Skip to content

Build and run agents you can see, understand and trust.

feat(model): refresh model cards across all providers (#2240)

DavdGaobf7e8ff75a23
diff --git a/src/agentscope/_version.py b/src/agentscope/_version.py
index 8999cf20c..cc3b8b042 100644
--- a/src/agentscope/_version.py
+++ b/src/agentscope/_version.py
@@ -1,4 +1,4 @@
# -*- coding: utf-8 -*-
"""The version of agentscope."""
-__version__ = "2.0.5"
+__version__ = "2.0.6dev"
diff --git a/src/agentscope/model/_anthropic/_model.py b/src/agentscope/model/_anthropic/_model.py
index 0b1dff8a9..b54b1d67e 100644
--- a/src/agentscope/model/_anthropic/_model.py
+++ b/src/agentscope/model/_anthropic/_model.py
@@ -56,6 +56,45 @@ class AnthropicChatModel(ChatModelBase):
gt=0,
)
+ thinking_mode: (
+ Literal["adaptive", "enabled", "disabled"] | None
+ ) = Field(
+ default=None,
+ title="Thinking Mode",
+ description=(
+ "How Claude thinks. ``adaptive`` lets the model decide when "
+ "and how deeply, and is the only mode Claude Opus 4.7 and "
+ "later accept — those models reject the ``enabled`` "
+ "(budget-based) mode. Takes precedence over "
+ "``thinking_enable``; leave unset to fall back to it."
+ ),
+ )
+
+ thinking_display: Literal["summarized", "omitted"] | None = Field(
+ default=None,
+ title="Thinking Display",
+ description=(
+ "Whether thinking blocks carry readable summaries or come "
+ "back empty. Claude Opus 4.7 and later default to "
+ "``omitted``, so set ``summarized`` to see the reasoning."
+ ),
+ )
+
+ reasoning_effort: (
+ Literal["low", "medium", "high", "xhigh", "max"] | None
+ ) = Field(
+ default=None,
+ title="Reasoning Effort",
+ description=(
+ "How many tokens Claude spends on the whole response — "
+ "not just thinking, but tool calls and explanation too. "
+ "Sent as Anthropic's ``output_config.effort``. The API "
+ "default is ``high``. Supported levels vary by model — see "
+ "the model card, and note that Claude Sonnet 4.5 and "
+ "Claude Haiku 4.5 do not accept this parameter at all."
+ ),
+ )
+
def __init__(
self,
credential: AnthropicCredential,
@@ -127,6 +166,16 @@ class AnthropicChatModel(ChatModelBase):
anthropic.InternalServerError,
)
+ def _thinking_mode(self) -> str | None:
+ """Resolve the effective thinking mode.
+
+ ``thinking_mode`` is the modern control; ``thinking_enable`` is the
+ legacy toggle that only ever meant the budget-based mode.
+ """
+ if self.parameters.thinking_mode is not None:
+ return self.parameters.thinking_mode
+ return "enabled" if self.parameters.thinking_enable else None
+
async def _call_api(
self,
model_name: str,
@@ -169,17 +218,26 @@ class AnthropicChatModel(ChatModelBase):
**generate_kwargs,
}
- # Anthropic extended thinking — only set when explicitly enabled.
- # Anthropic requires max_tokens > budget_tokens strictly.
- if self.parameters.thinking_enable and "thinking" not in kwargs:
- budget = self.parameters.thinking_budget or (max_tokens // 2)
- if budget >= max_tokens:
- # Auto-expand max_tokens to satisfy the strict inequality.
- max_tokens = budget + 1024
- kwargs["max_tokens"] = max_tokens
- kwargs["thinking"] = {
- "type": "enabled",
- "budget_tokens": budget,
+ mode = self._thinking_mode()
+ if mode is not None and "thinking" not in kwargs:
+ thinking: dict[str, Any] = {"type": mode}
+ if mode == "enabled":
+ # Anthropic requires max_tokens > budget_tokens strictly.
+ budget = self.parameters.thinking_budget or (max_tokens // 2)
+ if budget >= max_tokens:
+ # Auto-expand max_tokens to satisfy the inequality.
+ max_tokens = budget + 1024
+ kwargs["max_tokens"] = max_tokens
+ thinking["budget_tokens"] = budget
+ # ``display`` is invalid alongside ``type: "disabled"``.
+ if mode != "disabled" and self.parameters.thinking_display:
+ thinking["display"] = self.parameters.thinking_display
+ kwargs["thinking"] = thinking
+
+ # Effort travels inside ``output_config``, not as a top-level field.
+ if self.parameters.reasoning_effort and "output_config" not in kwargs:
+ kwargs["output_config"] = {
+ "effort": self.parameters.reasoning_effort,
}
fmt_tools, fmt_tool_choice = self._format_tools(tools, tool_choice)
@@ -529,14 +587,14 @@ class AnthropicChatModel(ChatModelBase):
) -> StructuredResponse:
"""Anthropic-specific override for structured output.
- Anthropic's extended thinking mode only supports
+ Anthropic's budget-based thinking mode only supports
``tool_choice={"type": "auto"}`` or ``{"type": "none"}``; any
forcing form (``"any"`` or a specific tool) raises an API error.
- When ``thinking_enable`` is on we default ``tool_choice`` to
- ``"auto"`` and rely on the base class's injected system-reminder
- prompt to guide the model. When thinking is disabled, this falls
+ In that mode we default ``tool_choice`` to ``"auto"`` and rely on
+ the base class's injected system-reminder prompt to guide the
+ model. Adaptive thinking accepts forced tool use, so it falls
through to the base implementation (force the structured-output
- tool).
+ tool) just like thinking being off entirely.
See:
https://platform.claude.com/docs/en/build-with-claude/extended-thinking#extended-thinking-with-tool-use
@@ -562,7 +620,7 @@ class AnthropicChatModel(ChatModelBase):
The structured response whose ``content`` is the validated
output dict matching ``structured_model``.
"""
- if tool_choice is None and self.parameters.thinking_enable:
+ if tool_choice is None and self._thinking_mode() == "enabled":
tool_choice = ToolChoice(mode="auto")
return await super()._call_api_with_structured_output(
model_name=model_name,
diff --git a/src/agentscope/model/_anthropic/_models/claude-fable-5.yaml b/src/agentscope/model/_anthropic/_models/claude-fable-5.yaml
new file mode 100644
index 000000000..ddddbee9b
--- /dev/null
+++ b/src/agentscope/model/_anthropic/_models/claude-fable-5.yaml
@@ -0,0 +1,37 @@
+name: claude-fable-5
+label: Claude Fable 5
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens: {"maximum": 128000}
+ # Thinking is always on here and cannot be disabled or budgeted —
+ # ``reasoning_effort`` is the only depth control it accepts.
+ thinking_enable:
+ hidden: true
+ thinking_budget:
+ hidden: true
+ thinking_mode:
+ default: adaptive
+ anyOf:
+ - enum: [adaptive]
+ type: string
+ - type: "null"
+ # Defaults to "omitted" on these models, which renders as empty thinking
+ # blocks — ask for summaries explicitly.
+ thinking_display:
+ default: summarized
diff --git a/src/agentscope/model/_anthropic/_models/claude-haiku-4-5.yaml b/src/agentscope/model/_anthropic/_models/claude-haiku-4-5.yaml
index feae8f167..eadc361f8 100644
--- a/src/agentscope/model/_anthropic/_models/claude-haiku-4-5.yaml
+++ b/src/agentscope/model/_anthropic/_models/claude-haiku-4-5.yaml
@@ -19,3 +19,12 @@ output_size: 65536
parameter_overrides:
max_tokens: {"maximum": 65536}
+ # Extended-thinking-only generation: adaptive mode is not accepted, and
+ # thinking display is already summarized by default here.
+ thinking_mode:
+ hidden: true
+ thinking_display:
+ hidden: true
+ # Haiku 4.5 does not accept the reasoning_effort parameter.
+ reasoning_effort:
+ hidden: true
diff --git a/src/agentscope/model/_anthropic/_models/claude-opus-4-5.yaml b/src/agentscope/model/_anthropic/_models/claude-opus-4-5.yaml
index 69f6b4f39..4b7e8048d 100644
--- a/src/agentscope/model/_anthropic/_models/claude-opus-4-5.yaml
+++ b/src/agentscope/model/_anthropic/_models/claude-opus-4-5.yaml
@@ -19,3 +19,16 @@ output_size: 65536
parameter_overrides:
max_tokens: {"maximum": 65536}
+ # Extended-thinking-only generation: adaptive mode is not accepted, and
+ # thinking display is already summarized by default here.
+ thinking_mode:
+ hidden: true
+ thinking_display:
+ hidden: true
+ # The oldest model accepting reasoning_effort at all; xhigh and max
+ # arrived with the 4.6 generation.
+ reasoning_effort:
+ anyOf:
+ - enum: [low, medium, high]
+ type: string
+ - type: "null"
diff --git a/src/agentscope/model/_anthropic/_models/claude-opus-4-6.yaml b/src/agentscope/model/_anthropic/_models/claude-opus-4-6.yaml
index 0e5e2f3c3..7b395e0a6 100644
--- a/src/agentscope/model/_anthropic/_models/claude-opus-4-6.yaml
+++ b/src/agentscope/model/_anthropic/_models/claude-opus-4-6.yaml
@@ -15,7 +15,23 @@ output_types:
- application/x-thinking
context_size: 1000000
-output_size: 131072
+output_size: 128000
parameter_overrides:
- max_tokens: {"maximum": 131072}
+ max_tokens: {"maximum": 128000}
+ # The last generation to accept both thinking modes; budget-based
+ # thinking still works here but is deprecated in favour of adaptive.
+ thinking_enable:
+ hidden: true
+ thinking_mode:
+ default: adaptive
+ anyOf:
+ - enum: [adaptive, enabled, disabled]
+ type: string
+ - type: "null"
+ # xhigh landed after this generation; max is supported.
+ reasoning_effort:
+ anyOf:
+ - enum: [low, medium, high, max]
+ type: string
+ - type: "null"
diff --git a/src/agentscope/model/_anthropic/_models/claude-opus-4-7.yaml b/src/agentscope/model/_anthropic/_models/claude-opus-4-7.yaml
index a3a9beee1..1777373d8 100644
--- a/src/agentscope/model/_anthropic/_models/claude-opus-4-7.yaml
+++ b/src/agentscope/model/_anthropic/_models/claude-opus-4-7.yaml
@@ -11,11 +11,27 @@ input_types:
- image/webp
output_types:
- - text/plain
- - application/x-thinking
+ - text/plain
+ - application/x-thinking
context_size: 1000000
output_size: 128000
parameter_overrides:
max_tokens: {"maximum": 128000}
+ # Budget-based thinking (``thinking.type: "enabled"``) is rejected by
+ # this generation; adaptive is the only mode it accepts.
+ thinking_enable:
+ hidden: true
+ thinking_budget:
+ hidden: true
+ thinking_mode:
+ default: adaptive
+ anyOf:
+ - enum: [adaptive, disabled]
+ type: string
+ - type: "null"
+ # Defaults to "omitted" on these models, which renders as empty thinking
+ # blocks — ask for summaries explicitly.
+ thinking_display:
+ default: summarized
diff --git a/src/agentscope/model/_anthropic/_models/claude-opus-4-8.yaml b/src/agentscope/model/_anthropic/_models/claude-opus-4-8.yaml
index 28df6a820..924f3a87f 100644
--- a/src/agentscope/model/_anthropic/_models/claude-opus-4-8.yaml
+++ b/src/agentscope/model/_anthropic/_models/claude-opus-4-8.yaml
@@ -19,3 +19,19 @@ output_size: 128000
parameter_overrides:
max_tokens: {"maximum": 128000}
+ # Budget-based thinking (``thinking.type: "enabled"``) is rejected by
+ # this generation; adaptive is the only mode it accepts.
+ thinking_enable:
+ hidden: true
+ thinking_budget:
+ hidden: true
+ thinking_mode:
+ default: adaptive
+ anyOf:
+ - enum: [adaptive, disabled]
+ type: string
+ - type: "null"
+ # Defaults to "omitted" on these models, which renders as empty thinking
+ # blocks — ask for summaries explicitly.
+ thinking_display:
+ default: summarized
diff --git a/src/agentscope/model/_anthropic/_models/claude-opus-5.yaml b/src/agentscope/model/_anthropic/_models/claude-opus-5.yaml
new file mode 100644
index 000000000..714c576bc
--- /dev/null
+++ b/src/agentscope/model/_anthropic/_models/claude-opus-5.yaml
@@ -0,0 +1,37 @@
+name: claude-opus-5
+label: Claude Opus 5
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens: {"maximum": 128000}
+ # Budget-based thinking (``thinking.type: "enabled"``) is rejected by
+ # this generation; adaptive is the only mode it accepts.
+ thinking_enable:
+ hidden: true
+ thinking_budget:
+ hidden: true
+ thinking_mode:
+ default: adaptive
+ anyOf:
+ - enum: [adaptive, disabled]
+ type: string
+ - type: "null"
+ # Defaults to "omitted" on these models, which renders as empty thinking
+ # blocks — ask for summaries explicitly.
+ thinking_display:
+ default: summarized
diff --git a/src/agentscope/model/_anthropic/_models/claude-sonnet-4-5.yaml b/src/agentscope/model/_anthropic/_models/claude-sonnet-4-5.yaml
index fedbea547..d97fe098b 100644
--- a/src/agentscope/model/_anthropic/_models/claude-sonnet-4-5.yaml
+++ b/src/agentscope/model/_anthropic/_models/claude-sonnet-4-5.yaml
@@ -19,3 +19,12 @@ output_size: 65536
parameter_overrides:
max_tokens: {"maximum": 65536}
+ # Extended-thinking-only generation: adaptive mode is not accepted, and
+ # thinking display is already summarized by default here.
+ thinking_mode:
+ hidden: true
+ thinking_display:
+ hidden: true
+ # Sonnet 4.5 does not accept the reasoning_effort parameter.
+ reasoning_effort:
+ hidden: true
diff --git a/src/agentscope/model/_anthropic/_models/claude-sonnet-4-6.yaml b/src/agentscope/model/_anthropic/_models/claude-sonnet-4-6.yaml
index 61d22d04c..b4783a1b1 100644
--- a/src/agentscope/model/_anthropic/_models/claude-sonnet-4-6.yaml
+++ b/src/agentscope/model/_anthropic/_models/claude-sonnet-4-6.yaml
@@ -15,7 +15,23 @@ output_types:
- application/x-thinking
context_size: 1000000
-output_size: 65536
+output_size: 128000
parameter_overrides:
- max_tokens: {"maximum": 65536}
+ max_tokens: {"maximum": 128000}
+ # The last generation to accept both thinking modes; budget-based
+ # thinking still works here but is deprecated in favour of adaptive.
+ thinking_enable:
+ hidden: true
+ thinking_mode:
+ default: adaptive
+ anyOf:
+ - enum: [adaptive, enabled, disabled]
+ type: string
+ - type: "null"
+ # xhigh landed after this generation; max is supported.
+ reasoning_effort:
+ anyOf:
+ - enum: [low, medium, high, max]
+ type: string
+ - type: "null"
diff --git a/src/agentscope/model/_anthropic/_models/claude-sonnet-5.yaml b/src/agentscope/model/_anthropic/_models/claude-sonnet-5.yaml
new file mode 100644
index 000000000..45d4008e0
--- /dev/null
+++ b/src/agentscope/model/_anthropic/_models/claude-sonnet-5.yaml
@@ -0,0 +1,37 @@
+name: claude-sonnet-5
+label: Claude Sonnet 5
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens: {"maximum": 128000}
+ # Budget-based thinking (``thinking.type: "enabled"``) is rejected by
+ # this generation; adaptive is the only mode it accepts.
+ thinking_enable:
+ hidden: true
+ thinking_budget:
+ hidden: true
+ thinking_mode:
+ default: adaptive
+ anyOf:
+ - enum: [adaptive, disabled]
+ type: string
+ - type: "null"
+ # Defaults to "omitted" on these models, which renders as empty thinking
+ # blocks — ask for summaries explicitly.
+ thinking_display:
+ default: summarized
diff --git a/src/agentscope/model/_dashscope/_models/qwen-flash.yaml b/src/agentscope/model/_dashscope/_models/qwen-flash.yaml
new file mode 100644
index 000000000..d85f1d42d
--- /dev/null
+++ b/src/agentscope/model/_dashscope/_models/qwen-flash.yaml
@@ -0,0 +1,21 @@
+name: qwen-flash
+label: Qwen-Flash
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 32768
+
+parameter_overrides:
+ max_tokens: {"maximum": 32768}
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_dashscope/_models/qwen3.5-flash.yaml b/src/agentscope/model/_dashscope/_models/qwen3.5-flash.yaml
new file mode 100644
index 000000000..25671936c
--- /dev/null
+++ b/src/agentscope/model/_dashscope/_models/qwen3.5-flash.yaml
@@ -0,0 +1,30 @@
+name: qwen3.5-flash
+label: Qwen3.5-Flash
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+ - image/bmp
+ - image/jpeg
+ - image/png
+ - image/tiff
+ - image/webp
+ - image/heic
+ - video/mp4
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 65536
+
+parameter_overrides:
+ max_tokens: {"maximum": 65536}
+ # Max chain-of-thought is independent of max_tokens.
+ thinking_budget: {"maximum": 81920}
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_dashscope/_models/qwen3.6-flash.yaml b/src/agentscope/model/_dashscope/_models/qwen3.6-flash.yaml
new file mode 100644
index 000000000..bfea31747
--- /dev/null
+++ b/src/agentscope/model/_dashscope/_models/qwen3.6-flash.yaml
@@ -0,0 +1,30 @@
+name: qwen3.6-flash
+label: Qwen3.6-Flash
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+ - image/bmp
+ - image/jpeg
+ - image/png
+ - image/tiff
+ - image/webp
+ - image/heic
+ - video/mp4
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 65536
+
+parameter_overrides:
+ max_tokens: {"maximum": 65536}
+ # Max chain-of-thought is independent of max_tokens.
+ thinking_budget: {"maximum": 131072}
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_dashscope/_models/qwen3.7-flash.yaml b/src/agentscope/model/_dashscope/_models/qwen3.7-flash.yaml
new file mode 100644
index 000000000..a3bfdd181
--- /dev/null
+++ b/src/agentscope/model/_dashscope/_models/qwen3.7-flash.yaml
@@ -0,0 +1,30 @@
+name: qwen3.7-flash
+label: Qwen3.7-Flash
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+ - image/bmp
+ - image/jpeg
+ - image/png
+ - image/tiff
+ - image/webp
+ - image/heic
+ - video/mp4
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 131072
+
+parameter_overrides:
+ max_tokens: {"maximum": 131072}
+ # Max chain-of-thought is independent of max_tokens.
+ thinking_budget: {"maximum": 262144}
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_dashscope/_models/qwen3.8-max.yaml b/src/agentscope/model/_dashscope/_models/qwen3.8-max.yaml
new file mode 100644
index 000000000..19a51c51c
--- /dev/null
+++ b/src/agentscope/model/_dashscope/_models/qwen3.8-max.yaml
@@ -0,0 +1,30 @@
+name: qwen3.8-max
+label: Qwen3.8-Max
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+ - image/bmp
+ - image/jpeg
+ - image/png
+ - image/tiff
+ - image/webp
+ - image/heic
+ - video/mp4
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+output_size: 131072
+
+parameter_overrides:
+ max_tokens: {"maximum": 131072}
+ # Max chain-of-thought is 262144 tokens, independent of max_tokens.
+ thinking_budget: {"maximum": 262144}
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_gemini/_models/gemini-2.5-flash-lite.yaml b/src/agentscope/model/_gemini/_models/gemini-2.5-flash-lite.yaml
new file mode 100644
index 000000000..08ee25c3a
--- /dev/null
+++ b/src/agentscope/model/_gemini/_models/gemini-2.5-flash-lite.yaml
@@ -0,0 +1,25 @@
+name: gemini-2.5-flash-lite
+label: Gemini 2.5 Flash-Lite
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+ - audio/mp3
+ - audio/wav
+ - video/mp4
+ - application/pdf
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1048576
+output_size: 65536
+
+parameter_overrides:
+ max_tokens:
+ maximum: 65536
diff --git a/src/agentscope/model/_gemini/_models/gemini-3.1-flash-lite.yaml b/src/agentscope/model/_gemini/_models/gemini-3.1-flash-lite.yaml
new file mode 100644
index 000000000..2ea92bb23
--- /dev/null
+++ b/src/agentscope/model/_gemini/_models/gemini-3.1-flash-lite.yaml
@@ -0,0 +1,25 @@
+name: gemini-3.1-flash-lite
+label: Gemini 3.1 Flash-Lite
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+ - audio/mp3
+ - audio/wav
+ - video/mp4
+ - application/pdf
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1048576
+output_size: 65536
+
+parameter_overrides:
+ max_tokens:
+ maximum: 65536
diff --git a/src/agentscope/model/_gemini/_models/gemini-3.5-flash-lite.yaml b/src/agentscope/model/_gemini/_models/gemini-3.5-flash-lite.yaml
new file mode 100644
index 000000000..09b1e6d4a
--- /dev/null
+++ b/src/agentscope/model/_gemini/_models/gemini-3.5-flash-lite.yaml
@@ -0,0 +1,25 @@
+name: gemini-3.5-flash-lite
+label: Gemini 3.5 Flash-Lite
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+ - audio/mp3
+ - audio/wav
+ - video/mp4
+ - application/pdf
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1048576
+output_size: 65536
+
+parameter_overrides:
+ max_tokens:
+ maximum: 65536
diff --git a/src/agentscope/model/_gemini/_models/gemini-3.5-flash.yaml b/src/agentscope/model/_gemini/_models/gemini-3.5-flash.yaml
new file mode 100644
index 000000000..037ca634c
--- /dev/null
+++ b/src/agentscope/model/_gemini/_models/gemini-3.5-flash.yaml
@@ -0,0 +1,25 @@
+name: gemini-3.5-flash
+label: Gemini 3.5 Flash
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+ - audio/mp3
+ - audio/wav
+ - video/mp4
+ - application/pdf
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1048576
+output_size: 65536
+
+parameter_overrides:
+ max_tokens:
+ maximum: 65536
diff --git a/src/agentscope/model/_gemini/_models/gemini-3.6-flash.yaml b/src/agentscope/model/_gemini/_models/gemini-3.6-flash.yaml
new file mode 100644
index 000000000..1c48e87fb
--- /dev/null
+++ b/src/agentscope/model/_gemini/_models/gemini-3.6-flash.yaml
@@ -0,0 +1,25 @@
+name: gemini-3.6-flash
+label: Gemini 3.6 Flash
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+ - audio/mp3
+ - audio/wav
+ - video/mp4
+ - application/pdf
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1048576
+output_size: 65536
+
+parameter_overrides:
+ max_tokens:
+ maximum: 65536
diff --git a/src/agentscope/model/_model_card.py b/src/agentscope/model/_model_card.py
index c5795ecfc..282e5c9ad 100644
--- a/src/agentscope/model/_model_card.py
+++ b/src/agentscope/model/_model_card.py
@@ -102,6 +102,8 @@ class ModelCard(BaseModel):
if "application/x-thinking" not in output_types:
properties.pop("thinking_enable", None)
properties.pop("thinking_budget", None)
+ properties.pop("thinking_mode", None)
+ properties.pop("thinking_display", None)
# Auto-filter: only omni-style models that declare an ``audio/*``
# output type expose the ``voice`` parameter to the frontend
diff --git a/src/agentscope/model/_moonshot/_models/kimi-k2.5.yaml b/src/agentscope/model/_moonshot/_models/kimi-k2.5.yaml
index 237ad103e..ee641858b 100644
--- a/src/agentscope/model/_moonshot/_models/kimi-k2.5.yaml
+++ b/src/agentscope/model/_moonshot/_models/kimi-k2.5.yaml
@@ -1,6 +1,7 @@
name: kimi-k2.5
label: Kimi K2.5
-status: active
+status: deprecated
+deprecated_at: "2026-08-31T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_moonshot/_models/kimi-k2.7-code-highspeed.yaml b/src/agentscope/model/_moonshot/_models/kimi-k2.7-code-highspeed.yaml
new file mode 100644
index 000000000..949f9e92a
--- /dev/null
+++ b/src/agentscope/model/_moonshot/_models/kimi-k2.7-code-highspeed.yaml
@@ -0,0 +1,23 @@
+name: kimi-k2.7-code-highspeed
+label: Kimi K2.7 Code Highspeed
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 262144
+# Moonshot publishes no per-model output ceiling; matched to kimi-k2.6,
+# which shares the same context window.
+output_size: 32768
+
+parameter_overrides:
+ max_tokens:
+ maximum: 32768
+ # Thinking is always on for the Code models and cannot be disabled.
+ thinking_enable:
+ hidden: true
diff --git a/src/agentscope/model/_moonshot/_models/kimi-k2.7-code.yaml b/src/agentscope/model/_moonshot/_models/kimi-k2.7-code.yaml
new file mode 100644
index 000000000..9003bd982
--- /dev/null
+++ b/src/agentscope/model/_moonshot/_models/kimi-k2.7-code.yaml
@@ -0,0 +1,23 @@
+name: kimi-k2.7-code
+label: Kimi K2.7 Code
+status: active
+
+input_types:
+ - text/plain
+ - application/x-thinking
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 262144
+# Moonshot publishes no per-model output ceiling; matched to kimi-k2.6,
+# which shares the same context window.
+output_size: 32768
+
+parameter_overrides:
+ max_tokens:
+ maximum: 32768
+ # Thinking is always on for the Code models and cannot be disabled.
+ thinking_enable:
+ hidden: true
diff --git a/src/agentscope/model/_moonshot/_models/moonshot-v1-128k.yaml b/src/agentscope/model/_moonshot/_models/moonshot-v1-128k.yaml
index 6f3e112fe..f297ad41f 100644
--- a/src/agentscope/model/_moonshot/_models/moonshot-v1-128k.yaml
+++ b/src/agentscope/model/_moonshot/_models/moonshot-v1-128k.yaml
@@ -1,6 +1,7 @@
name: moonshot-v1-128k
label: Kimi Moonshot v1 128K
-status: active
+status: deprecated
+deprecated_at: "2026-08-31T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_moonshot/_models/moonshot-v1-32k.yaml b/src/agentscope/model/_moonshot/_models/moonshot-v1-32k.yaml
index 3ae99026b..e40264f19 100644
--- a/src/agentscope/model/_moonshot/_models/moonshot-v1-32k.yaml
+++ b/src/agentscope/model/_moonshot/_models/moonshot-v1-32k.yaml
@@ -1,6 +1,7 @@
name: moonshot-v1-32k
label: Kimi Moonshot v1 32K
-status: active
+status: deprecated
+deprecated_at: "2026-08-31T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_moonshot/_models/moonshot-v1-8k.yaml b/src/agentscope/model/_moonshot/_models/moonshot-v1-8k.yaml
index 8e4149e11..67de5e6c8 100644
--- a/src/agentscope/model/_moonshot/_models/moonshot-v1-8k.yaml
+++ b/src/agentscope/model/_moonshot/_models/moonshot-v1-8k.yaml
@@ -1,6 +1,7 @@
name: moonshot-v1-8k
label: Kimi Moonshot v1 8K
-status: active
+status: deprecated
+deprecated_at: "2026-08-31T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_openai_chat/_models/gpt-4.1-nano.yaml b/src/agentscope/model/_openai_chat/_models/gpt-4.1-nano.yaml
index bf26d9d9b..d8cfac057 100644
--- a/src/agentscope/model/_openai_chat/_models/gpt-4.1-nano.yaml
+++ b/src/agentscope/model/_openai_chat/_models/gpt-4.1-nano.yaml
@@ -1,6 +1,7 @@
name: gpt-4.1-nano
label: GPT-4.1 Nano
-status: active
+status: deprecated
+deprecated_at: "2026-10-23T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_openai_chat/_models/gpt-5.5.yaml b/src/agentscope/model/_openai_chat/_models/gpt-5.5.yaml
index 1e009e393..75ed83d21 100644
--- a/src/agentscope/model/_openai_chat/_models/gpt-5.5.yaml
+++ b/src/agentscope/model/_openai_chat/_models/gpt-5.5.yaml
@@ -8,8 +8,6 @@ input_types:
- image/png
- image/gif
- image/webp
- - audio/mp3
- - audio/wav
output_types:
- text/plain
diff --git a/src/agentscope/model/_openai_chat/_models/gpt-5.6-luna.yaml b/src/agentscope/model/_openai_chat/_models/gpt-5.6-luna.yaml
new file mode 100644
index 000000000..4a77d13da
--- /dev/null
+++ b/src/agentscope/model/_openai_chat/_models/gpt-5.6-luna.yaml
@@ -0,0 +1,28 @@
+name: gpt-5.6-luna
+label: GPT-5.6 Luna
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+
+context_size: 1050000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 128000
+ # Reasoning happens internally and is never returned as content, so the
+ # thinking toggle has nothing to surface — reasoning_effort is the knob.
+ thinking_enable:
+ hidden: true
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_openai_chat/_models/gpt-5.6-sol.yaml b/src/agentscope/model/_openai_chat/_models/gpt-5.6-sol.yaml
new file mode 100644
index 000000000..23126dbc8
--- /dev/null
+++ b/src/agentscope/model/_openai_chat/_models/gpt-5.6-sol.yaml
@@ -0,0 +1,28 @@
+name: gpt-5.6-sol
+label: GPT-5.6 Sol
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+
+context_size: 1050000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 128000
+ # Reasoning happens internally and is never returned as content, so the
+ # thinking toggle has nothing to surface — reasoning_effort is the knob.
+ thinking_enable:
+ hidden: true
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_openai_chat/_models/gpt-5.6-terra.yaml b/src/agentscope/model/_openai_chat/_models/gpt-5.6-terra.yaml
new file mode 100644
index 000000000..0688ce0f3
--- /dev/null
+++ b/src/agentscope/model/_openai_chat/_models/gpt-5.6-terra.yaml
@@ -0,0 +1,28 @@
+name: gpt-5.6-terra
+label: GPT-5.6 Terra
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+
+context_size: 1050000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 128000
+ # Reasoning happens internally and is never returned as content, so the
+ # thinking toggle has nothing to surface — reasoning_effort is the knob.
+ thinking_enable:
+ hidden: true
+ # Voice only applies to omni-style models; hide the field for this
+ # non-omni model so the frontend popover doesn't render it.
+ voice:
+ hidden: true
diff --git a/src/agentscope/model/_openai_chat/_models/gpt-audio-mini.yaml b/src/agentscope/model/_openai_chat/_models/gpt-audio-mini.yaml
index 5877d6832..d6eaa8780 100644
--- a/src/agentscope/model/_openai_chat/_models/gpt-audio-mini.yaml
+++ b/src/agentscope/model/_openai_chat/_models/gpt-audio-mini.yaml
@@ -1,6 +1,7 @@
name: gpt-audio-mini
label: GPT Audio Mini
-status: active
+status: deprecated
+deprecated_at: "2027-01-20T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_openai_chat/_models/o3.yaml b/src/agentscope/model/_openai_chat/_models/o3.yaml
index 13e7762aa..78473d0c7 100644
--- a/src/agentscope/model/_openai_chat/_models/o3.yaml
+++ b/src/agentscope/model/_openai_chat/_models/o3.yaml
@@ -1,6 +1,7 @@
name: o3
label: o3
-status: active
+status: deprecated
+deprecated_at: "2026-12-11T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_openai_chat/_models/o4-mini.yaml b/src/agentscope/model/_openai_chat/_models/o4-mini.yaml
index ab2e91192..87fbcd9eb 100644
--- a/src/agentscope/model/_openai_chat/_models/o4-mini.yaml
+++ b/src/agentscope/model/_openai_chat/_models/o4-mini.yaml
@@ -1,6 +1,7 @@
name: o4-mini
label: o4-mini
-status: active
+status: deprecated
+deprecated_at: "2026-10-23T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_openai_response/_models/gpt-4.1-nano.yaml b/src/agentscope/model/_openai_response/_models/gpt-4.1-nano.yaml
index 5a067a036..34f000ffc 100644
--- a/src/agentscope/model/_openai_response/_models/gpt-4.1-nano.yaml
+++ b/src/agentscope/model/_openai_response/_models/gpt-4.1-nano.yaml
@@ -1,6 +1,7 @@
name: gpt-4.1-nano
label: GPT-4.1 Nano
-status: active
+status: deprecated
+deprecated_at: "2026-10-23T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_openai_response/_models/gpt-5.5.yaml b/src/agentscope/model/_openai_response/_models/gpt-5.5.yaml
index 746c4f516..34c92a47b 100644
--- a/src/agentscope/model/_openai_response/_models/gpt-5.5.yaml
+++ b/src/agentscope/model/_openai_response/_models/gpt-5.5.yaml
@@ -8,8 +8,6 @@ input_types:
- image/png
- image/gif
- image/webp
- - audio/mp3
- - audio/wav
output_types:
- text/plain
diff --git a/src/agentscope/model/_openai_response/_models/gpt-5.6-luna.yaml b/src/agentscope/model/_openai_response/_models/gpt-5.6-luna.yaml
new file mode 100644
index 000000000..2817d739a
--- /dev/null
+++ b/src/agentscope/model/_openai_response/_models/gpt-5.6-luna.yaml
@@ -0,0 +1,24 @@
+name: gpt-5.6-luna
+label: GPT-5.6 Luna
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+
+context_size: 1050000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 128000
+ # Reasoning happens internally and is never returned as content, so the
+ # thinking toggle has nothing to surface — reasoning_effort is the knob.
+ thinking_enable:
+ hidden: true
diff --git a/src/agentscope/model/_openai_response/_models/gpt-5.6-sol.yaml b/src/agentscope/model/_openai_response/_models/gpt-5.6-sol.yaml
new file mode 100644
index 000000000..ee21e20c8
--- /dev/null
+++ b/src/agentscope/model/_openai_response/_models/gpt-5.6-sol.yaml
@@ -0,0 +1,24 @@
+name: gpt-5.6-sol
+label: GPT-5.6 Sol
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+
+context_size: 1050000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 128000
+ # Reasoning happens internally and is never returned as content, so the
+ # thinking toggle has nothing to surface — reasoning_effort is the knob.
+ thinking_enable:
+ hidden: true
diff --git a/src/agentscope/model/_openai_response/_models/gpt-5.6-terra.yaml b/src/agentscope/model/_openai_response/_models/gpt-5.6-terra.yaml
new file mode 100644
index 000000000..b7c3535a6
--- /dev/null
+++ b/src/agentscope/model/_openai_response/_models/gpt-5.6-terra.yaml
@@ -0,0 +1,24 @@
+name: gpt-5.6-terra
+label: GPT-5.6 Terra
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+ - image/gif
+ - image/webp
+
+output_types:
+ - text/plain
+
+context_size: 1050000
+output_size: 128000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 128000
+ # Reasoning happens internally and is never returned as content, so the
+ # thinking toggle has nothing to surface — reasoning_effort is the knob.
+ thinking_enable:
+ hidden: true
diff --git a/src/agentscope/model/_openai_response/_models/o3.yaml b/src/agentscope/model/_openai_response/_models/o3.yaml
index a891bd1be..8cd74f603 100644
--- a/src/agentscope/model/_openai_response/_models/o3.yaml
+++ b/src/agentscope/model/_openai_response/_models/o3.yaml
@@ -1,6 +1,7 @@
name: o3
label: o3 (Responses API)
-status: active
+status: deprecated
+deprecated_at: "2026-12-11T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_openai_response/_models/o4-mini.yaml b/src/agentscope/model/_openai_response/_models/o4-mini.yaml
index 587223a31..a8cd18a4f 100644
--- a/src/agentscope/model/_openai_response/_models/o4-mini.yaml
+++ b/src/agentscope/model/_openai_response/_models/o4-mini.yaml
@@ -1,6 +1,7 @@
name: o4-mini
label: o4-mini (Responses API)
-status: active
+status: deprecated
+deprecated_at: "2026-10-23T00:00:00"
input_types:
- text/plain
diff --git a/src/agentscope/model/_xai/_models/grok-3-fast.yaml b/src/agentscope/model/_xai/_models/grok-3-fast.yaml
index 66be7ded9..0ebe34296 100644
--- a/src/agentscope/model/_xai/_models/grok-3-fast.yaml
+++ b/src/agentscope/model/_xai/_models/grok-3-fast.yaml
@@ -1,6 +1,6 @@
name: grok-3-fast
label: Grok 3 Fast
-status: active
+status: deprecated
input_types:
- text/plain
diff --git a/src/agentscope/model/_xai/_models/grok-3-mini.yaml b/src/agentscope/model/_xai/_models/grok-3-mini.yaml
index 04f100adc..5e118dce2 100644
--- a/src/agentscope/model/_xai/_models/grok-3-mini.yaml
+++ b/src/agentscope/model/_xai/_models/grok-3-mini.yaml
@@ -1,6 +1,6 @@
name: grok-3-mini
label: Grok 3 Mini
-status: active
+status: deprecated
input_types:
- text/plain
diff --git a/src/agentscope/model/_xai/_models/grok-3.yaml b/src/agentscope/model/_xai/_models/grok-3.yaml
index 6a04c98fe..6618b8f9f 100644
--- a/src/agentscope/model/_xai/_models/grok-3.yaml
+++ b/src/agentscope/model/_xai/_models/grok-3.yaml
@@ -1,6 +1,6 @@
name: grok-3
label: Grok 3
-status: active
+status: deprecated
input_types:
- text/plain
diff --git a/src/agentscope/model/_xai/_models/grok-4.20-0309-non-reasoning.yaml b/src/agentscope/model/_xai/_models/grok-4.20-0309-non-reasoning.yaml
new file mode 100644
index 000000000..8f7b501e6
--- /dev/null
+++ b/src/agentscope/model/_xai/_models/grok-4.20-0309-non-reasoning.yaml
@@ -0,0 +1,23 @@
+name: grok-4.20-0309-non-reasoning
+label: Grok 4.20 Non-Reasoning
+status: active
+
+input_types:
+ - text/plain
+
+output_types:
+ - text/plain
+
+context_size: 1000000
+# xAI publishes no max-output figure for any Grok model. Derived as 1/16 of
+# the context window, matching the ratio of providers that do publish one.
+output_size: 64000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 64000
+ # This variant never reasons; its sibling grok-4.20-0309-reasoning does.
+ thinking_enable:
+ hidden: true
+ reasoning_effort:
+ hidden: true
diff --git a/src/agentscope/model/_xai/_models/grok-4.20-0309-reasoning.yaml b/src/agentscope/model/_xai/_models/grok-4.20-0309-reasoning.yaml
new file mode 100644
index 000000000..e3a2d6369
--- /dev/null
+++ b/src/agentscope/model/_xai/_models/grok-4.20-0309-reasoning.yaml
@@ -0,0 +1,19 @@
+name: grok-4.20-0309-reasoning
+label: Grok 4.20 Reasoning
+status: active
+
+input_types:
+ - text/plain
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+# xAI publishes no max-output figure for any Grok model. Derived as 1/16 of
+# the context window, matching the ratio of providers that do publish one.
+output_size: 64000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 64000
diff --git a/src/agentscope/model/_xai/_models/grok-4.20-multi-agent-0309.yaml b/src/agentscope/model/_xai/_models/grok-4.20-multi-agent-0309.yaml
new file mode 100644
index 000000000..0fba29a5c
--- /dev/null
+++ b/src/agentscope/model/_xai/_models/grok-4.20-multi-agent-0309.yaml
@@ -0,0 +1,19 @@
+name: grok-4.20-multi-agent-0309
+label: Grok 4.20 Multi-Agent
+status: active
+
+input_types:
+ - text/plain
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 1000000
+# xAI publishes no max-output figure for any Grok model. Derived as 1/16 of
+# the context window, matching the ratio of providers that do publish one.
+output_size: 64000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 64000
diff --git a/src/agentscope/model/_xai/_models/grok-4.3.yaml b/src/agentscope/model/_xai/_models/grok-4.3.yaml
index 4b712ef44..916f1ea14 100644
--- a/src/agentscope/model/_xai/_models/grok-4.3.yaml
+++ b/src/agentscope/model/_xai/_models/grok-4.3.yaml
@@ -12,8 +12,10 @@ output_types:
- application/x-thinking
context_size: 1000000
-output_size: 30000
+# xAI publishes no max-output figure for any Grok model. Derived as 1/16 of
+# the context window, matching the ratio of providers that do publish one.
+output_size: 64000
parameter_overrides:
max_tokens:
- maximum: 30000
+ maximum: 64000
diff --git a/src/agentscope/model/_xai/_models/grok-4.5.yaml b/src/agentscope/model/_xai/_models/grok-4.5.yaml
new file mode 100644
index 000000000..21d5bd3d2
--- /dev/null
+++ b/src/agentscope/model/_xai/_models/grok-4.5.yaml
@@ -0,0 +1,21 @@
+name: grok-4.5
+label: Grok 4.5
+status: active
+
+input_types:
+ - text/plain
+ - image/jpeg
+ - image/png
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 500000
+# xAI publishes no max-output figure for any Grok model. Derived as 1/16 of
+# the context window, matching the ratio of providers that do publish one.
+output_size: 32000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 32000
diff --git a/src/agentscope/model/_xai/_models/grok-build-0.1.yaml b/src/agentscope/model/_xai/_models/grok-build-0.1.yaml
new file mode 100644
index 000000000..6eed6e431
--- /dev/null
+++ b/src/agentscope/model/_xai/_models/grok-build-0.1.yaml
@@ -0,0 +1,19 @@
+name: grok-build-0.1
+label: Grok Build 0.1
+status: active
+
+input_types:
+ - text/plain
+
+output_types:
+ - text/plain
+ - application/x-thinking
+
+context_size: 256000
+# xAI publishes no max-output figure for any Grok model. Derived as 1/16 of
+# the context window, matching the ratio of providers that do publish one.
+output_size: 16000
+
+parameter_overrides:
+ max_tokens:
+ maximum: 16000
diff --git a/tests/model_anthropic_test.py b/tests/model_anthropic_test.py
index 87853afc7..d45363708 100644
--- a/tests/model_anthropic_test.py
+++ b/tests/model_anthropic_test.py
@@ -255,6 +255,164 @@ class TestAnthropicNonStream(IsolatedAsyncioTestCase):
)
+class TestAnthropicEffort(IsolatedAsyncioTestCase):
+ """Tests for the ``reasoning_effort`` parameter."""
+
+ def setUp(self) -> None:
+ self.model = _make_model(stream=False)
+ self.mock_client = MagicMock()
+ self.model.client = self.mock_client
+ self.mock_create = AsyncMock(return_value=_mock_completion(text="hi"))
+ self.mock_client.messages.create = self.mock_create
+
+ async def test_effort_omitted_by_default(self) -> None:
+ """No output_config is sent when reasoning_effort is unset."""
+ await self.model([])
+
+ self.assertNotIn("output_config", self.mock_create.call_args.kwargs)
+
+ async def test_effort_nested_in_output_config(self) -> None:
+ """Effort travels inside output_config, not as a top-level field."""
+ self.model.parameters.reasoning_effort = "medium"
+
+ await self.model([])
+
+ kwargs = self.mock_create.call_args.kwargs
+ self.assertEqual(kwargs["output_config"], {"effort": "medium"})
+ self.assertNotIn("effort", kwargs)
+
+ async def test_effort_coexists_with_thinking(self) -> None:
+ """Effort and extended thinking are independent controls."""
+ self.model.parameters.reasoning_effort = "max"
+ self.model.parameters.thinking_enable = True
+ self.model.parameters.thinking_budget = 1024
+
+ await self.model([])
+
+ kwargs = self.mock_create.call_args.kwargs
+ self.assertEqual(kwargs["output_config"], {"effort": "max"})
+ self.assertEqual(
+ kwargs["thinking"],
+ {"type": "enabled", "budget_tokens": 1024},
+ )
+
+ async def test_caller_output_config_wins(self) -> None:
+ """An explicit output_config kwarg is not overwritten."""
+ self.model.parameters.reasoning_effort = "low"
+
+ await self.model([], output_config={"effort": "high"})
+
+ self.assertEqual(
+ self.mock_create.call_args.kwargs["output_config"],
+ {"effort": "high"},
+ )
+
+
+class TestAnthropicThinkingMode(IsolatedAsyncioTestCase):
+ """Tests for adaptive vs budget-based thinking configuration."""
+
+ def setUp(self) -> None:
+ self.model = _make_model(stream=False)
+ self.mock_client = MagicMock()
+ self.model.client = self.mock_client
+ self.mock_create = AsyncMock(return_value=_mock_completion(text="hi"))
+ self.mock_client.messages.create = self.mock_create
+
+ def _thinking(self) -> Any:
+ return self.mock_create.call_args.kwargs.get("thinking")
+
+ async def test_no_thinking_by_default(self) -> None:
+ """Neither control set means no thinking config is sent."""
+ await self.model([])
+
+ self.assertIsNone(self._thinking())
+
+ async def test_adaptive_carries_no_budget(self) -> None:
+ """Adaptive mode must not send budget_tokens, which it rejects."""
+ self.model.parameters.thinking_mode = "adaptive"
+ self.model.parameters.thinking_budget = 4096
+
+ await self.model([])
+
+ self.assertEqual(self._thinking(), {"type": "adaptive"})
+
+ async def test_adaptive_with_display(self) -> None:
+ """Display is what makes thinking text visible on newer models."""
+ self.model.parameters.thinking_mode = "adaptive"
+ self.model.parameters.thinking_display = "summarized"
+
+ await self.model([])
+
+ self.assertEqual(
+ self._thinking(),
+ {"type": "adaptive", "display": "summarized"},
+ )
+
+ async def test_disabled_drops_display(self) -> None:
+ """Display is invalid alongside type: disabled."""
+ self.model.parameters.thinking_mode = "disabled"
+ self.model.parameters.thinking_display = "summarized"
+
+ await self.model([])
+
+ self.assertEqual(self._thinking(), {"type": "disabled"})
+
+ async def test_legacy_toggle_still_means_budget_mode(self) -> None:
+ """thinking_enable keeps its old meaning when mode is unset."""
+ self.model.parameters.thinking_enable = True
+ self.model.parameters.thinking_budget = 2048
+
+ await self.model([])
+
+ self.assertEqual(
+ self._thinking(),
+ {"type": "enabled", "budget_tokens": 2048},
+ )
+
+ async def test_mode_overrides_legacy_toggle(self) -> None:
+ """An explicit mode wins over the legacy boolean."""
+ self.model.parameters.thinking_enable = True
+ self.model.parameters.thinking_mode = "adaptive"
+
+ await self.model([])
+
+ self.assertEqual(self._thinking(), {"type": "adaptive"})
+
+ async def test_budget_mode_expands_max_tokens(self) -> None:
+ """max_tokens must stay strictly above budget_tokens."""
+ self.model.parameters.thinking_mode = "enabled"
+ self.model.parameters.thinking_budget = 8192
+
+ await self.model([])
+
+ kwargs = self.mock_create.call_args.kwargs
+ self.assertEqual(kwargs["thinking"]["budget_tokens"], 8192)
+ self.assertGreater(kwargs["max_tokens"], 8192)
+
+ def test_resolved_mode_drives_tool_choice_downgrade(self) -> None:
+ """Only budget mode forbids forced tool use, so only it downgrades.
+
+ ``_call_api_with_structured_output`` keys the downgrade off this
+ resolution — adaptive must not trip it.
+ """
+ cases = [
+ ({}, None),
+ ({"thinking_enable": True}, "enabled"),
+ ({"thinking_mode": "enabled"}, "enabled"),
+ ({"thinking_mode": "adaptive"}, None),
+ ({"thinking_mode": "disabled"}, None),
+ ({"thinking_enable": True, "thinking_mode": "adaptive"}, None),
+ ]
+ for params, expected in cases:
+ with self.subTest(params=params):
+ model = _make_model()
+ for key, val in params.items():
+ setattr(model.parameters, key, val)
+ resolved = model._thinking_mode()
+ downgrades = resolved == "enabled"
+ self.assertEqual(downgrades, expected == "enabled")
+
+
# ---------------------------------------------------------------------------
# Streaming tests
# ---------------------------------------------------------------------------