fix: add verbosity parameter support for GPT-5 models across legacy and modern AI SDK (#11281)

* Initial plan

* feat: add verbosity parameter support for GPT-5 models in OpenAIAPIClient

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

* fix: ensure gpt-5-pro always uses 'high' verbosity

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

* refactor: move verbosity configuration to config/models as suggested

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

* refactor: encapsulate verbosity logic in getVerbosity method

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

* feat: add support for verbosity and reasoning options for GPT-5 Pro and GPT-5.1 models

* fix comment

* build: add @ai-sdk/google dependency

Add the @ai-sdk/google package to support Google AI SDK integration

* build: add @ai-sdk/anthropic dependency

* refactor(aiCore): update reasoning params handling for AI providers

- Add type imports for provider options
- Handle 'none' reasoning effort consistently across providers
- Improve type safety by using Pick with provider options
- Standardize disabled reasoning config for all providers

* fix: adjust none effort ratio from 0 to 0.01

Prevent potential division by zero errors by ensuring none effort ratio has a small positive value

* feat(reasoning): add support for GPT-5.1 series models

Handle 'none' reasoning effort for GPT-5.1 models and add model type check

* Update src/renderer/src/aiCore/utils/reasoning.ts

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>
Co-authored-by: suyao <sy20010504@gmail.com>
Co-authored-by: icarus <eurfelux@gmail.com>

package.json

@@ -108,8 +108,10 @@
     "@agentic/searxng": "^7.3.3",
     "@agentic/tavily": "^7.3.3",
     "@ai-sdk/amazon-bedrock": "^3.0.53",
+    "@ai-sdk/anthropic": "^2.0.44",
     "@ai-sdk/cerebras": "^1.0.31",
     "@ai-sdk/gateway": "^2.0.9",
+    "@ai-sdk/google": "^2.0.32",
     "@ai-sdk/google-vertex": "^3.0.62",
     "@ai-sdk/huggingface": "patch:@ai-sdk/huggingface@npm%3A0.0.8#~/.yarn/patches/@ai-sdk-huggingface-npm-0.0.8-d4d0aaac93.patch",
     "@ai-sdk/mistral": "^2.0.23",

src/renderer/src/aiCore/legacy/clients/BaseApiClient.ts

@@ -1,5 +1,6 @@
 import { loggerService } from '@logger'
 import {
+  getModelSupportedVerbosity,
   isFunctionCallingModel,
   isNotSupportTemperatureAndTopP,
   isOpenAIModel,
@@ -242,12 +243,18 @@ export abstract class BaseApiClient<
     return serviceTierSetting
   }
 
-  protected getVerbosity(): OpenAIVerbosity {
+  protected getVerbosity(model?: Model): OpenAIVerbosity {
     try {
       const state = window.store?.getState()
       const verbosity = state?.settings?.openAI?.verbosity
 
       if (verbosity && ['low', 'medium', 'high'].includes(verbosity)) {
+        // If model is provided, check if the verbosity is supported by the model
+        if (model) {
+          const supportedVerbosity = getModelSupportedVerbosity(model)
+          // Use user's verbosity if supported, otherwise use the first supported option
+          return supportedVerbosity.includes(verbosity) ? verbosity : supportedVerbosity[0]
+        }
         return verbosity
       }
     } catch (error) {

src/renderer/src/aiCore/legacy/clients/openai/OpenAIApiClient.ts

@@ -35,6 +35,7 @@ import {
   isSupportedThinkingTokenModel,
   isSupportedThinkingTokenQwenModel,
   isSupportedThinkingTokenZhipuModel,
+  isSupportVerbosityModel,
   isVisionModel,
   MODEL_SUPPORTED_REASONING_EFFORT,
   ZHIPU_RESULT_TOKENS
@@ -733,6 +734,13 @@ export class OpenAIAPIClient extends OpenAIBaseClient<
           ...modalities,
           // groq 有不同的 service tier 配置，不符合 openai 接口类型
           service_tier: this.getServiceTier(model) as OpenAIServiceTier,
+          ...(isSupportVerbosityModel(model)
+            ? {
+                text: {
+                  verbosity: this.getVerbosity(model)
+                }
+              }
+            : {}),
           ...this.getProviderSpecificParameters(assistant, model),
           ...reasoningEffort,
           ...getOpenAIWebSearchParams(model, enableWebSearch),

src/renderer/src/aiCore/legacy/clients/openai/OpenAIResponseAPIClient.ts

@@ -520,7 +520,7 @@ export class OpenAIResponseAPIClient extends OpenAIBaseClient<
           ...(isSupportVerbosityModel(model)
             ? {
                 text: {
-                  verbosity: this.getVerbosity()
+                  verbosity: this.getVerbosity(model)
                 }
               }
             : {}),

src/renderer/src/aiCore/utils/options.ts

@@ -1,6 +1,12 @@
 import { baseProviderIdSchema, customProviderIdSchema } from '@cherrystudio/ai-core/provider'
 import { loggerService } from '@logger'
-import { isOpenAIModel, isQwenMTModel, isSupportFlexServiceTierModel } from '@renderer/config/models'
+import {
+  getModelSupportedVerbosity,
+  isOpenAIModel,
+  isQwenMTModel,
+  isSupportFlexServiceTierModel,
+  isSupportVerbosityModel
+} from '@renderer/config/models'
 import { isSupportServiceTierProvider } from '@renderer/config/providers'
 import { mapLanguageToQwenMTModel } from '@renderer/config/translate'
 import type { Assistant, Model, Provider } from '@renderer/types'
@@ -191,6 +197,23 @@ function buildOpenAIProviderOptions(
       ...reasoningParams
     }
   }
+
+  if (isSupportVerbosityModel(model)) {
+    const state = window.store?.getState()
+    const userVerbosity = state?.settings?.openAI?.verbosity
+
+    if (userVerbosity && ['low', 'medium', 'high'].includes(userVerbosity)) {
+      const supportedVerbosity = getModelSupportedVerbosity(model)
+      // Use user's verbosity if supported, otherwise use the first supported option
+      const verbosity = supportedVerbosity.includes(userVerbosity) ? userVerbosity : supportedVerbosity[0]
+
+      providerOptions = {
+        ...providerOptions,
+        textVerbosity: verbosity
+      }
+    }
+  }
+
   return providerOptions
 }
 

src/renderer/src/aiCore/utils/reasoning.ts

@@ -1,3 +1,7 @@
+import type { BedrockProviderOptions } from '@ai-sdk/amazon-bedrock'
+import type { AnthropicProviderOptions } from '@ai-sdk/anthropic'
+import type { GoogleGenerativeAIProviderOptions } from '@ai-sdk/google'
+import type { XaiProviderOptions } from '@ai-sdk/xai'
 import { loggerService } from '@logger'
 import { DEFAULT_MAX_TOKENS } from '@renderer/config/constant'
 import {
@@ -7,6 +11,7 @@ import {
   isDeepSeekHybridInferenceModel,
   isDoubaoSeedAfter251015,
   isDoubaoThinkingAutoModel,
+  isGPT51SeriesModel,
   isGrok4FastReasoningModel,
   isGrokReasoningModel,
   isOpenAIDeepResearchModel,
@@ -56,13 +61,20 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
   }
   const reasoningEffort = assistant?.settings?.reasoning_effort
 
-  if (!reasoningEffort) {
+  // Handle undefined and 'none' reasoningEffort.
+  // TODO: They should be separated.
+  if (!reasoningEffort || reasoningEffort === 'none') {
     // openrouter: use reasoning
     if (model.provider === SystemProviderIds.openrouter) {
       // Don't disable reasoning for Gemini models that support thinking tokens
       if (isSupportedThinkingTokenGeminiModel(model) && !GEMINI_FLASH_MODEL_REGEX.test(model.id)) {
         return {}
       }
+      // 'none' is not an available value for effort for now.
+      // I think they should resolve this issue soon, so I'll just go ahead and use this value.
+      if (isGPT51SeriesModel(model) && reasoningEffort === 'none') {
+        return { reasoning: { effort: 'none' } }
+      }
       // Don't disable reasoning for models that require it
       if (
         isGrokReasoningModel(model) ||
@@ -117,6 +129,13 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
       return { thinking: { type: 'disabled' } }
     }
 
+    // Specially for GPT-5.1. Suppose this is a OpenAI Compatible provider
+    if (isGPT51SeriesModel(model) && reasoningEffort === 'none') {
+      return {
+        reasoningEffort: 'none'
+      }
+    }
+
     return {}
   }
 
@@ -371,7 +390,7 @@ export function getOpenAIReasoningParams(assistant: Assistant, model: Model): Re
 
 export function getAnthropicThinkingBudget(assistant: Assistant, model: Model): number {
   const { maxTokens, reasoning_effort: reasoningEffort } = getAssistantSettings(assistant)
-  if (reasoningEffort === undefined) {
+  if (reasoningEffort === undefined || reasoningEffort === 'none') {
     return 0
   }
   const effortRatio = EFFORT_RATIO[reasoningEffort]
@@ -393,14 +412,17 @@ export function getAnthropicThinkingBudget(assistant: Assistant, model: Model):
  * 获取 Anthropic 推理参数
  * 从 AnthropicAPIClient 中提取的逻辑
  */
-export function getAnthropicReasoningParams(assistant: Assistant, model: Model): Record<string, any> {
+export function getAnthropicReasoningParams(
+  assistant: Assistant,
+  model: Model
+): Pick<AnthropicProviderOptions, 'thinking'> {
   if (!isReasoningModel(model)) {
     return {}
   }
 
   const reasoningEffort = assistant?.settings?.reasoning_effort
 
-  if (reasoningEffort === undefined) {
+  if (reasoningEffort === undefined || reasoningEffort === 'none') {
     return {
       thinking: {
         type: 'disabled'
@@ -429,7 +451,10 @@ export function getAnthropicReasoningParams(assistant: Assistant, model: Model):
  * 注意：Gemini/GCP 端点所使用的 thinkingBudget 等参数应该按照驼峰命名法传递
  * 而在 Google 官方提供的 OpenAI 兼容端点中则使用蛇形命名法 thinking_budget
  */
-export function getGeminiReasoningParams(assistant: Assistant, model: Model): Record<string, any> {
+export function getGeminiReasoningParams(
+  assistant: Assistant,
+  model: Model
+): Pick<GoogleGenerativeAIProviderOptions, 'thinkingConfig'> {
   if (!isReasoningModel(model)) {
     return {}
   }
@@ -438,7 +463,7 @@ export function getGeminiReasoningParams(assistant: Assistant, model: Model): Re
 
   // Gemini 推理参数
   if (isSupportedThinkingTokenGeminiModel(model)) {
-    if (reasoningEffort === undefined) {
+    if (reasoningEffort === undefined || reasoningEffort === 'none') {
       return {
         thinkingConfig: {
           includeThoughts: false,
@@ -478,27 +503,35 @@ export function getGeminiReasoningParams(assistant: Assistant, model: Model): Re
  * @param model - The model being used
  * @returns XAI-specific reasoning parameters
  */
-export function getXAIReasoningParams(assistant: Assistant, model: Model): Record<string, any> {
+export function getXAIReasoningParams(assistant: Assistant, model: Model): Pick<XaiProviderOptions, 'reasoningEffort'> {
   if (!isSupportedReasoningEffortGrokModel(model)) {
     return {}
   }
 
   const { reasoning_effort: reasoningEffort } = getAssistantSettings(assistant)
 
-  if (!reasoningEffort) {
+  if (!reasoningEffort || reasoningEffort === 'none') {
     return {}
   }
 
-  // For XAI provider Grok models, use reasoningEffort parameter directly
+  switch (reasoningEffort) {
-  return {
+    case 'auto':
-    reasoningEffort
+    case 'minimal':
+    case 'medium':
+      return { reasoningEffort: 'low' }
+    case 'low':
+    case 'high':
+      return { reasoningEffort }
   }
 }
 
 /**
  * Get Bedrock reasoning parameters
  */
-export function getBedrockReasoningParams(assistant: Assistant, model: Model): Record<string, any> {
+export function getBedrockReasoningParams(
+  assistant: Assistant,
+  model: Model
+): Pick<BedrockProviderOptions, 'reasoningConfig'> {
   if (!isReasoningModel(model)) {
     return {}
   }
@@ -509,6 +542,14 @@ export function getBedrockReasoningParams(assistant: Assistant, model: Model): R
     return {}
   }
 
+  if (reasoningEffort === 'none') {
+    return {
+      reasoningConfig: {
+        type: 'disabled'
+      }
+    }
+  }
+
   // Only apply thinking budget for Claude reasoning models
   if (!isSupportedThinkingTokenClaudeModel(model)) {
     return {}

src/renderer/src/config/models/reasoning.ts

@@ -8,7 +8,7 @@ import type {
 import { getLowerBaseModelName, isUserSelectedModelType } from '@renderer/utils'
 
 import { isEmbeddingModel, isRerankModel } from './embedding'
-import { isGPT5SeriesModel, isGPT51SeriesModel } from './utils'
+import { isGPT5ProModel, isGPT5SeriesModel, isGPT51SeriesModel } from './utils'
 import { isTextToImageModel } from './vision'
 import { GEMINI_FLASH_MODEL_REGEX, isOpenAIDeepResearchModel } from './websearch'
 
@@ -26,6 +26,7 @@ export const MODEL_SUPPORTED_REASONING_EFFORT: ReasoningEffortConfig = {
   gpt5_codex: ['low', 'medium', 'high'] as const,
   gpt5_1: ['none', 'low', 'medium', 'high'] as const,
   gpt5_1_codex: ['none', 'medium', 'high'] as const,
+  gpt5pro: ['high'] as const,
   grok: ['low', 'high'] as const,
   grok4_fast: ['auto'] as const,
   gemini: ['low', 'medium', 'high', 'auto'] as const,
@@ -47,6 +48,7 @@ export const MODEL_SUPPORTED_OPTIONS: ThinkingOptionConfig = {
   o: MODEL_SUPPORTED_REASONING_EFFORT.o,
   openai_deep_research: MODEL_SUPPORTED_REASONING_EFFORT.openai_deep_research,
   gpt5: [...MODEL_SUPPORTED_REASONING_EFFORT.gpt5] as const,
+  gpt5pro: MODEL_SUPPORTED_REASONING_EFFORT.gpt5pro,
   gpt5_codex: MODEL_SUPPORTED_REASONING_EFFORT.gpt5_codex,
   gpt5_1: MODEL_SUPPORTED_REASONING_EFFORT.gpt5_1,
   gpt5_1_codex: MODEL_SUPPORTED_REASONING_EFFORT.gpt5_1_codex,
@@ -90,6 +92,9 @@ const _getThinkModelType = (model: Model): ThinkingModelType => {
       thinkingModelType = 'gpt5_codex'
     } else {
       thinkingModelType = 'gpt5'
+      if (isGPT5ProModel(model)) {
+        thinkingModelType = 'gpt5pro'
+      }
     }
   } else if (isSupportedReasoningEffortOpenAIModel(model)) {
     thinkingModelType = 'o'

src/renderer/src/config/models/utils.ts

@@ -240,6 +240,21 @@ export const isGPT51SeriesModel = (model: Model) => {
   return modelId.includes('gpt-5.1')
 }
 
+// GPT-5 verbosity configuration
+// gpt-5-pro only supports 'high', other GPT-5 models support all levels
+export const MODEL_SUPPORTED_VERBOSITY: Record<string, ('low' | 'medium' | 'high')[]> = {
+  'gpt-5-pro': ['high'],
+  default: ['low', 'medium', 'high']
+}
+
+export const getModelSupportedVerbosity = (model: Model): ('low' | 'medium' | 'high')[] => {
+  const modelId = getLowerBaseModelName(model.id)
+  if (modelId.includes('gpt-5-pro')) {
+    return MODEL_SUPPORTED_VERBOSITY['gpt-5-pro']
+  }
+  return MODEL_SUPPORTED_VERBOSITY.default
+}
+
 export const isGeminiModel = (model: Model) => {
   const modelId = getLowerBaseModelName(model.id)
   return modelId.includes('gemini')
@@ -256,3 +271,8 @@ export const ZHIPU_RESULT_TOKENS = ['<|begin_of_box|>', '<|end_of_box|>'] as con
 export const agentModelFilter = (model: Model): boolean => {
   return !isEmbeddingModel(model) && !isRerankModel(model) && !isTextToImageModel(model)
 }
+
+export const isGPT5ProModel = (model: Model) => {
+  const modelId = getLowerBaseModelName(model.id)
+  return modelId.includes('gpt-5-pro')
+}

src/renderer/src/pages/home/Tabs/components/OpenAISettingsGroup.tsx

@@ -1,5 +1,6 @@
 import Selector from '@renderer/components/Selector'
 import {
+  getModelSupportedVerbosity,
   isSupportedReasoningEffortOpenAIModel,
   isSupportFlexServiceTierModel,
   isSupportVerbosityModel
@@ -80,20 +81,24 @@ const OpenAISettingsGroup: FC<Props> = ({ model, providerId, SettingGroup, Setti
     }
   ]
 
-  const verbosityOptions = [
+  const verbosityOptions = useMemo(() => {
-    {
+    const allOptions = [
-      value: 'low',
+      {
-      label: t('settings.openai.verbosity.low')
+        value: 'low',
-    },
+        label: t('settings.openai.verbosity.low')
-    {
+      },
-      value: 'medium',
+      {
-      label: t('settings.openai.verbosity.medium')
+        value: 'medium',
-    },
+        label: t('settings.openai.verbosity.medium')
-    {
+      },
-      value: 'high',
+      {
-      label: t('settings.openai.verbosity.high')
+        value: 'high',
-    }
+        label: t('settings.openai.verbosity.high')
-  ]
+      }
+    ]
+    const supportedVerbosityLevels = getModelSupportedVerbosity(model)
+    return allOptions.filter((option) => supportedVerbosityLevels.includes(option.value as any))
+  }, [model, t])
 
   const serviceTierOptions = useMemo(() => {
     let baseOptions: { value: ServiceTier; label: string }[]
@@ -155,6 +160,15 @@ const OpenAISettingsGroup: FC<Props> = ({ model, providerId, SettingGroup, Setti
     }
   }, [provider.id, serviceTierMode, serviceTierOptions, setServiceTierMode])
 
+  useEffect(() => {
+    if (verbosity && !verbosityOptions.some((option) => option.value === verbosity)) {
+      const supportedVerbosityLevels = getModelSupportedVerbosity(model)
+      // Default to the highest supported verbosity level
+      const defaultVerbosity = supportedVerbosityLevels[supportedVerbosityLevels.length - 1]
+      setVerbosity(defaultVerbosity)
+    }
+  }, [model, verbosity, verbosityOptions, setVerbosity])
+
   if (!isOpenAIReasoning && !isSupportServiceTier && !isSupportVerbosity) {
     return null
   }

src/renderer/src/types/index.ts

@@ -86,6 +86,7 @@ const ThinkModelTypes = [
   'gpt5_1',
   'gpt5_codex',
   'gpt5_1_codex',
+  'gpt5pro',
   'grok',
   'grok4_fast',
   'gemini',
@@ -113,7 +114,7 @@ export function isThinkModelType(type: string): type is ThinkingModelType {
 }
 
 export const EFFORT_RATIO: EffortRatio = {
-  none: 0,
+  none: 0.01,
   minimal: 0.05,
   low: 0.05,
   medium: 0.5,

src/renderer/src/types/sdk.ts

@@ -126,6 +126,10 @@ export type OpenAIExtraBody = {
     source_lang: 'auto'
     target_lang: string
   }
+  // for gpt-5 series models verbosity control
+  text?: {
+    verbosity?: 'low' | 'medium' | 'high'
+  }
 }
 // image is for openrouter. audio is ignored for now
 export type OpenAIModality = OpenAI.ChatCompletionModality | 'image'

yarn.lock

@@ -102,7 +102,7 @@ __metadata:
   languageName: node
   linkType: hard
 
-"@ai-sdk/anthropic@npm:2.0.44":
+"@ai-sdk/anthropic@npm:2.0.44, @ai-sdk/anthropic@npm:^2.0.44":
   version: 2.0.44
   resolution: "@ai-sdk/anthropic@npm:2.0.44"
   dependencies:
@@ -206,6 +206,18 @@ __metadata:
   languageName: node
   linkType: hard
 
+"@ai-sdk/google@npm:^2.0.32":
+  version: 2.0.32
+  resolution: "@ai-sdk/google@npm:2.0.32"
+  dependencies:
+    "@ai-sdk/provider": "npm:2.0.0"
+    "@ai-sdk/provider-utils": "npm:3.0.17"
+  peerDependencies:
+    zod: ^3.25.76 || ^4.1.8
+  checksum: 10c0/052de16f1f66188e126168c8a9cc903448104528c7e44d6867bbf555c9067b9d6d44a4c4e0e014838156ba39095cb417f1b76363eb65212ca4d005f3651e58d2
+  languageName: node
+  linkType: hard
+
 "@ai-sdk/google@patch:@ai-sdk/google@npm%3A2.0.31#~/.yarn/patches/@ai-sdk-google-npm-2.0.31-b0de047210.patch":
   version: 2.0.31
   resolution: "@ai-sdk/google@patch:@ai-sdk/google@npm%3A2.0.31#~/.yarn/patches/@ai-sdk-google-npm-2.0.31-b0de047210.patch::version=2.0.31&hash=9f3835"
@@ -9891,8 +9903,10 @@ __metadata:
     "@agentic/searxng": "npm:^7.3.3"
     "@agentic/tavily": "npm:^7.3.3"
     "@ai-sdk/amazon-bedrock": "npm:^3.0.53"
+    "@ai-sdk/anthropic": "npm:^2.0.44"
     "@ai-sdk/cerebras": "npm:^1.0.31"
     "@ai-sdk/gateway": "npm:^2.0.9"
+    "@ai-sdk/google": "npm:^2.0.32"
     "@ai-sdk/google-vertex": "npm:^3.0.62"
     "@ai-sdk/huggingface": "patch:@ai-sdk/huggingface@npm%3A0.0.8#~/.yarn/patches/@ai-sdk-huggingface-npm-0.0.8-d4d0aaac93.patch"
     "@ai-sdk/mistral": "npm:^2.0.23"