Stay up to date with the latest features, improvements, and fixes in RouterBench.
gpt-image-2 image-to-image edits now work end to end — supplying an input image no longer returns HTTP 500. The gateway routes image-edit requests to OpenAI's multipart edits endpoint. Text-to-image generation is unchanged. The unstable routerbenchQwopus3.6-27B model has been retired.
Read more →Generate music through RouterBench. Two Google Lyria 3 models (Clip and Pro) are live behind a new synchronous /v1/music API — text-to-music with optional reference images, custom lyrics, and an instrumental toggle. Flat per-track billing, a dedicated Music Studio in the Playground with history, and 30-day media storage that auto-expires (your chat and usage history are untouched).
Read more →Generate video through RouterBench. Six Google Veo models (3.1, 3.1 Fast, 3.1 Lite, 3, 3 Fast, 2) are live behind a new async /v1/videos API — text-to-video, image-to-video, first/last-frame interpolation, up to 3 reference images, and video extension. Per-second metered billing, plus a dedicated Video Studio in the Playground.
Read more →Three new image-generation families are live — OpenAI gpt-image-2, Google Imagen 4 (Fast / Standard / Ultra), and Nano Banana Pro. Gemini catalog updated against Google's live model list. Bug fix: gateway no longer leaks OpenAI-only fields to Gemini.
Read more →Added Moonshot's new Kimi K2.6 flagship (image + video + reasoning, 262K context) with direct Moonshot pricing. Fixed Kimi K2.5 on Moonshot — reasoning capability was previously unflagged in the catalog even though the model supports it. Removed the Groq mapping for Kimi K2 — Groq no longer serves moonshotai/kimi-k2-instruct.
Read more →Added 1h, 5h, Today (IST midnight), and 3-day filters to the admin panel's Top Users page so ops can spot spikes and incidents without scanning the full 7/30/90-day view.
Read more →Marked the Neysa mapping for qwen3.5-122b-a10b and qwen3.5-397b-a17b as unstable and added Groq (qwen/qwen3-32b) as the working provider. Auto-selected Qwen3.5 requests now go to Groq until the Neysa backend is restored. Also removed a redundant duplicate mapping — it shared the same offline backend.
Read more →Kimi K2.5 now defaults to Moonshot direct after Fireworks capacity issues caused 429 'service is overloaded' errors. Fireworks remains available when explicitly requested.
Read more →New Fireworks models (Qwen 3P6 Plus, Kimi K2.5), free Gemma 4 models with reasoning, removed deprecated Gemma 3 models, and critical fixes for Moonshot and Fireworks provider handling.
Read more →