@ai-sdk/openai

Patch Changes

• openai: Preserve valid JSON string tool arguments and serialize BigInt and cycles in tool calls and results, retaining malformed argument replay fallbacks

• Forward languages and keywords transcription hints as repeated multipart fields instead of silently dropping them. Warn when these REST-only options are supplied to streaming transcription.

Patch Changes

• Support promises and synchronous or asynchronous callbacks in provider API key and token settings. Resolve credentials per authenticated request while preserving existing authentication behavior and synchronous credential serialization.

  The experimental OpenAI Live getServerWebSocketConfig() method can now resolve asynchronously. Await its result before using the connection settings.

• Introduce a model entry contract with named capability sections and request configuration. Consolidate the OpenAI and Anthropic model catalogs so existing request behavior and generated capability documentation use the same source of truth. Catalogs also declare the provider options each provider accepts and which models support them.

Patch Changes

• openai: Send native JSON input and prompt cache keys to Decisions

• openai: Map top-level reasoning to efforts supported by GPT-5.x models instead of sending unsupported values that OpenAI rejects with a 400. max is now only sent on the Responses API, and GPT-5.6 on Chat Completions sends reasoning_effort: 'none' when function tools are used, since it rejects function tools with reasoning. A warning is emitted when a reasoning level was requested.

• Map the top-level reasoning option consistently across providers

  Providers now resolve reasoning the same way:

  1. A provider-specific reasoning setting in providerOptions is sent unchanged. OpenAI no longer drops providerOptions.openai.reasoningEffort values that are not in its list for the model, and Fireworks no longer rewrites providerOptions.fireworks.reasoningEffort. Fireworks also sends top-level reasoning values directly instead of rounding them to 'high' at most, except for model families known to reject 'none' (GPT-OSS, GLM-5.3, Muse Spark) or 'minimal' (Nemotron, Inkling).

  2. Otherwise, reasoning maps to the closest level the model supports, rounding down, and rounds up only when there is no lower level. For example, 'xhigh' now maps to 'high' instead of 'max' on Anthropic and Bedrock models without xhigh effort, and 'medium' maps to 'low' instead of 'high' on DeepSeek, Kimi K3, and Mistral GLM-5.3 models.

  3. 'none' on a model that cannot disable reasoning maps to the model's lowest level with a compatibility warning, instead of being dropped or sending a request the API rejects. This affects GPT-6 models without 'none', xAI grok-4.5 and later, Groq GPT-OSS, Kimi K3, Perplexity, QuiverAI, Bedrock OpenAI gpt-oss models, and Gemini 2.5 Pro (minimum thinking budget). Gemini 3 and Anthropic and Bedrock between_tools mappings now emit the compatibility warning too.

  mapReasoningToProviderEffort now accepts any reasoning value, returns an optional providerOptionsEffort unchanged, and rounds levels that are missing from the effort map to the closest supported level. mapReasoningToProviderBudget maps 'none' to the minimum reasoning budget, always returns a budget, and uses the default percentage for levels missing from budgetPercentages. OpenAICompatibleChatConfig adds an optional mapReasoning function for mapping the top-level reasoning option.

Patch Changes

• openai: Add GPT-6.1 Terra model support

• openai: Read transcription duration from usage for duration-billed json responses

  durationInSeconds now falls back to usage.seconds when usage.type is duration and the response has no top-level duration (e.g. gpt-transcribe with the default json response format).

• Support streaming transcription with gpt-live-transcribe and gpt-transcribe. For these models the language provider option is sent as a single-entry languages list. gpt-transcribe warns about and omits streaming.delay, which it does not support. gpt-live-transcribe rejects non-streaming transcription, since it has no REST endpoint.

• openai: Strip propertyNames schemas that omit type: 'string' (e.g. Zod 3 records with constrained keys) instead of throwing an unsupported functionality error

• openai: Report unfinished web searches as tool errors

• openai: Only send transcription response format and timestamp defaults when needed

  Transcription requests no longer send response_format=verbose_json or timestamp_granularities=segment unless the caller sets them or the model needs them. Previously, any providerOptions.openai value (such as a prompt) caused verbose_json to be sent for models like gpt-transcribe, which rejected the request.

Patch Changes

• Add azure.tools.toolSearch and support azure provider options on responses function tools

• Fix string interpolation in Amazon Bedrock unsupported provider tool warnings and OpenAI / OpenAI-compatible completion errors for unexpected system messages.

Patch Changes

• Fix non-streaming Responses API structured output failing when multiple assistant messages contain JSON. Providers can supply separate structured-output text for parsing, allowing OpenAI and Azure Responses models to select the last final_answer while preserving all generated content, annotations, conversation history, and token usage.

Patch Changes

• Report interrupted when a stream ends cleanly without its provider completion signal, preserving partial output without emitting an error solely for the interruption.

  This applies to OpenAI (Chat, Responses, and Completions), Anthropic, Google (Generate Content and Interactions), Alibaba, Amazon Bedrock Converse, Cohere, DeepSeek, Groq, Hugging Face Responses, Mistral Chat, Moonshot AI, Open Responses, Perplexity, and xAI Responses. It also covers cutoffs in later Anthropic messages and Google automatic continuations.

  Compatibility notice: These streams previously fell back to other or omitted the finish part. Applications must now handle interrupted in finish-reason switches, mappings, and custom validation. Check finishReason === 'interrupted' to detect incomplete responses; interruption alone does not trigger error handling.

  Explicit unknown completion reasons still map to other. Actual provider, parsing, and transport errors retain their error paths; affected error paths that previously left the finish reason as other now report error. OpenAI-compatible remains unchanged.

Patch Changes

• Add a shared encodePathSegment helper and use it to encode provider-returned identifiers in credentialed follow-up request paths.

• Defer construction of provider validation schemas until first use to reduce package import time.

• openai: Support 12h prompt cache writes and cache prewarming

Patch Changes

• Add maxOutputTokens and reasoning to experimental decision calls and forward them to language-model-backed decisions. Language-model adapters keep reasoning disabled when omitted. Native decision providers warn when either setting is unsupported.

• openai: Preserve detected language codes in transcription responses

• openai: Forward shell_call timeout_ms and max_output_length to the shell tool and round-trip max_output_length

Patch Changes

• Add an optional modelId setting to Azure OpenAI language-model factories for
  accurate telemetry and capability detection with custom deployment names. Keep
  the deployment ID for API routing and preserve existing behavior when the model
  ID is omitted.

• perplexity: Add decisions api support

Patch Changes

• deprecated

  openai: Deprecate 'webSearchPreview()' tool

• openai: Support ordered 'additional_tools' from tool results

Patch Changes

• Restore raw OpenAI custom-tool input deltas for consumers that forward them to Responses API clients. Mark streamed input with inputEncoding: 'text' so UI clients can display it progressively without JSON parsing. Preserve the encoding when partial UI messages are saved and resumed. Final provider tool-call inputs remain JSON-encoded.

Patch Changes

• openai: Replay web search results when response storage is disabled

• Expose complete retrieved web-search URLs and file-search documents as sources, and normalize inline references separately as citations on text parts and text-end events. Preserve citation-derived sources when retrieval data is unavailable, including web-search-preview and Azure responses. Keep raw annotations, citation titles, and text locations available. Deduplicate file-search passages by file ID.

• openai: JSON-encode custom tool input deltas for progressive UI updates

Patch Changes

• Map the top level embedding dimensions to provider settings

• openai: Send image detail in decisions

• Report failed and incomplete web searches as tool errors instead of successful tool results in OpenAI Responses API calls and batch results.

Patch Changes

• Add ordered text, file, and JSON parts to experimental decision state. Support image input in OpenAI Decisions and language model adapters. Gateway retains its existing string and JSON request format and rejects files, as does TypeSafe AI.

  Arrays passed directly as state now contain decision state parts. Wrap JSON arrays in an object or a json part.

  Normalize all public state forms into an array of parts before calling decision providers. Providers receive text and JSON objects as text and json parts.

  Serialize decision file bytes as base64 in OpenTelemetry state attributes. Model-call spans record normalized state parts, while outer spans retain public state inputs.

  Preserve native JSON state in TypeSafe AI when state contains one JSON part. Label shared state in language-model decision prompts.

Patch Changes

• provider: Report decision refusals as refusal answers

Patch Changes

• Limit jsonl rows to 64mb during batch result downloads

• Replace the OpenAI structured-output wrapper for decide with the native Decisions API. Only gpt-6-luna is currently supported by this endpoint; other Responses models no longer work with openai.decisionModel() or its deprecated evaluationModel() alias.

  Responses API options, such as reasoningEffort, are no longer applied and produce unsupported-option warnings. The Decisions API supports providerOptions.openai.safetyIdentifier. Native token usage and confidence values are preserved, and per-question refusals fail the decision call.

• provider: Add a portable max reasoning level with native provider mappings, compatibility coercions, and budget-based fallback support.

  Compatibility notes:

  • TypeScript source compatibility: Adding max widens LanguageModelV4CallOptions['reasoning'] and the derived ReasoningLevel type. Third-party providers and consumers with exhaustive switches or Record<ReasoningLevel, ...> mappings must handle max before upgrading. Ordinary calls using existing reasoning levels remain accepted.

  • Amazon Bedrock: Existing Nova 2 calls using portable reasoning: 'xhigh' now send maxReasoningEffort: 'high' instead of the unsupported 'max', and return a compatibility warning. Anthropic and OpenAI Bedrock effort mappings retain their existing xhigh behavior. Partial Bedrock reasoning configurations continue to preserve explicit values and derive missing fields; explicit effort/budget overrides and disabled thinking no longer emit warnings for portable mappings that are not sent.

  • Fireworks: Existing portable minimal and xhigh calls keep their request mappings to low and high, respectively, and now return compatibility warnings in generation and streaming results. Portable max maps to high with the same warning behavior.

  Release classification: This remains a patch changeset under the repository's explicit release policy, which uses patch releases for both fixes and features. The TypeScript source-compatibility caveat for exhaustive consumers is disclosed above. Maintainers can override this classification with a major label if they decide to align the provider-spec addition with a future AI SDK major release.

Patch Changes

• Rename evaluate to decide

• openai: Warn when reasoningSummary is used with chat models

Patch Changes

• ai: Add telemetry to speech generation and transcription, including provider usage propagation and experimental streaming lifecycle callbacks

Patch Changes

• openai: Add GPT-6.1 Sol model support

Patch Changes

• Use standards-compliant User-Agent header

Patch Changes

• Fix streamed tool calls with missing, blank, or repeated IDs.

• provider: Advertise image model file and mask input support

  Use confirmed model IDs for capability declarations so unrecognized model names remain unknown. Include Together AI FLUX.2 Pro and Flex single-image editing, and allow asynchronous capability lookups and middleware overrides to resolve to unknown.

  Advertise QuiverAI Arrow 2 and Arrow 2 Telos file-input support, and mark Together AI Gemini image inputs unsupported by the current single-image request mapping.

Patch Changes

• Default OpenAI Responses function tools to strict: false when tool strict mode is omitted.

Patch Changes

• openai: Classify Responses stream errors with null codes

Patch Changes

• openai: Expose Chat Completions audio transcripts as generated text

• GPT-6 Sol and Luna model IDs

Patch Changes

• openai: Cancel image edit URL downloads when the request is aborted

Patch Changes

• openai: Preserve provider file references in Responses tool results

• openai: Send assistant text as Responses easy input messages

Patch Changes

• Add experimental Choice, Score, and Boolean evaluations through openai.evaluationModel() and a shared structured language-model evaluation adapter in @ai-sdk/provider-utils/experimental-evaluation. Preserve exact labels and metadata, validate score bounds, and return prompted Boolean P(true) estimates validated to be in [0, 1]. Boolean estimates are not guaranteed to be calibrated; application code chooses thresholds.

Patch Changes

• Add optional browser-direct WebRTC for experimental client-delegated Live conversations alongside the existing WebSocket path. Exchange SDP through an application endpoint with api.session, configure server-owned data-channel permissions, and preserve committed React session ownership. Capture follows the selected sender track, borrowed tracks remain caller-owned, and disconnect recovery and finalization stay bounded. Applications continue to handle client delegation and submit context; Live session updates and Responses delegation remain unsupported.

  Serialize microphone sender changes and close the peer if detachment fails, without stopping borrowed tracks. Validate nonempty SDP setup answers with a bounded response body, and document the same-origin broker authentication contract.

• Add experimental OpenAI Live provider support through the unified openai.experimental_realtime factory for server WebSocket sessions with client delegation. Applications own their agents and tools, receive continuous audio/transcript events and delegation metadata, and return context through validated channels. Support immutable startup options, microphone mute controls, graceful session close, and cumulative voice usage. Responses delegation and Live session updates reject before sending.

  Route known Live model IDs to Live, allow an OpenAI-specific api override for early-access models, and preserve legacy Realtime defaults for unknown IDs. Token minting follows the same selection rules and rejects Live before requesting unsupported credentials. Extend the realtime v4 specification with optional server WebSocket configuration, per-connection raw-event parsers, and model-wide startup/finalization capabilities. This provider layer supplies connection settings and protocol mapping for server adapters; browser lifecycle and UI integration belong to the core runtime and framework hooks.

  Preserve Realtime client event IDs for session updates and audio appends, and correlate server errors with the originating client event.

Patch Changes

• Reject unsupported batch request types

• Support image generation requests in batches

• openai: Allow providers to disable web search source includes

• openai: Remove propertyNames from JSON Schema

Patch Changes

• openai: Add batch cancellation and listing

Patch Changes

• openai: Return an AI SDK error for empty chat completion choices

Patch Changes

• Include explicit message item types in Azure AI Foundry Responses requests.

• openai: Add GPT Image 2.5 Flare and Sunburst model IDs

• openai: Reject denied programmatic tool calls

• openai: Report apply patch responses with a tool-calls finish reason

Patch Changes

• Support per-request models in batch

• openai: Add async tool calling