Changelog
New
- Models: launched span-02, an open-weight decision model with anytime reasoning. Give it a state and typed questions in one call and it answers with choices, yes/no, scores, and numbers, each with a confidence. Set
budget_msto cap thinking time and it returns its best decision so far, or leavereasoningoff and it thinks only when it's unsure. Available on the Respan gateway with Respan credits, with weights on Hugging Face. - Security: added optional multi-factor authentication with authenticator-app enrollment, sign-in challenges, recovery codes, and security notifications.
- Gateway: added video generation across supported OpenAI Sora, Google Veo, Azure OpenAI, xAI, and OpenRouter models, including Sora remix.
- Gateway: expanded image generation, added OpenAI image editing and streaming, and enabled audio and video inputs in chat.
- Batch processing: added BYOK batch support for Anthropic, Gemini, Mistral, DeepInfra, and xAI, with individual request logs showing input, output, and cost.
- Workspace exports: logs, traces, and customer users can now be exported across the workspace.
- Models: Gemini Argon and ParetoInference GLM-5.3-Flash are now available with Respan credits.
Improved
- Logs: images, audio, and video are stored privately and can be viewed in log details.
- Models: added quantization filtering and sorting, refreshed model descriptions, and rebuilt public model discovery, comparisons, benchmarks, and provider-status pages.
- Rankings: added public model and provider rankings based on Gateway request and token volumes.
- Inbox: uncertain behavior classifications now appear as labeling requests, while samples from disabled behaviors are hidden.
- Billing: added separate image and audio pricing for more accurate multimodal usage accounting.
- Mobile: time-range pickers and filter menus now fit and remain accessible on smaller screens.
Fixed
- Traces: gateway-linked calls now display complete content, usage, and cost while preserving the trace's own metadata and customer details.
- Tracing: spans reported with an OpenTelemetry error status now appear as failed.
- Billing: BYOK tracing spans no longer deduct credits, and invoice-billed organizations can continue serving requests with a negative credit balance.
- Metrics: corrected tokens-per-second, average time-to-first-token, and dashboard date labels.
- Logs: fixed media links and nested content rendering in log details.
- Errors: the All errors row now correctly filters to failed spans.
- Models: custom models no longer appear in the public catalog.
- Gateway: streaming responses that close early now surface the provider's error message.
New
- Guardrails: launched Guardrails V2 for configuring prompt protections.
- Models: launched span-01-security, a specialized Span-01 model for prompt guardrails.
- Saved views: added personal and project visibility options when creating, saving, and duplicating views.
- Settings: added self-service account deletion.
- Simulation: scenario generation now lives in Datasets, and simulation runs appear alongside experiments in Experiments.
Improved
- Models: added modality filters, model tabs, custom-model filtering, and modality selection during model setup.
- Models: added benchmark filters and improved foundation-model names and catalog organization.
- Behaviors: improved Span-01 routing performance and classification throughput to keep behavior labels current.
- Mobile: improved the home page, settings, routing configuration, dashboards, list pages, and view tabs on phones and tablets.
- Docs and articles: standardized reading layouts, page navigation, diagrams, and FAQs across blog posts and customer stories.
Fixed
- Threads: fixed exports crashing or returning no rows.
- Prompts: fixed backup restoration missing stored backup keys.
- Models: fixed custom models disappearing when a catalog entry lacked a setting, and corrected the model-age filter.
- Logs: context-window failures are now labeled clearly. Behavior filters appear only in supported views, with confirmation before clearing them when changing grouping.
- Billing: corrected Anthropic one-hour prompt-cache write pricing.
- Gateway: chat requests now honor the provider load-balancing header, and fallback routing preserves models absent from the catalog.
- Customers: fixed customer filtering and ID sorting across pages.
- Global search: removed the unavailable Load Balancing destination.
New
- Home: launched a redesigned home page with a clearer overview of activity and easier access to key workflows.
- Logs: launched a reworked Logs experience with Behaviors integrated into log investigation and streamlined Spans, Threads, and Traces views.
- Traces: added dedicated actions to identify, copy, and navigate to the selected span.
- API keys: admin API keys can now manage limit policies, while keys remain unable to raise their own caps.
Improved
- Logs: content hidden by plan limits now displays an explanation and upgrade link instead of appearing as blank rows.
- Mobile: improved onboarding, settings, and empty-state layouts on smaller screens.
- Models: custom models now create organization-owned model entries. Expanded catalog synchronization to AWS Bedrock and Google Vertex.
- Privacy: PII redaction now also covers prompt templates and prompt variables in logs.
Fixed
- Threads: prevented unsupported OR filter combinations while preserving supported expressions.
- Simulation: run reports now retain the scenario used at execution, even if it is edited afterward.
- Billing: corrected pricing for fallback and overridden models, and ensured embedding requests use the requested model's rate.
- Gateway: request-supplied custom pricing is now consistently ignored across Gateway endpoints. Custom pricing remains supported for logging and tracing.
- Usage limits: retries and fallbacks now count once toward request limits. Customer and user limits are enforced consistently across supported request fields.
- API keys: setting spend, usage, or rate limits to zero now saves correctly.
- Gateway: fixed Gemini requests failing when custom metadata is included, and added clearer errors for requests without credentials or with unknown model names.
- Reliability: corrupted authentication and prompt-version cache entries now rebuild automatically. Improved serving and ingestion stability during heavy traffic.
- Self-hosting: fixed unavailable storage images preventing self-hosted installations from starting.
New
- SQL editor: added query previews and background CSV or JSONL exports.
- API keys: added Read, Write, Gateway, and Admin permission scopes for more precise access control.
Improved
- Playground: made Run easier to access, improved tool controls, and added automatic restoration of your last Playground with an option to start fresh.
- Models: model details now lead with providers and pricing, with API examples in a separate tab. Added an All / Discounted toggle for easier price comparison.
- Saved views: selected time ranges, visible columns, and sorting now persist in saved views.
- Logs: added thread and trace aggregate filtering and restored grouped-view refresh, counts, exports, and shareable view state.
- Behaviors: refined behavior definitions and detail views, improved breakdown charts and scrolling, and moved enable and disable controls into the list menu.
- Notifications: notifications now support parallel delivery across multiple methods, with duplicate methods prevented during setup.
Fixed
- Playground: saved response schemas now remain intact across reruns, continuations, and restored sessions.
- Prompt variables: long values now scroll correctly without clipping the first line.
- Threads: fixed Customer ID and API key filters returning server errors.
- Logs: provider service tiers are now recorded on passthrough requests.
- Provider settings: credential validation no longer creates duplicate internal API keys or splits activity across views.
- Reliability: requests in progress now finish during server replacements, reducing failures during deployments and restarts.
- Red Team: remote campaigns can resume from the last completed goal after a restart or interruption.
New
- SQL editor: the SQL editor is now available to all platform users for querying and exploring their data.
- Custom ID: the Custom ID page is now available to all platform users for exploring activity grouped by custom identifiers.
Improved
- Billing: reorganized the Usage tab so the billing period applies clearly to both usage and its breakdown, with a more consistent logs usage display.
- Red Team: confirmed findings now display directly from campaign events, and CLI scan and login commands provide clearer explanations when the server refuses a connection.
- Demo: product screenshots on the homepage now open the corresponding screen in the live demo.
Fixed
- Gateway: Responses API requests without a model now use your default model instead of returning a server error.
- Billing: corrected cached-token pricing for OpenAI Flex, OpenAI Fast, and Gemini Priority so charges and spend-limit usage reflect the selected service tier.
- Project members: added typed-email confirmation before removing a member to help prevent accidental removal.
- Log ingestion: isolated retry queues by deployment so failed messages cannot be replayed into the wrong deployment.
New
- Simulation: simulation is now available on the Prompts page, with scenario generation, runs, and results.
- Filters: added full-text search to help users find matching content directly.
Improved
- Models: model details now show original and discounted prices together, including percentage savings for each provider.
- Simulation: organized scenario details into Setup, Behavior, and Evaluation sections and improved scrolling to keep panel content stable.
- Side panels: standardized toolbar and content tooltips across the platform.
- Gateway: organizations without credits or a provider key now receive a clear activation-required error.
- API: retired the legacy
/api/generateendpoint. Use/api/chat/completionsfor chat completion requests. - Log ingestion: improved resilience against memory-related interruptions during heavy traffic.
Fixed
- Prompts: fixed prompt detail and version parameters causing unsupported filter errors when returning to prompt lists.
- Exports: corrected export time-range handling and filter mismatches so requests reflect the selected range and filters.
- Queries: large queries that exceed resource limits now return clear errors instead of generic server failures, including during pagination.
- Playground: new organizations now receive three free Playground calls to help them get started.
New
- PII redaction: added organization-level redaction settings with separate controls for Gateway requests and stored telemetry, configurable categories, sensitivity controls, and previews.
- Model leaderboard: added a Top models by spend chart to help users understand Gateway usage on the landing page.
- Observability: combined Tracing and Observability into a single experience while preserving existing links and workflows.
- Simulation: added a streamlined simulation flow where users can select a prompt and version directly, generate scenarios, run simulations, and inspect results.
Improved
- Gateway: BYOK requests can now fall back to managed credits when provider credentials encounter rate limits.
- Model leaderboard: legends now reflect the entire selected time range, model colors remain consistent, and individual models can be isolated for comparison.
- PII detection: improved redaction reliability for short telemetry fields, Unicode content, street addresses, private IPv4 addresses, and loopback IPv6 addresses.
- Tables: the first column now remains consistently frozen, preventing alignment from shifting across views.
- Behavior classification: improved processing reliability so behavior results remain current during traffic spikes.