Changelog

Product changes

• Rankings percentage toggle -- Bar charts on the rankings page now have a menu button that lets you switch between percentage-normalized and raw token-volume views.
• Activity deep links to individual messages -- You can now deep-link to a specific message inside the Activity prompt overlay by appending ?message=<n> to the URL, and the selected row stays in sync as you navigate with prev/next buttons. Docs
• Fixed: media_resolution parameter silently dropped for Gemini models -- Sending media_resolution (e.g. MEDIA_RESOLUTION_MEDIUM) to Gemini models now correctly forwards the value to Gemini's generationConfig, changing image token counts as expected.
• Fixed: image generation returning empty errors for some providers -- Image generation requests to providers that return non-streaming binary data (e.g. base64 images) now correctly deliver the generated image instead of an empty {} error.

New models

• Microsoft: MAI-Voice-2
• Microsoft: MAI-Transcribe 1.5
• Microsoft: MAI-Image-2.5

Product changes

• May Release Spotlight blog post -- Published the May Release Spotlight covering speech and transcription APIs, Model Fusion, private models, enterprise workspace controls, and new model launches.
• Fixed: model-specific default parameters not shown on API tab -- The Parameters table on model detail pages now displays the model's actual default overrides (e.g. top_p: 0.95) instead of generic global defaults. Docs
• /llms.txt and /llms-full.txt now resolve at the domain root -- openrouter.ai/llms.txt and openrouter.ai/llms-full.txt now redirect to their generated equivalents under /docs/, following the llms.txt specification.
• Fixed: verbosity parameter default in docs -- The verbosityparameter reference no longer lists medium as the default; the effective default varies by upstream provider.

Product changes

• Fusion Router documentation and model page -- Added a dedicated Fusion Router documentation page and a model page describing the full panel → judge → synthesis pipeline, configuration options, and failure behavior.
• MiniMax M3 reasoning support -- MiniMax M3 now supports the thinking parameter, enabling extended reasoning for complex tasks.
• Fixed: Azure forbidden error mapping -- Azure provider responses with a forbidden status in server-sent events are now correctly surfaced as HTTP 403 instead of being misclassified.

New models

• MiniMax: MiniMax M3

Product changes

• Fallback errors in error metadata -- When all provider fallbacks fail, the error response metadata now includes a previous_errors array showing every attempted provider's error code, message, and name, making multi-provider failures easier to debug.
• Fixed: 404 page layout -- The 404 page no longer displays visual artifacts (stray borders, duplicate footer, off-center content).
• Fixed: multipart/form-data error code -- Sending multipart/form-data to JSON-only endpoints (e.g. speech-to-text) now returns a clear 400 Bad Request instead of a misleading 500 Internal Server Error.
• Fixed: playground reasoning panel scroll -- Opening the reasoning panel in the playground now scrolls to the top instead of an arbitrary position.

Product changes

• Preset ID in generation responses -- The GET /api/v1/generation endpoint now returns preset_id when a preset was used, letting you trace which preset produced each completion. Docs
• Fixed: activity overview date range -- The activity overview widget no longer breaks when selecting a date range exceeding 365 days.

Product changes

• IP address filtering for enterprise plans -- Enterprise organizations can now restrict all API key usage to approved IP address ranges from Privacy settings, reducing the blast radius of leaked keys.
• Negation filters in activity and logs -- Filters now support exclusion — click "is" on any filter chip to toggle it to "is not", letting you exclude specific models, providers, API keys, or workspaces from your view. Docs
• Presets: Responses API support -- Creating presets from inference now supports the Responses API via POST /api/v1/presets/:slug/responses, and new documentation covers creating and updating presets directly from inference request bodies. Docs
• Prompt caching: session_id sticky routing documentation -- Added docs on using session_id to control sticky provider routing, which activates on any successful request rather than requiring a cache hit. Docs
• Anthropic thinking tokens in API responses -- API responses for Anthropic models now include output_tokens_details.thinking_tokens, providing precise reasoning token counts for Claude Opus 4.8 and later.
• Fixed: Responses API streaming crashes in OpenAI SDKs -- Streaming keep-alive heartbeats on the Responses API no longer cause union_tag_invalid (Python) or invalid_union (TypeScript) validation errors in the OpenAI SDKs.

New models

• Anthropic: Claude Opus 4.8
• Anthropic: Claude Opus 4.8 (Fast)
• StepFun: Step 3.7 Flash

Product changes

• Series B announcement -- Published the $113M Series B announcement.
• API tab redesign on model pages -- Each model's API tab now shows the /responses and /messages endpoint references alongside /chat/completions, making it easier to discover alternative request formats. Docs
• Presets: Anthropic Messages API support -- Created presets from inference now support POST /api/v1/presets/:slug/messages, the Anthropic Messages counterpart to the existing chat/completions preset route. Docs
• Stream inactivity timeout -- Streaming requests that produce no chunks for 30 seconds are now aborted immediately instead of waiting up to 5 minutes, catching dead connections and empty streams faster.
• Guardrails API documentation -- Added a guide for updating your workspace default guardrail programmatically via the API. Docs
• Fixed: EU routing for non-Google models on Vertex -- Non-Google models (Claude, Llama, DeepSeek, etc.) routed through eu.openrouter.ai now use the correct endpoint path, fixing 400 errors caused by a hostname/path mismatch.

Product changes

• Model comparison page -- Compare up to five models side by side on pricing, context length, and benchmark scores with the new comparison tool, including a "Highlight best" toggle to call out the top value in each row.
• Redesigned model detail API tab -- The API tab on each model page now walks you through numbered steps (get an API key, make a request, enable streaming) with an endpoint reference and interactive parameters table.
• Rankings daily dataset endpoint -- A new GET /api/v1/datasets/rankings-daily endpoint returns the top-50 models by daily token volume, matching the data behind the public rankings chart. SDK reference
• IP allowlist enforcement -- API keys configured with an IP allowlist now actively block requests from non-allowlisted IPs with a 403 response, upgrading the previous observe-only mode. Docs
• Auto router cost_quality_tradeoff parameter -- The auto router now accepts a cost_quality_tradeoff integer (0–10) instead of the previous binary tradeoff, giving finer control over the cost-versus-quality balance when routing requests. Docs
• Fixed: dashboard not refreshing on account switch -- Switching between accounts in the dashboard now correctly remounts the page so data reflects the selected account.
• Fixed: Bedrock BYOK region prefix mismatch -- Bring-your-own-key Bedrock requests now normalize the inference profile region prefix to match the customer's configured region, fixing "invalid model identifier" errors when the endpoint and key regions differed.

Product changes

• Activity and Logs in workspace sidebar -- Organizations with multiple workspaces now see Activity and Logs links directly in each workspace's sidebar, pre-filtered to that workspace.
• API key links in activity logs -- The Activity page and generation detail pane now display the API key name as a clickable link to its detail page, with org-scoped access control.
• Fixed: video playback errors in Logs -- Video generation results in Logs now show specific error messages based on the failure type (expired, codec unsupported, network error) and always provide a Download button as fallback.

Product changes

• Fixed: activity chart color consistency -- Spend, Requests, and Tokens charts on the activity page now use consistent colors for the same series across all three charts.

Product changes

• Compare page benchmark visualization -- The compare page now displays provider-coded landscape benchmark charts for Intelligence, Coding, and Agentic metrics with author-brand color gradients, replacing the previous data grid.
• Redesigned model selector -- The model selector now features full-screen search, improved filtering, and integrated apps discovery, making it faster to find and switch between models across the site.

Product changes

• Profile link in navigation -- Added a Profile link to the user dropdown menu for quicker access to profile settings.
• Compare page toolbar layout -- Moved the comparison controls to the left of model cards and the "Add model" button to the top-right title bar on the compare page.
• Fixed: service_tier response normalization -- API responses now consistently return "default" instead of sometimes "standard" for the base service tier, and explicitly return null when no tier is present. Docs
• Fixed: profile API keys scoping in organizations -- API keys on the profile page now correctly display only the current user's keys when viewing within an organization context.

Product changes

• Export all chats from playground -- Added an "Export All Chats" option to the chatroom sidebar menu to download all conversations as a single JSON file.
• Filter menu badge counts -- The activity filter menu now displays numeric badge counts for active selections per category and uses checkmarks instead of dots for selected items.
• API keys on profile page -- All API keys across workspaces are now visible on the user profile page with click-through navigation to each key's detail page.
• Improved coding agent attribution -- Requests from Cursor, GitHub Copilot, Cline, RooCode, Kilo Code, Zed, and OpenCode are now properly identified and labeled in activity logs.
• Fixed: chat export missing rooms -- Exporting chats now includes all rooms, including those not recently visited.

New models

• Qwen: Qwen3.7 Max

Product changes

• Web fetch: Parallel engine -- The web fetch server tool now supports Parallel as an engine option, bringing engine parity with web search.
• Delete all chats in chatroom -- Added a "Delete All Chats" option in the chatroom sidebar with a confirmation dialog for bulk-deleting all chat rooms at once.
• Pareto Router default settings -- The Pareto Router can now be configured with a default coding quality tier in workspace plugin settings, removing the need to pass min_coding_score on every API request.
• Fixed: HTML-encoded characters in xAI tool-call arguments -- xAI Grok models that intermittently emit &, <, or > inside tool-call argument JSON now have those entities decoded server-side so downstream clients receive clean arguments.
• Fixed: compare page horizontal scroll at 4–5 models -- The compare page now properly enables horizontal scrolling when comparing four or five models.

New models

• xAI: Grok Build 0.1

Product changes

• Download fusion responses as markdown -- Fusion responses in the chatroom can now be downloaded as .md files with human-readable filenames based on the prompt title and timestamp.
• Profile page improvements -- User profiles now show per-model breakdowns for token and prompt counts, a discoverable profile picture change button, and a refreshed header layout.
• Request ID filter in logs -- The logs page now displays request IDs in generation details and supports filtering by request ID alongside generation ID and session ID.
• Unified date range picker on activity page -- The activity page now uses the same date range preset picker as the logs page for consistent date selection.
• Collapsible citation icons in chatroom -- When a chatroom response has more than five citation icons, only the first three are shown with an expander button to reveal the rest.
• Model pinning for router sessions -- Router models like openrouter/auto now pin to the same concrete model across requests sharing a session_id, improving prompt cache hit rates for multi-turn conversations.
• Fixed: router fallback blocked by generation start events -- Protocol-level generation start events no longer prevent provider fallback, so the router correctly retries with another provider when the first one errors before delivering content.

New models

• Google: Gemini 3.5 Flash

Product changes

• Move API keys between workspaces -- You can now transfer API keys from one workspace to another, making it easy to reorganize keys as your team structure evolves.
• Presets API for creating presets from inference requests -- Added POST /api/v1/presets/:slug/chat/completions for creating or versioning a preset directly from a standard chat completions request body, now also available in the TypeScript and Python SDKs.
• Date format preference -- Added a date format selector in Settings with Default, Short, ISO, and Relative options that apply across the dashboard, logs, and API key pages.
• Long-horizon agents cookbook -- New cookbook guide covering how to build long-running agent tasks with the OpenRouter Agent SDK, including stop-condition ceilings, resumable state, streaming progress, and voice input.
• Compare page empty state redesign -- The compare page now shows interactive slot cards in the empty state, letting you select models directly without a separate search step.
• Profile page layout -- Profile pages now display name, avatar, email, role, and workspace info at the top of the page for a cleaner layout.
• Fixed: Anthropic native web_fetch passthrough -- Anthropic models with native web_fetch support now correctly receive the tool in upstream requests, restoring full web content fetching capability.
• Fixed: compare page section headers not sticking on scroll -- Section headers on the compare page now correctly dock under the model cards while scrolling.
• Fixed: playground side panel losing generation state on close -- Closing and reopening the playground side panel on model detail pages now preserves in-progress generation output.

New models

• xAI: Grok Imagine Video
• xAI: Grok Imagine Image Quality

Product changes

• Fixed: rankings filter state not persisted in URL -- Filters on the rankings page are now stored in the URL via query parameters, making filtered views shareable and bookmarkable.

Product changes

• Bring-your-own-key management API -- New REST endpoints for programmatically listing, retrieving, creating, updating, and deleting BYOK credentials across workspaces via management key authentication.
• Observability destinations management API -- New REST endpoints for listing, retrieving, creating, updating, and deleting observability destinations (Datadog, Langfuse, LangSmith, and more) via management key authentication.
• Fixed: dashboard growth percentages exceeding display bounds -- Very large growth percentages on the activity dashboard are now capped at >999% for readability.

Product changes

• Cursor integration guide -- Added a guide for configuring Cursor to route through OpenRouter via the Override OpenAI Base URL feature.
• Fixed: /api/v1/models under-reporting context_length -- Models whose cheapest endpoint has a lower context window than the model itself (e.g. moonshotai/kimi-k2.6 showing 32K instead of 256K) now correctly report the full model context length in the models API.
• Fixed: web search unbounded context growth -- The web search server tool now defaults max_total_results to 50 when not specified, preventing runaway token accumulation in agentic loops. Docs
• Fixed: Mistral speech-to-text billing unit -- Corrected the billing unit for Mistral speech-to-text models from per-second to per-minute to match the provider's actual pricing. Docs

New models

• Mistral: Voxtral Mini Transcribe

Product changes

• Playground side panel on model pages -- Model detail pages now include a "Playground" button in the header that opens a slide-out panel for testing the model inline, with support for all available modalities.
• OpenCode integration guide -- Added a guide for configuring OpenCode with OpenRouter to the coding agents documentation.
• Fixed: BYOK model filter showing incomplete results -- Bring-your-own-key provider pages now show all models a provider serves, not just models whose cheapest endpoint is from that provider.
• Fixed: 500 errors for router model variant mismatches -- Requests to router models with unavailable variants (e.g. openrouter/auto:thinking) now return a descriptive 404 instead of a generic 500 error.
• Fixed: vertical scrollbar on models page modality tabs -- Removed an unwanted vertical scrollbar on the models page modality filter tabs.

New models

• xAI: Grok Voice TTS 1.0