@ai-sdk/openai

batch: Surface the uploaded input file on the batch start result (providerMetadata.<provider>.inputFileId / inputFileExpiresAt) and accept an inputFileExpiresAfter provider option on the OpenAI and xAI batch input file upload

openai: Preserve complete Responses API raw usage objects

openai: Implement getFileMetadata, downloadFile (streaming), and deleteFile on the OpenAI files interface, support streaming uploads via { type: 'stream' } data, expose byteSize/createdAt/expiresAt on upload results, and thread abortSignal/headers through all file operations; blank and dot-segment file ids are rejected/encoded so they cannot retarget request paths

openai: Handle null usage in responses

Add tool calling support to batch

Fix workflow deserialization for OpenAI Responses models.

Preserve explicit prompt cache breakpoints on scalar Responses tool results.

Fix Responses history serialization for regular functions named tool_search.

Normalize non-object replayed Chat Completions tool arguments to empty objects.

Add batch completion webhooks. experimental_startTextBatch accepts a webhookUrl, and the gateway provider registers it through the batch callbackUrl contract and exports typed async-job metadata. Direct Anthropic and OpenAI batch providers return an unsupported warning when the option is provided.

Prevent negative text output token counts when providers report reasoning tokens. Perplexity reasoning tokens are now treated as separate from completion tokens.

openai: Expand internal parallel tool call wrappers from the Responses API while preserving stateful continuation and streaming fallbacks

Handle empty string tool call IDs

openai: Support built-in and provider-defined tools in the Responses allowedTools option

allowedTools emitted every allow-list entry as { type: 'function', name }, but OpenAI identifies built-in tools by type. Allow-listing a declared provider-defined tool (web search, image generation, MCP, custom, ...) therefore failed with Tool choice '<name>' not found in 'tools' parameter. Entries are now derived from the declared tool, including the MCP server label and custom tool name.

Tools that OpenAI cannot allow-list (the tool search tool, deferred tools, and namespaced tools) are dropped from the allow-list with a warning, and an error is thrown if that would leave the allow-list empty rather than silently sending an unrestricted request.

Ambiguous names are now reported instead of resolved silently. A name that matches both a declared tool and another tool's provider tool name resolves to the declared tool and warns; a provider tool name shared by several tools in the same request (two MCP servers, for example) is dropped with a warning. A name that matches no declared tool keeps its existing behavior and is now warned about.

provider/openai: Support explicit Responses compaction triggers

Reconstruct provider-executed shell calls when continuing OpenAI Responses with storage disabled.

Avoid duplicate MCP approval request references when continuing stored OpenAI Responses.

Keep client-executed function calls paired with their outputs when chaining OpenAI Responses with a previous response ID.

openai: Serialize tool text outputs when an output schema is configured

Accept serviceTier: 'fast' on OpenAI chat and responses models. OpenAI renamed priority processing to Fast mode and accepts service_tier: 'fast' and 'priority' interchangeably, so 'fast' is now passed through verbatim and gated on the same model capability as 'priority'.

provider/openai: Correlate rotating Responses API item IDs by output index

openai: Support batch APIs with experimental_startTextBatch

Fix streamed tool calls with non-zero, non-contiguous, reused, or missing indexes.

Serialize file upload expiry settings in the multipart field shape accepted by OpenAI.

provider/openai: Resolve responses doStream at response.in_progress instead of first output token

The early-stream-error peek treated response.in_progress as an unknown chunk, so doStream did not resolve until the first output item arrived — delaying stream availability (and downstream TTFB for proxies/gateways) by the model's full time-to-first-token. response.in_progress is now modeled in the chunk schema and marks the request as accepted: the peek keeps watching for error frames for a short grace window (50ms) so quota/rate-limit errors flushed alongside response.in_progress still throw as retryable APICallErrors, while healthy streams become available right after upstream acknowledges the request.

Add experimental_toolCaller routing to generateText for code mode

Add blocked domain filters to the OpenAI and Azure Responses API web search tools.

Add experimental streaming speech translation models (openai.translation('gpt-realtime-translate') over the OpenAI Realtime translations WebSocket and google.translation('gemini-3.5-live-translate-preview') over the Gemini Live API). connectToWebSocket in @ai-sdk/provider-utils now passes close code and reason to onClose (additive, optional parameter).

Add OpenAI Responses Programmatic Tool Calling with hosted program tools, function caller controls, structured output schemas, and multi-step continuation support.

Apply reasoning, service tier, and image defaults to recognizable future OpenAI model family versions.

Preserve stored tool search output item IDs from provider metadata.

provider/openai, provider/open-responses: Throw a descriptive error when the Responses API returns a 200 with no output

A successful (200) Responses body missing the output array previously threw an opaque output is not iterable TypeError from doGenerate. Both providers now surface a clear APICallError ("Responses API returned no output …"), including the incomplete-details reason (and status, for open-responses) when present. When the body includes a response.error, its message is surfaced first so upstream error details aren't masked by the generic fallback. This makes malformed/incomplete upstream responses actionable instead of a cryptic crash.

Add the client-executed OpenAI Responses API computer tool with batched actions and screenshot outputs.

Fix realtime transcription auth header handling: per-call authorization headers now override configuration headers regardless of header-key casing (last case-variant wins), and the Bearer scheme is matched case-insensitively.

Reject empty OpenAI, Anthropic, and Replicate base URLs with a helpful AI SDK
invalid argument error.

provider/openai: Add GPT-5.6 reasoning and prompt cache controls

provider/openai,provider/gateway: Add gpt-5.6 model ids

Fix StreamingToolCallTracker finalizing streaming tool calls on parsable partial JSON. Tool calls now only finalize during stream flush, restoring the behavior of #13137: a parsable argument buffer can still be the prefix of a longer argument string, so finalizing early could act on truncated tool inputs.

Send inline image file parts in OpenAI chat requests as data URLs instead of bare base64 strings.

Add experimental streaming transcription support for transcription models, including OpenAI gpt-realtime-whisper and xAI WebSocket STT.