Release history

The 17 October 2026 beta is published as @yourgpt/[email protected] and @yourgpt/[email protected]. These versions use npm's beta tag. The beta publication left the latest tags unchanged. For the prepared stable promotion, whose npm publication is separate, see the 2.6.0 upgrade guide.

Install the beta

Pin both packages when validating an upgrade:

npm install @yourgpt/[email protected] @yourgpt/[email protected]

Alternatively, use @beta on both package names to follow the beta channel. Install the optional client for your chat provider, such as openai or @anthropic-ai/sdk. The OpenAI speech/transcription helpers use their own HTTP transport.

What shipped

Feature or fix	Available behavior	Guide
Model parameters	GPT-5/GPT-6/o-series parameter normalization; unsupported reasoning temperature omitted with a warning	OpenAI
Compatibility fallback	Opt into advancing on a specific unsupported parameter without retrying the same invalid request	Fallback
Anthropic system prompts	Runtime preserves system-role instructions; explicit `systemPrompt` remains supported	System prompts
OpenAI Responses	Explicit transport selection, stored/stateless continuation, and provider-state replay	Responses
Image tool results	Native model-visible image/text results with OpenAI Responses and Anthropic	Image tool results
Speech generation	Buffered OpenAI TTS with typed audio results and cancellation	Speech
Transcription	OpenAI recorded-file STT, model-specific timestamps and speaker labels	Transcription
Microphone input	Explicit recording, cancellation, upload, and transcript review	Voice input

The preceding beta also fixed duplicate generation across stream readers, interleaved OpenAI tool arguments, trailing streaming usage, cancellation before later tools, screenshot attachments, mixed browser/server tool turns, and CJS/ESM declarations. The current release retains those fixes. MCP image outputs now retain their accompanying text.

Upgrade behavior to review

Anthropic: a system-role message that was previously ignored by Runtime now takes effect. The effective systemPrompt comes first, followed by distinct system instructions. Beta.3 deduplicates text shared by that prompt and a system message, and tolerates null/unsupported non-text system content while preserving valid text.
Reasoning models: the SDK does not silently disable reasoning to honor temperature. Use supported reasoningEffort: 'none' explicitly when sampling is required. Custom compatible endpoints can opt into passthrough rules.
Transport/storage: api: 'auto' keeps compatibility routing. Select Responses explicitly when needed, and choose store deliberately. Persist reduced continuation metadata and native tool output when replaying conversations; display text alone is insufficient. Remove top-level Runtime billing usage and debug events at your application's public response boundary after accounting.
Image tools: ordinary JSON results and the screenshot attachment path retain their behavior. Native image results require a compatible endpoint. Image input to a model and image output from a tool use different content shapes.
Audio: TTS/STT are separate from chat streaming. The recorder does not automatically send a chat message. Keep the transcript editable before submission.

Still pending

Beta.3 also restores raw structured tool results instead of coercing them to [object Object], classifies malformed Responses tool calls for configured fallback before output/tool execution, and handles models that omit toolCalls.

Work	Status
Native Decisions for Jev/TypeSafe, OpenAI, OpenRouter, Cloudflare	Implemented on `main`, outside these published versions; #116. Direct Jev/Cloudflare live validation remains pending.
Cross-provider prompt caching	Partial: OpenAI Responses option forwarding exists. Broader controls, cache accounting, and cold/warm validation remain; #114, #104.
Hosted YourGPT Responses replay	Full opaque metadata restoration is not yet supported by the hosted adapter; storage limits.
Native MCP/background lifecycle and broader reasoning contracts	Follow-ups in #119 and #118.
Google transport/Interactions and provider-native compaction	Planned in #117 and #120.
Audio providers beyond OpenAI	No built-in implementation in this release. Custom providers can implement `SpeechModel` / `TranscriptionModel`.
Physical microphone/browser testing	Automated lifecycle coverage exists; physical device verification remains.
Backend `TRAINING_CONFLICT` adoption	SDK parameter/fallback mechanisms shipped; application schema, fallback policy, and deployment remain application work (#122).
Realtime voice calls and image generation	Implemented on `main`, outside these published versions. Realtime is experimental and tracked in #126.

Selected compatibility and UI adaptations were consolidated into main in PR #131. Hosted MCP, concurrent threads and single-stream work remain separate follow-ups; see the roadmap.

Validation and production preparation

The isolated beta.3 hotfix passed 371 automated tests on Node 18 and Node 26, SDK builds/type checks, 22 live OpenAI/Anthropic/OpenRouter checks, 41 packed-consumer regressions and all 13 public entry-point imports. The downloaded npm archive matched the tested release archive. Post-merge GitHub CI and the docs/playground deployments passed. The previous image/audio beta validation remains recorded separately; Google, xAI and Fireworks compatibility paths used transport fixtures rather than live credentials.

Live media checks covered inline images and WAV/MP3 audio; other supported media representations have fixture coverage. Prompt delivery checks do not guarantee strict JSON formatting or application-level guardrail accuracy.

Before promoting your application, validate its actual provider/model combinations, structured-output parsing, persisted-history reloads, and microphone targets. The beta release itself did not promote npm beta to latest or deploy your backend.