npx @buildinternet/releases get openai-node-sdkA Live response's Decisions question instructions can now be a string or an array of text, image, or JSON parts, so code like question.instructions.trim() can fail on arrays, and Decisions request inputs accept new message roles and tool-history items. DecisionInputMessage is now a union, and code extending the old interface must use DecisionInputMessage.DecisionInputUserMessage instead.
Added a voice deletion API. Fixed agents doing repeated ancestor-path hashing and excessive work when redacting long log values.
The Responses API can now be called through Vertex AI.
Adds custom voice consent management, 30-minute cache-write usage counts, and compaction trigger IDs. Fixes include large WebSocket messages being truncated, request encoding and response parsing alignment, and directshow input naming on Windows.
The Realtime API now supports opt-in bounded connection recovery. Moderation examples were switched to neutral wording.
Adds typed Live sideband events to the API surface.
Adds a degraded external storage status field to the API. Also adds shutdown notices for GPT-4 model aliases, which are being retired.
Image API requests and responses now include safety identifiers.
Adds prompt-based voice creation and Decisions prompt caching options, along with Decisions support and audio fixture playback in Live sessions, project API key scope updates, and response error diagnostics and model notices. Fixed Realtime response error message declarations.
Adds regional processing residency options to the API. Remaining changes are dev dependency bumps and a clarification of Decisions request and response fields.
Added browser and computer toolsets, beta session spending limits, OCI external-storage provider support, and organization Decisions usage reporting. Also added a billing_not_active session error code, vault updates with metadata filters, and image URLs with detailed safety alert explanations.
Adds agent environment suspension and expiration, and preserves detailed token usage in totalUsage for runners. Fixed a migration guide troubleshooting link and bumped a Next.js dependency in the Vercel Edge ecosystem tests.
The API now supports prewarmed hosted environments.
Fixed incorrect custom voice creation parameters, structured text parsing in Responses to apply only in final message phases, and safe decoding of inferred URL filenames in uploads. Also refreshed development and ecosystem dependencies.
Adds standalone Decisions support to the API surface.
Agents can now observe local tool failures and run local tools during session creation streams. Typed function helpers support tool search via defer_loading=true.
The API now supports custom voice creation and agent session events.
The agents module can now prepare hosted files and download result artifacts.
Added agent session trace listing and binding of typed application functions to Agents tools. Fixed Chat Completions rejecting original image detail, corrected the eval run cancellation endpoint, and made Responses WebSocket reconnects refresh credentials and reserve lane IDs across session replacements.
