Experiential changelog: what's new each month
Every stable Experiential release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 3 months covered; the current month updates daily.
November 2026 · so far
24 releases: 0.7.226 → 0.7.261
November brought decision-model routing capabilities, an open-source coding agent release, and multiple improvements to gateway handling and billing across model providers. The month focused on reliability fixes, expanded model support, and clearer cost tracking for image and reasoning operations.
Highlights
- Open-source exp-code coding and collaboration agent released with self-contained sandbox runner.
- Decision-model category scores and tag-based routing for more sophisticated request handling.
- Image operation billing refinement for GPT and xAI edits with separate input and output tokens.
- Gateway reliability improvements including stuck state fixes and better error classification.
- Input moderation now accepts configurable wait time for parallel classifier checks.
New
- Decision-model category scores provided for every answered check.
- Project aliases can be routed by decision-model tags via new tag-rules router policy authoring.
- Fleet-wide lane limits through GatewayRungDispatchPolicy.
- Parallel input moderation accepts explicit additional wait time from 0 to 5000 ms.
- Optional content-free evidence collection for online input checks in terminal ledger writes.
- Claude Code billing-header block automatically prepended to system prompts on plan rungs.
- Image output token tracking with separate image_tokens subset in settlement details.
- Gemini continuation routing to its signature's issuer for better request handling.
Improved
- Dropped gateway Postgres connections now return retryable 503 instead of failing silently.
- Non-streaming reasoning tool turns cut by token limits now settle as length rather than 502 error.
- Route unavailable at admission escalates as retryable 503 with Retry-After header instead of anonymous 500.
- Unclassified boundary exceptions now logged with type and raising frame for better diagnostics.
- Warm-cache throttle fix allows long Retry-After values to advance failover routes.
- Gemini 4 Argon thinking levels (none, low, medium, high) now supported with proper budget translation.
- Azure AI Foundry chat reasoning count now read at top-level admission.
- Malformed Gemini function calls redial with VALIDATED calling for recovery.
Fixed
- Responses WebSocket captures no longer expire while waiting for HTTP relay metadata.
- Gateway exits for five stuck states including guardrail pool worker leak on cancel and health registry pruning.
- Alias reload backoff and requests left open by crashes now properly closed.
- Hard capture memory bounds for spilling collectors with explicit background-writer budget and oversized body handling.
- Gemini output blocks via OpenRouter now filed as refusals instead of errors.
- Repeatedly broken tool calls now answer with 400 instead of continuing to fail.
October 2026
43 releases: 0.7.137 → 0.7.225
October brought major improvements to request handling, reasoning display, priority management, and provider integrations. The month focused on enhanced reliability, cost transparency, and broader model support across the gateway.
- Priority overflow caps changed: Pro callers now overflow to 1.5x (was 2.0x) and paying callers to 1.25x (was 1.5x).
- Enforcing policy and deterministic-adapter revisions now participate in replay identity; stale replay contracts fail closed.
- Per-request token and retry-inclusive cost reservations remain finite; explicit spending_limit_usd=None removes only the aggregate cap.
- Catalog-backed evaluation spending limits are opt-in only; new evaluations have the limit off by default.
Highlights
- Reasoning from models is now displayed by default in Chat and Responses, with encrypted fallback for carriers.
- Priority admission tiers (free, paying, Pro) now have fairer overflow caps and better capacity handling.
- Custom tool declarations with grammars are now supported on Chat and Responses.
- Web search on Responses is fixed and now returns proper result lists.
- Audio endpoints (speech and transcription) are now available on the OpenAI wire.
New
- Calling-app attribution now tracks User-Agent, originator, and client app across decisions, images, and embeddings.
- OpenAI Decisions API endpoint with predicate, choice, and score support.
- Custom tools with text, Lark, or regex grammar declarations in Chat and Responses.
- Authored scenarios can be added to evaluations alongside mined tasks.
- Guardian observation mode for policies with bounded inspection and separate accounting.
- Native audio surfaces on POST /v1/audio/speech and POST /v1/audio/transcriptions.
- Guardrails observation input with content-free outcomes while keeping enforcement as default.
- Access programs opt-in injection for Responses on gateway rungs.
Improved
- Kernel file locks for text and sandbox cells now protect processes during execution with clear failure on unsupported filesystems.
- Evaluation spending limits are now opt-in; new evaluations display estimated cost with limit off by default.
- Fairness improvements to rollout scheduling across models under shared concurrency.
- Retry-After and HTTP-date hints are now respected with bounded jitter.
- Token pricing schedules freeze at evaluation preparation time for clearer cost reporting.
- Cache affinity sessions now preserve fallback rungs instead of one-rung routes for better failover.
- Pre-first-byte budget added as configurable request-level timeout (default 90s) to support failover chains.
- Reasoning summaries are now best-effort with disclosure when a route cannot provide them.
Fixed
- Credential-related responses are now preserved instead of discarded when discussing password management or authorization.
- Mistral tools without descriptions no longer send null descriptions upstream.
- Responses web search include filters now accept web_search_call.action.sources.
- OpenRouter Responses rungs no longer receive refused include parameters.
- Replayed reasoning items correctly lose output-only status to avoid OpenAI parameter errors.
- Tool calls with incomplete JSON in length-limited responses are now classified as interrupted and can recover in fresh rollouts.
September 2026
37 releases: 0.7.98 → 0.7.136
September brought major improvements to token handling, capture infrastructure, and provider compatibility across Experiential. The month focused on precise accounting, reasoning preservation, and expanded model support including new Claude and OpenAI features.
- Removed arbitrary 4,096-token fallback for omitted output limits; required caps now derive from declared per-route maxima.
- Unsupported numeric-budget translations now refused rather than silently converted to advisory effort.
- Truncated tool-call arguments and length termination no longer have empty JSON invented or tools auto-completed.
- Stream: false parameter now accepted but not forwarded upstream on embeddings endpoint.
Highlights
- First-token stall bounds now have their own timeout allowance, separate from overall request budgets, improving streaming reliability.
- Capture system now supports batched destinations and asynchronous transfers, with durable acknowledgement tracking and per-turn usage preservation.
- Added support for token probabilities on OpenAI-compatible and Anthropic lanes, with proper qualification controls and provider-specific translations.
- Implemented ordered model fallback chains for Chat Completions, Responses and Messages, allowing explicit model references and recovery strategies.
- Private reasoning content now stays uncommitted until output or completion, enabling policy-authorized failover without leaking reasoning.
New
- Token probability support for Chat Completions and Responses with provider-specific field translation.
- Generated images from Gemini and compatible models now survive streaming and buffered responses with validated raster data.
- Native capture collector with typed hosted destination interface preserving conversation context, tool definitions, and provider reasoning.
- Embeddings endpoint now accepts text, text batches, token-ID arrays and token-ID batches with proper accounting.
- Desktop Capture for macOS with live request metrics and independent watchdog recovery.
- Connection support for ChatGPT and Claude subscription plans with OAuth-based sign-in and usage window tracking.
- Automatic Vertex prefix caching for whole-message prompts without client markers, with US and EU jurisdictional endpoint support.
- Explicit asynchronous capture transfers for already-admitted records to background writers without waiting for delivery capacity.
Improved
- Default per-lane in-flight bounds now refuse overflow instead of queuing, preventing unbounded concurrency.
- Idle model generation now bounded by genuine progress rather than transport keepalives like pings and comments.
- Tool-search rounds and reasoning repair now use hard physical-call accounting with inclusive attempt ceilings and configurable backoff modes.
- Stop-sequence draining, downstream backpressure, and cancellation now preserved through provider tool work.
- Cache handling on disconnected streams now prices at organization's recent cache share instead of entirely fresh.
- Eligible hosted prompts now checkpointed durably after winning lane is frozen and before releasing output.
- TypeSafe input and output token counts now independently validated even when decision answer validation fails.
- Gemini content-policy refusals now properly recognized in both HTTP and SSE error envelopes.
Fixed
- Local-process sandbox pipes at high descriptor numbers now work correctly using resource-scoped selectors.
- Synthesized tool-search names now properly handled in named tool choice.
- Invalid generated images from Gemini no longer trigger automatic retry, avoiding duplicate billing.
- Gateway rejection of independently rounded TypeSafe score responses now accepts hundredth-valued rounding-compatible responses.
- Encrypted reasoning items without item IDs now accepted during stateless replay on Responses.
- Preserved-thinking carriers from closed rungs are now dropped and disclosed as unavailable instead of causing errors.
Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Experiential: pricing, features and alternatives · All changelogs