docs: document provider reasoning request formats

2026-06-26 09:37:21 -05:00
parent d6e5057cfa
commit 4e85eac00a
533 changed files with 2155 additions and 6 deletions
@@ -1,5 +1,11 @@
 name = "302.AI"
 env = ["302AI_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Audited POST https://api.302.ai/v1/chat/completions. The provider's API guide
+# documents model/messages only; no reasoning toggle, effort, or numeric budget
+# request field is documented. Do not infer passthrough from upstream APIs.
+# Sources:
+# https://doc.302.ai/
 doc = "https://doc.302.ai"
 api = "https://api.302.ai/v1"
@@ -1,5 +1,11 @@
 name = "Abacus"
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Audited POST https://routellm.abacus.ai/v1/chat/completions. The provider API
+# reference documents no reasoning toggle, effort, or numeric budget request
+# field. Do not infer behavior from the routed model developer's API.
+# Sources:
+# https://abacus.ai/help/api
 env = ["ABACUS_API_KEY"]
 doc = "https://abacus.ai/help/api"
 api = "https://routellm.abacus.ai/v1"
@@ -1,4 +1,10 @@
 name = "Abliterated Model"
+# Reasoning HTTP format (accessed 2026-06-25):
+# This model thinks by default. On POST /v1/chat/completions or /v1/messages,
+# top-level `thinking: false` skips thinking; omission keeps it enabled.
+# Sources:
+# https://docs.abliteration.ai/models
+# https://docs.abliteration.ai/capabilities/thinking
 release_date = "2026-01-06"
 last_updated = "2026-01-06"
 attachment = true
@@ -1,5 +1,12 @@
 name = "abliteration.ai"
 env = ["ABLIT_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# POST /v1/chat/completions and POST /v1/messages: top-level `thinking` is true
+# by default; false skips thinking. POST /v1/responses has no thinking toggle.
+# No effort or numeric reasoning-budget request field is documented.
+# Sources:
+# https://docs.abliteration.ai/capabilities/thinking
+# https://docs.abliteration.ai/compatibility-matrix
 api = "https://api.abliteration.ai/v1"
 doc = "https://docs.abliteration.ai/models"
@@ -5,6 +5,7 @@ last_updated = "2026-03-13"
 attachment = true
 reasoning = true
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "max"] }, { type = "budget_tokens", min = 1_024 }]
+# Native Messages prefers $.thinking.type = "adaptive" with $.output_config.effort = "low"|"medium"|"high"|"max"; enabled budget_tokens >= 1024 is deprecated and must be < $.max_tokens. https://docs.aihubmix.com/cn/api/Claude-Native (accessed 2026-06-25)
 temperature = true
 tool_call = true
 structured_output = true
@@ -5,6 +5,7 @@ last_updated = "2026-04-16"
 attachment = true
 reasoning = true
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]
+# Native Messages uses $.thinking.type = "adaptive" and $.output_config.effort = "low"|"medium"|"high"|"xhigh"|"max"; manual budget_tokens is rejected. https://docs.aihubmix.com/cn/blogs/Claude-Opus4.7 (accessed 2026-06-25)
 temperature = false
 tool_call = true
 structured_output = true
@@ -5,6 +5,7 @@ last_updated = "2026-03-13"
 attachment = true
 reasoning = true
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "max"] }, { type = "budget_tokens", min = 1_024 }]
+# Native Messages prefers $.thinking.type = "adaptive" with $.output_config.effort = "low"|"medium"|"high"; enabled budget_tokens >= 1024 is deprecated and must be < $.max_tokens. Chat effort "max" maps to native "high". https://docs.aihubmix.com/cn/api/Claude-Native (accessed 2026-06-25)
 temperature = true
 tool_call = true
 structured_output = true
@@ -5,6 +5,7 @@ last_updated = "2025-06-05"
 attachment = true
 reasoning = true
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }, { type = "budget_tokens", min = 0, max = 24_576 }]
+# Native Gemini uses $.generationConfig.thinkingConfig.thinkingBudget: 0 disables, -1 is dynamic, and manual budgets are 1..24576. https://cloud.google.com/vertex-ai/generative-ai/docs/thinking (accessed 2026-06-25)
 temperature = true
 tool_call = true
 structured_output = true
@@ -5,6 +5,7 @@ last_updated = "2025-06-05"
 attachment = true
 reasoning = true
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
+# Native Gemini uses $.generationConfig.thinkingConfig.thinkingBudget: -1 is dynamic and manual budgets are 128..32768; 0/off is unsupported. https://cloud.google.com/vertex-ai/generative-ai/docs/thinking (accessed 2026-06-25)
 temperature = true
 tool_call = true
 structured_output = true
@@ -1,5 +1,9 @@
 name = "AIHubMix"
 npm = "@aihubmix/ai-sdk-provider"
+# Raw Chat: $.reasoning_effort = "none"|"minimal"|"low"|"medium"|"high"|"xhigh"; aliases are $.reasoning.effort and integer $.reasoning.max_tokens. "none" disables models that permit it. https://docs.aihubmix.com/cn/api/unified-inference (accessed 2026-06-25)
+# Raw Responses: $.reasoning.effort carries effort; this endpoint has no reasoning-token budget field. https://docs.aihubmix.com/cn/api-reference/openai-compatible/create-a-model-response (accessed 2026-06-25)
+# Raw Messages: $.thinking.type = "enabled"|"disabled"|"adaptive"; enabled uses $.thinking.budget_tokens >= 1024, and $.output_config.effort = "low"|"medium"|"high"|"xhigh"|"max" subject to model support. https://docs.aihubmix.com/cn/api-reference/anthropic-compatible/create-a-message (accessed 2026-06-25)
+# Raw Gemini native: $.generationConfig.thinkingConfig uses integer thinkingBudget (-1 dynamic; 0 off where supported) or string thinkingLevel; model bounds differ below. https://docs.aihubmix.com/cn/api-reference/google-vertex-ai-compatible/generate-content (accessed 2026-06-25)
 env = ["AIHUBMIX_API_KEY"]
 doc = "https://docs.aihubmix.com"

@@ -1,5 +1,20 @@
 name = "Alibaba (China)"
 env = ["DASHSCOPE_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# OpenAI Chat: POST /compatible-mode/v1/chat/completions; top-level
+# `enable_thinking` is true or false and `thinking_budget` is an integer token
+# cap (Qwen3 in thinking mode and Kimi only; default is the model maximum).
+# Responses: POST /compatible-mode/v1/responses; `reasoning.effort` is none,
+# minimal, low, medium (default), or high; no numeric thinking budget is accepted.
+# Anthropic: POST /apps/anthropic/v1/messages; `thinking.type` is enabled or
+# disabled and `thinking.budget_tokens` is an integer used only when enabled.
+# `reasoning_effort` is high or max only for DeepSeek V4 Pro/Flash.
+# DashScope text: POST /api/v1/services/aigc/text-generation/generation; put
+# `enable_thinking` and `thinking_budget` under `parameters`, not at JSON root.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://help.aliyun.com/zh/model-studio/compatibility-with-openai-responses-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages
 doc = "https://www.alibabacloud.com/help/en/model-studio/models"
 api = "https://dashscope.aliyuncs.com/compatible-mode/v1" 
@@ -1,4 +1,10 @@
 name = "MiniMax-M2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model; Alibaba-deployed MiniMax-M2.5 is
+# thinking-only. No toggle, effort, or numeric budget request field is documented.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/coding-plan
+# https://help.aliyun.com/zh/model-studio/deep-thinking
 family = "minimax"
 release_date = "2026-02-12"
 last_updated = "2026-02-12"
@@ -1,4 +1,10 @@
 name = "GLM-5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. Hybrid thinking is on by default and
+# top-level `enable_thinking`: true|false toggles it. No effort/budget documented.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/coding-plan
+# https://help.aliyun.com/zh/model-studio/deep-thinking
 family = "glm"
 release_date = "2026-02-11"
 last_updated = "2026-02-11"
@@ -1,4 +1,11 @@
 name = "Kimi K2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. Hybrid thinking uses top-level
+# `enable_thinking`: true enables; false (default) disables. No effort or numeric
+# budget bound is documented specifically for the plan endpoint.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/coding-plan
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
 family = "kimi-k2"
 release_date = "2026-01-27"
 last_updated = "2026-01-27"
@@ -1,4 +1,9 @@
 name = "Qwen3 Coder Plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. The plan docs document no reasoning
+# toggle, effort, or numeric budget field for this model.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/coding-plan
 family = "qwen"
 release_date = "2025-07-23"
 last_updated = "2025-07-23"
@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.7-plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. Hybrid thinking is on by default;
+# top-level `enable_thinking`: true|false toggles it. The plan docs do not state
+# a numeric budget bound or a plan-specific effort control.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/coding-plan
+# https://help.aliyun.com/zh/model-studio/deep-thinking
 reasoning_options = [{ type = "toggle" }]

 [cost]
@@ -1,5 +1,15 @@
 name = "Alibaba Coding Plan (China)"
 env = ["ALIBABA_CODING_PLAN_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# OpenAI plan base https://coding.dashscope.aliyuncs.com/v1 uses POST
+# /chat/completions with top-level `enable_thinking`: true or false.
+# Anthropic plan base https://coding.dashscope.aliyuncs.com/apps/anthropic uses
+# POST /v1/messages with `thinking.type`: enabled or disabled and optional
+# integer `thinking.budget_tokens`. No plan-specific effort field is documented.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/coding-plan
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages
 doc = "https://help.aliyun.com/zh/model-studio/coding-plan"
 api = "https://coding.dashscope.aliyuncs.com/v1"
@@ -1,4 +1,10 @@
 name = "MiniMax-M2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model; Alibaba-deployed MiniMax-M2.5 is
+# thinking-only. No toggle, effort, or numeric budget request field is documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/coding-plan
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 family = "minimax"
 release_date = "2026-02-12"
 last_updated = "2026-02-12"
@@ -1,4 +1,10 @@
 name = "GLM-5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. Hybrid thinking is on by default and
+# top-level `enable_thinking`: true|false toggles it. No effort/budget documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/coding-plan
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 family = "glm"
 release_date = "2026-02-11"
 last_updated = "2026-02-11"
@@ -1,4 +1,11 @@
 name = "Kimi K2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. Hybrid thinking uses top-level
+# `enable_thinking`: true enables; false (default) disables. No effort or numeric
+# budget bound is documented specifically for the plan endpoint.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/coding-plan
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
 family = "kimi-k2"
 release_date = "2026-01-27"
 last_updated = "2026-01-27"
@@ -1,4 +1,9 @@
 name = "Qwen3 Coder Plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. The plan docs document no reasoning
+# toggle, effort, or numeric budget field for this model.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/coding-plan
 family = "qwen"
 release_date = "2025-07-23"
 last_updated = "2025-07-23"
@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.7-plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Coding Plan exact-string allowlist model. Hybrid thinking is on by default;
+# top-level `enable_thinking`: true|false toggles it. The plan docs do not state
+# a numeric budget bound or a plan-specific effort control.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/coding-plan
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 reasoning_options = [{ type = "toggle" }]

 [cost]
@@ -1,5 +1,15 @@
 name = "Alibaba Coding Plan"
 env = ["ALIBABA_CODING_PLAN_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# OpenAI plan base https://coding-intl.dashscope.aliyuncs.com/v1 uses POST
+# /chat/completions with top-level `enable_thinking`: true or false.
+# Anthropic plan base https://coding-intl.dashscope.aliyuncs.com/apps/anthropic
+# uses POST /v1/messages with `thinking.type`: enabled or disabled and optional
+# integer `thinking.budget_tokens`. No plan-specific effort field is documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/coding-plan
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages
 doc = "https://www.alibabacloud.com/help/en/model-studio/coding-plan"
 api = "https://coding-intl.dashscope.aliyuncs.com/v1"
@@ -1,4 +1,9 @@
 base_model = "minimax/MiniMax-M2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Alibaba-deployed MiniMax-M2.5 is thinking-only. No toggle, effort, or numeric
+# reasoning-budget request field is documented for this model.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/deep-thinking
 reasoning_options = []

 [interleaved]
@@ -1,4 +1,11 @@
 base_model = "deepseek/deepseek-v4-flash"
+# Reasoning HTTP format (accessed 2026-06-25):
+# This model toggles with `enable_thinking`: true or false. Chat Completions
+# `reasoning_effort` accepts high (default) or max; low/medium map to high and
+# xhigh maps to max. Anthropic Messages defaults `reasoning_effort` to max.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deepseek-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages

 [[reasoning_options]]
 type = "toggle"
@@ -1,4 +1,11 @@
 base_model = "deepseek/deepseek-v4-pro"
+# Reasoning HTTP format (accessed 2026-06-25):
+# This model toggles with `enable_thinking`: true or false. Chat Completions
+# `reasoning_effort` accepts high (default) or max; low/medium map to high and
+# xhigh maps to max. Anthropic Messages defaults `reasoning_effort` to max.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deepseek-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages

 [[reasoning_options]]
 type = "toggle"
@@ -1,4 +1,10 @@
 base_model = "moonshotai/kimi-k2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid model: top-level `enable_thinking` true enables thinking; false is the
+# default. `thinking_budget` is an integer token cap; no bounds are documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 base_model_omit = ["structured_output"]
 family = "kimi-k2"
 attachment = true
@@ -1,4 +1,10 @@
 base_model = "moonshotai/kimi-k2.6"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid model: top-level `enable_thinking` true enables thinking; false is the
+# default. `thinking_budget` is an integer token cap; no bounds are documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 base_model_omit = ["structured_output"]
 family = "kimi-k2"

@@ -1,4 +1,10 @@
 base_model = "moonshotai/kimi-k2.7-code"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Thinking-only model: `enable_thinking` defaults to true and cannot disable
+# thinking. `thinking_budget` is an integer token cap; no bounds are documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 base_model_omit = ["structured_output"]
 family = "kimi-k2"
 reasoning_options = [{ type = "toggle" }, { type = "budget_tokens" }]
@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.6-flash"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 81920; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://help.aliyun.com/zh/model-studio/compatibility-with-openai-responses-api

 [[reasoning_options]]
 type = "toggle"
@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.6-plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 81920; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://help.aliyun.com/zh/model-studio/compatibility-with-openai-responses-api

 [[reasoning_options]]
 type = "toggle"
@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.7-max"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 262144; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://help.aliyun.com/zh/model-studio/compatibility-with-openai-responses-api

 [[reasoning_options]]
 type = "toggle"
@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.7-plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 262144; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://help.aliyun.com/zh/model-studio/compatibility-with-openai-responses-api

 [[reasoning_options]]
 type = "toggle"
@@ -1,5 +1,17 @@
 name = "Alibaba Token Plan (China)"
 env = ["ALIBABA_TOKEN_PLAN_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# OpenAI plan base .../compatible-mode/v1 uses POST /chat/completions with
+# top-level `enable_thinking`: true or false and integer `thinking_budget`.
+# POST /responses uses `reasoning.effort`: none, minimal, low, medium, or high;
+# `thinking_budget` is not accepted there. Anthropic plan base .../apps/anthropic
+# uses POST /v1/messages with `thinking.type`: enabled or disabled, integer
+# `thinking.budget_tokens`, and DeepSeek V4-only `reasoning_effort`: high or max.
+# Sources:
+# https://help.aliyun.com/zh/model-studio/token-plan-quickstart
+# https://help.aliyun.com/zh/model-studio/deep-thinking
+# https://help.aliyun.com/zh/model-studio/compatibility-with-openai-responses-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages
 doc = "https://www.alibabacloud.com/help/zh/model-studio/token-plan-overview"
 api = "https://token-plan.cn-beijing.maas.aliyuncs.com/compatible-mode/v1"
@@ -1,4 +1,9 @@
 base_model = "minimax/MiniMax-M2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Alibaba-deployed MiniMax-M2.5 is thinking-only. No toggle, effort, or numeric
+# reasoning-budget request field is documented for this model.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking

 reasoning_options = []

@@ -1,4 +1,11 @@
 base_model = "deepseek/deepseek-v4-flash"
+# Reasoning HTTP format (accessed 2026-06-25):
+# This model toggles with `enable_thinking`: true or false. Chat Completions
+# `reasoning_effort` accepts high (default) or max; low/medium map to high and
+# xhigh maps to max. Anthropic Messages defaults `reasoning_effort` to max.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deepseek-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages

 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["high", "max"] }]

@@ -1,4 +1,11 @@
 base_model = "deepseek/deepseek-v4-pro"
+# Reasoning HTTP format (accessed 2026-06-25):
+# This model toggles with `enable_thinking`: true or false. Chat Completions
+# `reasoning_effort` accepts high (default) or max; low/medium map to high and
+# xhigh maps to max. Anthropic Messages defaults `reasoning_effort` to max.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deepseek-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages

 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["high", "max"] }]

@@ -1,4 +1,10 @@
 base_model = "moonshotai/kimi-k2.5"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid model: top-level `enable_thinking` true enables thinking; false is the
+# default. `thinking_budget` is an integer token cap; no bounds are documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 base_model_omit = ["structured_output"]
 family = "kimi-k2"
 attachment = true
@@ -1,4 +1,10 @@
 base_model = "moonshotai/kimi-k2.6"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid model: top-level `enable_thinking` true enables thinking; false is the
+# default. `thinking_budget` is an integer token cap; no bounds are documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 base_model_omit = ["structured_output"]
 family = "kimi-k2"
 reasoning_options = [{ type = "toggle" }, { type = "budget_tokens" }]
@@ -1,4 +1,10 @@
 base_model = "moonshotai/kimi-k2.7-code"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Thinking-only model: `enable_thinking` defaults to true and cannot disable
+# thinking. `thinking_budget` is an integer token cap; no bounds are documented.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/kimi-api
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
 base_model_omit = ["structured_output"]
 family = "kimi-k2"
 reasoning_options = [{ type = "toggle" }, { type = "budget_tokens" }]
@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.6-flash"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 81920; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api

 reasoning_options = [{ type = "toggle" }, { type = "budget_tokens", max = 81_920 }]

@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.6-plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 81920; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api

 reasoning_options = [{ type = "toggle" }, { type = "budget_tokens", max = 81_920 }]

@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.7-max"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 262144; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api

 reasoning_options = [{ type = "toggle" }, { type = "budget_tokens", max = 262_144 }]

@@ -1,4 +1,11 @@
 base_model = "alibaba/qwen3.7-plus"
+# Reasoning HTTP format (accessed 2026-06-25):
+# Hybrid, thinking on by default. Chat/DashScope `enable_thinking`: true|false;
+# integer `thinking_budget` maximum 262144; no minimum or disable sentinel is
+# documented. Responses instead uses `reasoning.effort` and no numeric budget.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api

 reasoning_options = [{ type = "toggle" }, { type = "budget_tokens", max = 262_144 }]

@@ -1,5 +1,17 @@
 name = "Alibaba Token Plan"
 env = ["ALIBABA_TOKEN_PLAN_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# OpenAI plan base .../compatible-mode/v1 uses POST /chat/completions with
+# top-level `enable_thinking`: true or false and integer `thinking_budget`.
+# POST /responses uses `reasoning.effort`: none, minimal, low, medium, or high;
+# `thinking_budget` is not accepted there. Anthropic plan base .../apps/anthropic
+# uses POST /v1/messages with `thinking.type`: enabled or disabled, integer
+# `thinking.budget_tokens`, and DeepSeek V4-only `reasoning_effort`: high or max.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/token-plan-quickstart
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages
 doc = "https://www.alibabacloud.com/help/en/model-studio/token-plan-overview"
 api = "https://token-plan.ap-southeast-1.maas.aliyuncs.com/compatible-mode/v1"
@@ -1,5 +1,20 @@
 name = "Alibaba"
 env = ["DASHSCOPE_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Reasoning HTTP format (accessed 2026-06-25):
+# OpenAI Chat: POST /compatible-mode/v1/chat/completions; top-level
+# `enable_thinking` is true or false and `thinking_budget` is an integer token
+# cap (Qwen3 in thinking mode and Kimi only; default is the model maximum).
+# Responses: POST /compatible-mode/v1/responses; `reasoning.effort` is none,
+# minimal, low, medium (default), or high; no numeric thinking budget is accepted.
+# Anthropic: POST /apps/anthropic/v1/messages; `thinking.type` is enabled or
+# disabled and `thinking.budget_tokens` is an integer used only when enabled.
+# `reasoning_effort` is high or max only for DeepSeek V4 Pro/Flash.
+# DashScope text: POST /api/v1/services/aigc/text-generation/generation; put
+# `enable_thinking` and `thinking_budget` under `parameters`, not at JSON root.
+# Sources:
+# https://www.alibabacloud.com/help/en/model-studio/deep-thinking
+# https://www.alibabacloud.com/help/en/model-studio/compatibility-with-openai-responses-api
+# https://www.alibabacloud.com/help/en/model-studio/anthropic-api-messages
 doc = "https://www.alibabacloud.com/help/en/model-studio/models"
 api = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1" 
@@ -1,3 +1,5 @@
+# Ambient documents no toggle, effort, or budget field for this model; reasoning
+# output does not itself establish a control. https://docs.ambient.xyz
 reasoning_options = []
 base_model = "moonshotai/kimi-k2.6"

@@ -1,5 +1,7 @@
 base_model = "zhipuai/glm-5.1"
 name = "GLM 5.1"
+# Ambient documents no toggle, effort, or budget field for this model; reasoning
+# output does not itself establish a control. https://docs.ambient.xyz
 reasoning_options = []

 [interleaved]
@@ -1,5 +1,11 @@
 name = "Ambient"
 env = ["AMBIENT_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Raw HTTP reasoning controls (sources accessed 2026-06-25):
+# POST https://api.ambient.xyz/v1/chat/completions. Ambient's developer
+# surface documents no reasoning toggle, effort, or token-budget request field.
+# A reasoning-capable model listing alone does not establish a usable control.
+# https://ambient.xyz/developers
+# https://docs.ambient.xyz
 api = "https://api.ambient.xyz/v1"
 doc = "https://ambient.xyz"
@@ -1,2 +1,6 @@
 base_model = "anthropic/claude-haiku-4-5"
+# AnyAPI accepts thinking.type="enabled" with budget_tokens, but documents no
+# model-specific 1,024..63,999 bounds or off mapping; these remain upstream facts.
+# https://docs.anyapi.ai/guides/parameters
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }, { type = "budget_tokens", min = 1_024, max = 63_999 }]
@@ -1,4 +1,8 @@
 base_model = "anthropic/claude-opus-4-6"
+# AnyAPI accepts thinking.type="enabled" with budget_tokens, but documents no
+# model-specific 1,024..127,999 bounds or off mapping; these remain upstream facts.
+# https://docs.anyapi.ai/guides/parameters
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }, { type = "budget_tokens", min = 1_024, max = 127_999 }]

 [experimental.modes.fast]
@@ -1,4 +1,7 @@
 base_model = "anthropic/claude-opus-4-7"
+# AnyAPI documents generic effort and thinking fields, but no Opus 4.7-specific
+# mapping or budget bounds; accepted schema does not prove effective control.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }]

 [experimental.modes.fast]
@@ -1,2 +1,6 @@
 base_model = "anthropic/claude-sonnet-4-5"
+# AnyAPI accepts thinking.type="enabled" with budget_tokens, but documents no
+# model-specific 1,024..63,999 bounds or off mapping; these remain upstream facts.
+# https://docs.anyapi.ai/guides/parameters
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }, { type = "budget_tokens", min = 1_024, max = 63_999 }]
@@ -1,2 +1,6 @@
 base_model = "anthropic/claude-sonnet-4-6"
+# AnyAPI accepts thinking.type="enabled" with budget_tokens, but documents no
+# model-specific 1,024..63,999 bounds or off mapping; these remain upstream facts.
+# https://docs.anyapi.ai/guides/parameters
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }, { type = "budget_tokens", min = 1_024, max = 63_999 }]
@@ -1,2 +1,5 @@
 base_model = "openai/gpt-5-mini"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
@@ -1,2 +1,5 @@
 base_model = "openai/gpt-5.1"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
@@ -1,2 +1,5 @@
 base_model = "openai/gpt-5.2"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
@@ -1,4 +1,7 @@
 base_model = "openai/gpt-5.4"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]

 [experimental.modes.fast]
@@ -1,2 +1,5 @@
 base_model = "openai/gpt-5"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
@@ -1,2 +1,5 @@
 base_model = "openai/o3-mini"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
@@ -1,2 +1,5 @@
 base_model = "openai/o3"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
@@ -1,2 +1,5 @@
 base_model = "openai/o4-mini"
+# AnyAPI accepts low/medium/high generically but documents no model-specific
+# effectiveness; unsupported parameters may be silently ignored.
+# https://docs.anyapi.ai/guides/parameters
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]
@@ -1,5 +1,13 @@
 name = "AnyAPI"
 npm = "@ai-sdk/openai-compatible"
+# Raw HTTP reasoning controls (sources accessed 2026-06-25):
+# POST https://api.anyapi.ai/v1/chat/completions accepts top-level
+# reasoning_effort = "low" | "medium" | "high" and Anthropic-style
+# thinking = { type = "enabled", budget_tokens = <integer> }. No budget bounds
+# or explicit disabled value are documented. Unsupported model parameters may
+# be silently ignored, so schema acceptance is not meaningful model support.
+# https://docs.anyapi.ai/guides/parameters
+# https://docs.anyapi.ai/openapi.json
 env = ["ANYAPI_API_KEY"]
 api = "https://api.anyapi.ai/v1"
 doc = "https://docs.anyapi.ai"
@@ -1,5 +1,13 @@
 name = "Atomic Chat"
 env = ["ATOMIC_CHAT_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Raw HTTP reasoning controls (sources accessed 2026-06-25):
+# POST http://127.0.0.1:1337/v1/chat/completions. Current source accepts
+# chat_template_kwargs.enable_thinking = true | false for template-supported
+# models. The former process-level --reasoning-budget accepted -1 (unlimited)
+# or 0 (disabled), but was removed; no effort or token-budget HTTP field is
+# documented. Request-schema acceptance still depends on the loaded backend.
+# https://github.com/AtomicBot-ai/Atomic-Chat/commit/92703bceb2c65e3218f81a15b4e9058f171d4ff2
+# https://github.com/AtomicBot-ai/Atomic-Chat/commit/742e731e966b1b59cedad36301e32fc7613e4bce
 api = "http://127.0.0.1:1337/v1"
 doc = "https://atomic.chat"
@@ -1,5 +1,8 @@
 base_model = "anthropic/claude-opus-4-6"

+# Claude 4.6 uses adaptive thinking; Auriko maps effort and "off" to its native
+# control. No caller-selected token budget is exposed.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "max"] }]

 [cost]
@@ -1,5 +1,8 @@
 base_model = "anthropic/claude-opus-4-7"

+# Auriko's schema accepts these levels and "off", but the provider support table
+# does not yet document an Opus 4.7-specific translation.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]

 [cost]
@@ -1,5 +1,8 @@
 base_model = "anthropic/claude-sonnet-4-6"

+# Claude 4.6 uses adaptive thinking; Auriko maps effort and "off" to its native
+# control. No caller-selected token budget is exposed.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "max"] }]

 [cost]
@@ -1,5 +1,8 @@
 base_model = "deepseek/deepseek-v4-flash"

+# Also aliased as deepseek-chat and deepseek-reasoner; aliases are not modes.
+# Effort controls a derived budget and "off" requests non-thinking output.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]

 [interleaved]
@@ -1,5 +1,7 @@
 base_model = "deepseek/deepseek-v4-pro"

+# Auriko derives a provider thinking budget from effort; "off" disables it.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]

 [interleaved]
@@ -1,5 +1,8 @@
 base_model = "google/gemini-2.5-flash"

+# Auriko derives a Gemini thinking budget from effort; "off" disables thinking.
+# Provider extension keys google/google_ai/googleai/gemini are aliases.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]

 [cost]
@@ -1,5 +1,8 @@
 base_model = "google/gemini-2.5-pro"

+# Auriko derives a Gemini thinking budget from effort. Although schema-level
+# "off" exists, this model metadata does not claim a toggle.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]

 [cost]
@@ -1,5 +1,8 @@
 base_model = "google/gemini-3.1-pro-preview"

+# Gemini 3.x maps only low/medium/high to thinking levels; xhigh/max normalize
+# to high. Provider extension keys google/google_ai/googleai/gemini are aliases.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }]

 [cost]
@@ -1,5 +1,8 @@
 base_model = "zhipuai/glm-5.1"

+# Auriko's schema accepts reasoning_effort, but its provider support table gives
+# no GLM 5.1 mapping; no meaningful toggle, effort, or budget is documented.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = []

 [interleaved]
@@ -1,5 +1,8 @@
 base_model = "xai/grok-4.3"

+# Grok 4.3 meaningfully supports native low/medium/high; xhigh/max normalize to
+# high, and "off" is Auriko's normalized disable value.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high"] }]

 [cost]
@@ -1,5 +1,7 @@
 base_model = "moonshotai/kimi-k2.5"

+# Auriko maps effort to Moonshot's native control; "off" disables thinking.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]

 [interleaved]
@@ -1,5 +1,7 @@
 base_model = "moonshotai/kimi-k2.6"

+# Auriko maps effort to Moonshot's native control; "off" disables thinking.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = [{ type = "toggle" }, { type = "effort", values = ["low", "medium", "high", "xhigh", "max"] }]

 [interleaved]
@@ -1,5 +1,8 @@
 base_model = "minimax/MiniMax-M2.7-highspeed"

+# Auriko documents M2-series reasoning as built in and drops reasoning_effort;
+# schema acceptance is therefore not a meaningful control for this model.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = []

 [cost]
@@ -1,5 +1,8 @@
 base_model = "minimax/MiniMax-M2.7"

+# Auriko documents M2-series reasoning as built in and drops reasoning_effort;
+# schema acceptance is therefore not a meaningful control for this model.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = []

 [cost]
@@ -1,5 +1,8 @@
 base_model = "alibaba/qwen3.6-plus"

+# Auriko's schema accepts reasoning_effort, but its provider support table gives
+# no Qwen 3.6 mapping; no meaningful toggle, effort, or budget is documented.
+# https://docs.auriko.ai/guides/extensions-and-thinking
 reasoning_options = []

 [cost]
@@ -1,5 +1,14 @@
 name = "Auriko"
 env = ["AURIKO_API_KEY"]
 npm = "@ai-sdk/openai-compatible"
+# Raw HTTP reasoning controls (sources accessed 2026-06-25):
+# POST https://api.auriko.ai/v1/chat/completions accepts top-level
+# reasoning_effort = "low" | "medium" | "high" | "xhigh" | "max" | "off";
+# "off" disables thinking. Auriko translates values per model/provider and may
+# clamp them with a warning. No caller-specified reasoning-token budget exists;
+# translated budgets are internal, and budget fields in extensions are replaced.
+# Provider extension aliases google, google_ai, googleai, and gemini are equal.
+# https://docs.auriko.ai/guides/extensions-and-thinking
+# https://docs.auriko.ai/api-reference/chat-completions
 api = "https://api.auriko.ai/v1"
 doc = "https://docs.auriko.ai"
@@ -4,6 +4,9 @@ release_date = "2025-11-18"
 last_updated = "2025-11-18"
 attachment = true
 reasoning = true
+# POST /anthropic/v1/messages uses thinking.type="enabled" and integer
+# thinking.budget_tokens >=1024; budget must be below max_tokens.
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "budget_tokens", min = 1_024 }]
 temperature = true
 knowledge = "2025-02-31"
@@ -4,6 +4,9 @@ release_date = "2025-11-18"
 last_updated = "2025-11-18"
 attachment = true
 reasoning = true
+# POST /anthropic/v1/messages uses thinking.type="enabled" and integer
+# thinking.budget_tokens >=1024; budget must be below max_tokens.
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "budget_tokens", min = 1_024 }]
 temperature = true
 knowledge = "2025-03-31"
@@ -4,6 +4,9 @@ release_date = "2025-11-24"
 last_updated = "2025-08-01"
 attachment = true
 reasoning = true
+# POST /anthropic/v1/messages accepts thinking.budget_tokens >=1024 and effort
+# low/medium/high for this model; a raw budget must remain below max_tokens.
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high"] }, { type = "budget_tokens", min = 1_024 }]
 temperature = true
 tool_call = true
@@ -4,6 +4,9 @@ release_date = "2026-02-05"
 last_updated = "2026-02-05"
 attachment = true
 reasoning = true
+# POST /anthropic/v1/messages supports adaptive thinking with output_config.effort
+# low/medium/high/max; manual thinking.budget_tokens is integer >=1024.
+# https://docs.anthropic.com/en/docs/build-with-claude/adaptive-thinking
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high", "max"] }, { type = "budget_tokens", min = 1_024 }]
 temperature = true
 tool_call = true
@@ -1,6 +1,9 @@
 base_model = "anthropic/claude-opus-4-8"
 temperature = true
 knowledge = "2025-12-31"
+# POST /anthropic/v1/messages uses adaptive thinking with output_config.effort
+# low/medium/high/max; Azure catalog availability alone is not a control.
+# https://docs.anthropic.com/en/docs/build-with-claude/adaptive-thinking
 reasoning_options = [{ type = "effort", values = ["low", "medium", "high", "max"] }]

 [cost]
@@ -4,6 +4,9 @@ release_date = "2025-11-18"
 last_updated = "2025-11-18"
 attachment = true
 reasoning = true
+# POST /anthropic/v1/messages uses thinking.type="enabled" and integer
+# thinking.budget_tokens >=1024; budget must be below max_tokens.
+# https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking
 reasoning_options = [{ type = "budget_tokens", min = 1_024 }]
 temperature = true
 knowledge = "2025-07-31"
@@ -1,4 +1,7 @@
 base_model = "openai/gpt-5.4-mini"
+# Azure OpenAI Chat uses top-level reasoning_effort; this model accepts
+# none/low/medium/high/xhigh. No separate token-budget field is documented.
+# https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning
 reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
 name = "GPT-5.4 Mini"

@@ -1,4 +1,7 @@
 base_model = "openai/gpt-5.4-nano"
+# Azure OpenAI Chat uses top-level reasoning_effort; this model accepts
+# none/low/medium/high/xhigh. No separate token-budget field is documented.
+# https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning
 reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]
 name = "GPT-5.4 Nano"

@@ -1,4 +1,7 @@
 base_model = "openai/gpt-5.4-pro"
+# Azure documents this Responses-only model's reasoning.effort subset as
+# medium/high/xhigh; no toggle or reasoning-token budget field is exposed.
+# https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning
 reasoning_options = [{ type = "effort", values = ["medium", "high", "xhigh"] }]

 [cost]
@@ -1,4 +1,7 @@
 base_model = "openai/gpt-5.4"
+# Azure OpenAI Chat uses top-level reasoning_effort; this model accepts
+# none/low/medium/high/xhigh. No separate token-budget field is documented.
+# https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning
 reasoning_options = [{ type = "effort", values = ["none", "low", "medium", "high", "xhigh"] }]

 [cost]
@@ -1,4 +1,7 @@
 base_model = "moonshotai/kimi-k2.5"
+# Azure documents POST /models/chat/completions and reasoning output, but no
+# toggle, effort, or budget request field for Kimi K2.5. Catalog modes are not
+# request controls. https://learn.microsoft.com/en-us/azure/ai-foundry/foundry-models/how-to/use-chat-reasoning
 reasoning_options = [{ type = "toggle" }]
 temperature = true
 interleaved = true
@@ -1,5 +1,8 @@
 base_model = "moonshotai/kimi-k2.6"
 attachment = false
+# Azure documents POST /models/chat/completions and reasoning output, but no
+# toggle, effort, or budget request field for Kimi K2.6. Catalog modes are not
+# request controls. https://learn.microsoft.com/en-us/azure/ai-foundry/foundry-models/how-to/use-chat-reasoning
 reasoning_options = [{ type = "toggle" }]
 interleaved = true

@@ -1,4 +1,12 @@
 name = "Azure Cognitive Services"
 env = ["AZURE_COGNITIVE_SERVICES_RESOURCE_NAME", "AZURE_COGNITIVE_SERVICES_API_KEY"]
 npm = "@ai-sdk/azure"
+# Raw HTTP reasoning controls (sources accessed 2026-06-25):
+# Azure OpenAI Chat: POST https://<resource>.openai.azure.com/openai/v1/chat/completions
+# uses top-level reasoning_effort; model-specific values are documented below.
+# Anthropic Messages: POST https://<resource>.services.ai.azure.com/anthropic/v1/messages
+# uses thinking.type and thinking.budget_tokens. Azure model "modes" alone are
+# not request controls. Azure documents no common toggle or budget across APIs.
+# https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/reasoning
+# https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/models-sold-directly-by-azure
 doc = "https://learn.microsoft.com/en-us/azure/ai-services/openai/concepts/models"
@@ -1,3 +1,6 @@
+# Direct endpoint: POST /models/chat/completions?api-version=2024-05-01-preview.
+# Its published request schema documents no toggle, effort, or reasoning budget.
+# https://learn.microsoft.com/en-us/rest/api/microsoftfoundry/model-inference/get-chat-completions/get-chat-completions (accessed 2026-06-25)
 base_model = "deepseek/deepseek-v4-flash"
 name = "DeepSeek-V4-Flash"
 tool_call = false
@@ -1,3 +1,6 @@
+# Direct endpoint: POST /models/chat/completions?api-version=2024-05-01-preview.
+# Its published request schema documents no toggle, effort, or reasoning budget.
+# https://learn.microsoft.com/en-us/rest/api/microsoftfoundry/model-inference/get-chat-completions/get-chat-completions (accessed 2026-06-25)
 base_model = "deepseek/deepseek-v4-pro"
 name = "DeepSeek-V4-Pro"
 tool_call = false
@@ -1,3 +1,6 @@
+# Direct endpoint: POST /models/chat/completions?api-version=2024-05-01-preview.
+# Its published request schema documents no toggle, effort, or reasoning budget.
+# https://learn.microsoft.com/en-us/rest/api/microsoftfoundry/model-inference/get-chat-completions/get-chat-completions (accessed 2026-06-25)
 name = "Kimi K2.5"
 family = "kimi-k2"
 release_date = "2026-02-06"
@@ -1,3 +1,6 @@
+# Direct endpoint: POST /models/chat/completions?api-version=2024-05-01-preview.
+# Its published request schema documents no toggle, effort, or reasoning budget.
+# https://learn.microsoft.com/en-us/rest/api/microsoftfoundry/model-inference/get-chat-completions/get-chat-completions (accessed 2026-06-25)
 name = "Kimi K2.6"
 family = "kimi-k2"
 release_date = "2026-04-22"
@@ -4,6 +4,8 @@ release_date = "2025-10"
 last_updated = "2025-10"
 attachment = false
 reasoning = true
+# The Bailing request surface documents no toggle, effort, or budget field for
+# Ring-1T. https://alipaytbox.yuque.com/sxs0ba/ling/intro
 reasoning_options = []
 temperature = true
 knowledge = "2024-06"
@@ -1,5 +1,10 @@
 name = "Bailing"
 env = ["BAILING_API_TOKEN"]
 npm = "@ai-sdk/openai-compatible"
+# Raw HTTP reasoning controls (source accessed 2026-06-25):
+# POST https://api.tbox.cn/api/llm/v1/chat/completions. The provider request
+# surface documents no reasoning toggle, effort, or token-budget request field;
+# Ring-1T being a reasoning model does not by itself establish such a control.
+# https://alipaytbox.yuque.com/sxs0ba/ling/intro
 doc = "https://alipaytbox.yuque.com/sxs0ba/ling/intro"
 api = "https://api.tbox.cn/api/llm/v1/chat/completions"
@@ -5,6 +5,9 @@ release_date = "2026-02-12"
 last_updated = "2026-02-12"
 attachment = false
 reasoning = true
+# This deprecated model is absent from Baseten's current reasoning support table;
+# no toggle, effort, or budget request field is documented for it.
+# https://docs.baseten.co/inference/model-apis/reasoning
 reasoning_options = []
 temperature = true
 tool_call = true
--- a/Show More
+++ b/Show More