          ``thinking`` block. When adaptive thinking is being dropped and the
+          effort level itself isn't supported by the model (e.g. ``xhigh``/``max``
+          on Opus 4.5, which only accepts low/medium/high, while ``xhigh`` is
+          Claude Code's default), fall through to the legacy translation below
+          instead of forwarding a level Anthropic would reject. Effort-only
+          requests are always left untouched: provider subclasses own their level
+          normalization (bedrock clamps ``xhigh`` to the model's ceiling after
+          this base transform runs).
+        - Thinking-capable but neither (``supports_reasoning``, e.g. Haiku/Sonnet
+          4.5): map effort to legacy ``thinking={type: enabled, budget_tokens}`` via
+          ``AnthropicConfig._map_reasoning_effort``, capped below ``max_tokens``
+          (Anthropic requires ``max_tokens > budget_tokens``) and dropped when
+          ``max_tokens`` can't fit even the minimum budget.
+        - No reasoning support: ``thinking`` is dropped.
+
+        For the last two, only the consumed ``effort`` key is removed from
+        ``output_config``; any residual (e.g. ``format``) is left for provider
+        subclasses to handle.
+        """
+        from litellm.exceptions import BadRequestError as _BadRequestError
+        from litellm.llms.anthropic.chat.transformation import AnthropicConfig
+
+        if AnthropicConfig._is_adaptive_thinking_model(model):
+            return
+
+        output_config = optional_params.get("output_config")
+        thinking = optional_params.get("thinking")
+        effort = output_config.get("effort") if isinstance(output_config, dict) else None
+        adaptive_thinking = isinstance(thinking, dict) and thinking.get("type") == "adaptive"
+        if effort is None and not adaptive_thinking:
+            return
+
+        if AnthropicConfig._model_supports_effort_param(model) and (
+            not adaptive_thinking or AnthropicConfig._validate_effort_for_model(model, effort) is None
+        ):
+            if adaptive_thinking:
+                optional_params.pop("thinking", None)
+            return
+
+        supports_thinking = AnthropicModelInfo._supports_model_capability(model, "supports_reasoning")
+        try:
+            legacy_thinking = (
+                AnthropicConfig._map_reasoning_effort(reasoning_effort=effort or "medium", model=model)
+                if supports_thinking
+                else None
+            )
+        except _BadRequestError as e:
+            raise AnthropicError(message=str(e.message), status_code=400)
+        capped_thinking = (
+            AnthropicMessagesConfig._cap_thinking_budget_to_max_tokens(legacy_thinking, max_tokens)
+            if legacy_thinking is not None
+            else None
+        )
+
+        if capped_thinking is not None:
+            optional_params["thinking"] = capped_thinking
+        else:
+            verbose_logger.warning(DROP_UNSUPPORTED_ADAPTIVE_EFFORT_WARNING, model)
+            optional_params.pop("thinking", None)
+
+        if isinstance(output_config, dict) and "effort" in output_config:
+            residual = {k: v for k, v in output_config.items() if k != "effort"}
+            if residual:
+                optional_params["output_config"] = residual
+            else:
+                optional_params.pop("output_config", None)
+
+    @staticmethod
+    def _cap_thinking_budget_to_max_tokens(thinking: Dict, max_tokens: Optional[int]) -> Optional[Dict]:
+        """Cap a legacy ``thinking.budget_tokens`` below ``max_tokens`` (Anthropic
+        requires ``max_tokens > budget_tokens``). Returns the (possibly capped)
+        thinking dict, or ``None`` when ``max_tokens`` is too small to fit even the
+        minimum thinking budget and thinking should be dropped."""
+        budget = thinking.get("budget_tokens")
+        if max_tokens is None or not isinstance(budget, int):
+            return thinking
+        if max_tokens <= ANTHROPIC_MIN_THINKING_BUDGET_TOKENS:
+            return None
+        if budget < max_tokens:
+            return thinking
+        return {**thinking, "budget_tokens": max_tokens - 1}
+
     def transform_anthropic_messages_request(
         self,
         model: str,
@@ -284,6 +396,12 @@ def transform_anthropic_messages_request(
             optional_params=anthropic_messages_optional_request_params,
         )
 
+        self._translate_adaptive_effort_for_non_adaptive_model(
+            model=model,
+            optional_params=anthropic_messages_optional_request_params,
+            max_tokens=max_tokens,
+        )
+
         system_param = anthropic_messages_optional_request_params.get("system")
         if self.should_strip_billing_metadata() and system_param is not None:
             filtered_system = self._filter_billing_headers_from_system(system_param)

tests/test_litellm/llms/anthropic/experimental_pass_through/messages/test_anthropic_messages_effort.py

@@ -0,0 +1,189 @@
+import pytest
+
+from litellm.constants import (
+    DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
+    DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
+    DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET,
+)
+from litellm.llms.anthropic.common_utils import AnthropicError
+from litellm.llms.anthropic.experimental_pass_through.messages.transformation import (
+    AnthropicMessagesConfig,
+)
+
+
+def _claude_code_payload(effort="medium", max_tokens=8192, **output_config_extra):
+    """The exact adaptive-thinking shape Claude Code (claude-cli) sends."""
+    output_config = {"effort": effort, **output_config_extra}
+    return {
+        "max_tokens": max_tokens,
+        "thinking": {"type": "adaptive"},
+        "output_config": output_config,
+    }
+
+
+def _transform(model, params, litellm_params=None):
+    return AnthropicMessagesConfig().transform_anthropic_messages_request(
+        model=model,
+        messages=[{"role": "user", "content": "Hello"}],
+        anthropic_messages_optional_request_params=dict(params),
+        litellm_params=litellm_params or {},
+        headers={},
+    )
+
+
+def test_effort_translated_to_legacy_thinking_for_haiku_4_5():
+    """Core regression: Claude Code sends adaptive thinking + effort to Haiku 4.5
+    (thinking-capable, pre-4.6). Effort must be translated to legacy extended
+    thinking rather than forwarded raw (which Anthropic rejects with "This model
+    does not support the effort parameter")."""
+    result = _transform("claude-haiku-4-5", _claude_code_payload(effort="medium"))
+
+    assert result["thinking"] == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
+    }
+    assert "output_config" not in result
+
+
+def test_effort_high_maps_to_high_budget_for_sonnet_4_5():
+    result = _transform("claude-sonnet-4-5", _claude_code_payload(effort="high"))
+
+    assert result["thinking"] == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_HIGH_THINKING_BUDGET,
+    }
+    assert "output_config" not in result
+
+
+def test_adaptive_effort_passes_through_untouched_for_4_6():
+    """4.6+ natively supports the adaptive interface, so it must not be rewritten."""
+    result = _transform("claude-sonnet-4-6", _claude_code_payload(effort="high"))
+
+    assert result["thinking"] == {"type": "adaptive"}
+    assert result["output_config"] == {"effort": "high"}
+
+
+def test_thinking_and_effort_dropped_for_non_reasoning_model():
+    """A model with no reasoning support cannot take thinking or effort, so both are
+    silently dropped (no drop_params required) so the request still succeeds."""
+    result = _transform("claude-3-5-haiku-latest", _claude_code_payload(effort="medium"))
+
+    assert "thinking" not in result
+    assert "output_config" not in result
+
+
+def test_residual_output_config_preserved_after_effort_translation():
+    """output_config may carry `format` (structured outputs) alongside effort. Only
+    the consumed effort key is removed; the residual is left for provider subclasses
+    (bedrock/vertex) to handle, and effort is translated to legacy thinking."""
+    result = _transform(
+        "claude-haiku-4-5",
+        _claude_code_payload(effort="medium", format={"type": "json_schema"}),
+    )
+
+    assert result["thinking"] == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_MEDIUM_THINKING_BUDGET,
+    }
+    assert result["output_config"] == {"format": {"type": "json_schema"}}
+
+
+def test_opus_4_5_keeps_effort_but_drops_adaptive_thinking():
+    """Regression: Opus 4.5 advertises supports_output_config (accepts
+    output_config.effort) but is NOT adaptive, so thinking:{type:adaptive} is
+    rejected by Anthropic. The effort must be kept and only the adaptive thinking
+    block dropped, rather than early-returning and forwarding adaptive thinking raw."""
+    result = _transform("claude-opus-4-5", _claude_code_payload(effort="medium"))
+
+    assert result["output_config"] == {"effort": "medium"}
+    assert "thinking" not in result
+
+
+def test_opus_4_5_preserves_native_effort_without_adaptive_thinking():
+    """A caller sending output_config.effort alone (no adaptive thinking) to Opus 4.5
+    must pass through untouched, since the model supports it natively."""
+    result = AnthropicMessagesConfig().transform_anthropic_messages_request(
+        model="claude-opus-4-5",
+        messages=[{"role": "user", "content": "Hello"}],
+        anthropic_messages_optional_request_params={
+            "max_tokens": 8192,
+            "output_config": {"effort": "high"},
+        },
+        litellm_params={},
+        headers={},
+    )
+
+    assert result["output_config"] == {"effort": "high"}
+    assert "thinking" not in result
+
+
+def test_opus_4_5_unsupported_effort_level_translated_to_legacy_thinking():
+    """Opus 4.5 accepts output_config.effort but only levels low/medium/high;
+    Claude Code defaults to xhigh on newer models, and forwarding that level raw
+    would be rejected with "effort='xhigh' is not supported by this model". An
+    unsupported level must fall through to the legacy translation (budget-based
+    thinking, effort stripped) instead of being preserved."""
+    result = _transform("claude-opus-4-5", _claude_code_payload(effort="xhigh", max_tokens=64000))
+
+    assert result["thinking"] == {
+        "type": "enabled",
+        "budget_tokens": DEFAULT_REASONING_EFFORT_XHIGH_THINKING_BUDGET,
+    }
+    assert "output_config" not in result
+
+
+def test_opus_4_5_effort_only_unsupported_level_left_for_provider_normalization():
+    """An effort-only request (no adaptive thinking) must pass through untouched even
+    when the level exceeds what the model supports: provider subclasses own their
+    level normalization (bedrock clamps xhigh to the model's ceiling after this base
+    transform runs), so consuming the effort here breaks that contract."""
+    result = _transform(
+        "claude-opus-4-5",
+        {"max_tokens": 4096, "output_config": {"effort": "xhigh"}},
+    )
+
+    assert result["output_config"] == {"effort": "xhigh"}
+    assert "thinking" not in result
+
+
+def test_budget_capped_below_max_tokens():
+    """Adaptive thinking carries no budget, so the translated legacy budget must be
+    capped below max_tokens (Anthropic requires max_tokens > budget_tokens). A
+    high-effort budget (4096) with max_tokens=3000 must be capped to 2999."""
+    result = _transform("claude-haiku-4-5", _claude_code_payload(effort="high", max_tokens=3000))
+
+    assert result["thinking"] == {"type": "enabled", "budget_tokens": 2999}
+
+
+def test_thinking_dropped_when_max_tokens_too_small_for_min_budget():
+    """When max_tokens can't fit even the minimum thinking budget, thinking is
+    silently dropped so the request still succeeds rather than being rejected."""
+    result = _transform("claude-haiku-4-5", _claude_code_payload(effort="medium", max_tokens=512))
+
+    assert "thinking" not in result
+    assert "output_config" not in result
+
+
+def test_unrecognized_effort_raises_clean_400():
+    """An unrecognized effort value (e.g. a future Anthropic tier) must surface as a
+    clean AnthropicError 400, matching _translate_reasoning_effort_to_anthropic,
+    rather than leaking litellm's internal BadRequestError."""
+    with pytest.raises(AnthropicError) as exc_info:
+        _transform("claude-haiku-4-5", _claude_code_payload(effort="turbo"))
+
+    assert exc_info.value.status_code == 400
+
+
+def test_non_adaptive_request_without_effort_is_untouched():
+    """A non-adaptive model receiving a request with no adaptive interface (no
+    effort, no adaptive thinking) must pass through untouched."""
+    result = AnthropicMessagesConfig().transform_anthropic_messages_request(
+        model="claude-haiku-4-5",
+        messages=[{"role": "user", "content": "Hello"}],
+        anthropic_messages_optional_request_params={"max_tokens": 1024},
+        litellm_params={},
+        headers={},
+    )
+
+    assert "thinking" not in result
+    assert "output_config" not in result