For AI agents: a documentation index is available at the root level at /llms.txt. Append /llms.txt to any URL for a page-level index, or .md for the markdown version of any page.
Evaluator results: LLM graders now provide a written rationale alongside each score, making it easier to understand why an evaluation passed or failed.
Experiments: refined evaluation analytics and comparison views while reducing unnecessary trace loading, and added backend status visibility and automatic refreshes for active experiments.
Evaluator editor: added expanded editors for LLM prompts and Python code, with immediate synchronization and smoother keyboard interactions.
Error Tracking: improved how errors are classified into incidents for more accurate investigation.
Model discovery: models now default to newest-first ordering, with clearer release dates and lifecycle information.
Model pages: improved layouts across compact, expanded, sparse, and data-rich views.
Experiment performance: compressed large experiment-log responses to reduce loading time and transferred data.
Hosted MCP: improved sign-in continuity for hosted MCP connections with a dedicated OAuth session flow.
Logs API: removed internal platform fields from log, trace, and span API responses.
Reports: reports can now include Pulse summaries for richer incident and behavior reporting.
Logs: request activity can now be grouped by Threads or Traces directly from Logs, with filtering, sorting, refreshing, and detail panels available in each grouped view.
Experiments: experiment lists now include evaluator scores and model details, making results easier to compare without opening each experiment.
Behaviors: Behaviors is now available to regular platform users, with a complete draft, training, and review workflow.
Home: added a recently visited section for quickly returning to prompts and datasets.
Red Team: improved campaign setup, engine performance, CLI presentation, documentation, and error handling.
Experiments: conditional evaluator results can now return N/A. Aggregate scores exclude unscored rows and show the number of scored results, while histograms display N/A separately.
Evaluators: refined the grader editor, expanded Code evaluation access, and improved how LLM and Code configurations are displayed and saved.
Behaviors: improved classification accuracy, reduced false positives, added controls for flagging incorrect matches, and made behavior charts clearer across time ranges.
Error tracking: improved error classification and added automatic clustering for errors occurring around the same time.
Model catalog: improved model information and catalog performance, including richer descriptions, release dates, and coding benchmark scores.
Request logs: customer names and email addresses are now preserved in request-log lists, and nearby-request navigation more accurately follows the selected event’s timestamp.
Dashboard sessions: bookmarked dashboards now refresh expired sessions automatically instead of unexpectedly returning users to sign-in.
Metrics: empty breakdown charts now remain in a stable no-data state without flashing.
Prompts: new drafts are created only after an edit, keeping prompt version history cleaner.
Dataset imports: imported message histories now appear as conversations, and expected outputs are stored and displayed as formatted JSON.
Experiments: boolean evaluator results now display false values correctly, and multi-turn prompt experiments retain the complete conversation history.
Experiment workflows: added task validation and made configured evaluator execution more consistent.
Prompts: fixed commit detection and redeployment behavior across all behavior-affecting settings. Load-balancing configuration is now preserved during JSON editing, and dropped parameters are persisted so they no longer reappear.
Evaluators: fixed commit messages overwriting evaluator descriptions and ensured each version displays its own commit message.
Security: closed an unauthorized live-event stream vulnerability and resolved an additional high-severity security issue.
Red Team: Red Team is now available online, with tools to create and run assessments, review results, and generate reports. The open-source Red Team engine is also available on PyPI.
Experiments: added filters for datasets, prompts, evaluators, and models, along with model, prompt, and evaluator details directly in experiment lists.
Dataset automation: automatic insert workflows now support trace datasets.
Dashboard autosave: named dashboard views now save automatically, making it easier to preserve dashboard changes as you work.
Dashboard charts: chart cards now stay above neighboring tiles when hovered or keyboard-focused, keeping chart controls accessible in dense dashboards.
Command palette: aliases now resolve during search, and the favorites action is now labeled “Pin.”
Onboarding: added a sign-out option during onboarding so users can switch accounts without getting stuck.
Dashboard charts: restored chart spacing, grid behavior, refresh controls, timezone selector behavior, and full-window calendar presets for daily, weekly, monthly, and compare-period views.
Onboarding: improved checklist steps, migrated onboarding variables, and fixed cases where users could get stuck or see visual flashes during onboarding.
Logs display: tool call usage and tool call results are now displayed more clearly in the Spans and Threads pages.
Limits and webhooks settings: improved cache and typing behavior, and fixed schema mismatch issues.
Dashboard charts: improved metrics controls with timezone-aware ranges, chart expansion, drag-to-zoom, synced hover, layout reset behavior, and graph customizations.
Experiments: improved the creation flow, model preset selection, log table wrapping, running indicators, and live result refresh while traces are still being written.
Prompt and Playground navigation: prompt identity is now driven by URL parameters for more reliable sharing, reloads, and navigation.
Dataset uploads and experiment results: fixed a regression where dataset uploads could silently fail and experiment runs could complete without results.
OAuth login: fixed a rate-limiting issue that could break OAuth or social login.
Gateway prompt variables: prompt variables now resolve with consistent precedence across Chat and Responses APIs.
Logging reliability: streaming logs now retry failed batch writes and route undeliverable batches to a recoverable queue instead of dropping them.
Streaming reliability: stalled streaming consumers now recover automatically, reducing delayed logs and background processing gaps.
Gateway credentials: dedicated deployments for one model no longer block shared credentials for other models under the same provider.
