OpenRouter

OpenRouter is the one aggregator key that can hold almost the whole of Brut: every text model (planning and Marcel — chat, ambient narration, acting), the full Nano Banana image family (incl. Nano Banana 2 Lite, the starter graph's model), GPT Image 2, GPT Image 2.5 Flare and Sunburst, the Seedream 5.0 trio (Pro / Flash / Lite at 1K / 2K, Lite 2K / 4K), and the video models — Kling 3.0 and O1, Veo 3.1 / Fast / Lite, Seedance 2.0 and 2.5. Outside it: the upscalers (they aren't resold anywhere, by design), Gemini Omni Flash and the other Kling tiers (3.0 Omni / Turbo / 2.6 / 2.5 and older), which OpenRouter doesn't carry. Like every route, it adds no models to the menu and is always visible: routed nodes show ⇄ runs via OpenRouter, and Marcel's composer says so too.

If you want the simplest possible seat: one OpenRouter key is a working studio. Native keys added later always take their models back automatically, usually at slightly better prices.

#Text — planning and Marcel

Every text model in Brut's catalog also runs here, at pass-through provider list rates — the same per-token prices as the native APIs (OpenRouter's revenue is a ~5.5% fee when you buy credits, not a token markup). Planning that reads your pictures works, the GPT reasoning dial rides along (including while Marcel acts on the canvas), and Anthropic prompt caching passes through — Marcel's long system context caches exactly as on a native Anthropic key, so routed conversations cost the same as native ones. New models usually arrive here too.

#Images — native knobs, no compromises

The Nano Banana family runs with Brut's own menus — the real 512/1K/2K/4K sizes and the full aspect list. Approximate inpaint masks work (the Gemini convention). Billing is pass-through: every routed image model bills the same as on the maker's own key, to the tenth of a cent. Nano Banana Pro costs $0.13 at 1K or 2K and $0.24 at 4K; Nano Banana 2 costs $0.07 / $0.10 / $0.15 at 1K / 2K / 4K; Lite $0.034 and the original Nano Banana $0.039. So the estimate before a run is the real price, and every image job records its exact billed amount on the ledger.

Two OpenRouter-specific limits worth knowing:

• Very large pictures may be sent lightly compressed for that one request; your files are untouched. If even that can't fit, the run stops with a clear message; export the source smaller or use a native key.
• Heavy 4K image jobs can be slow. OpenRouter's queue sometimes stalls on large 4K multi-image jobs (a 4K edit that takes ~55 s on a native Gemini key can take much longer there). Brut gives up after about ten minutes. If that happens: retry, pick a smaller size, or use the native key, which takes over automatically the moment it's configured.

The one exception is GPT Image, which OpenRouter renders at about 1K only: there's no size knob, so 2K and 4K need your OpenAI key. The aspect ratio and the quality do come through, on GPT Image 2 and on the GPT Image 2.5 pair alike: 2.5 keeps its full low to max quality menu, and image-to-image with aspect auto follows your input's shape. Masks don't: inpainting on GPT Image needs the OpenAI key. Measured: a 1:1 image costs $0.006 at low and $0.053 at high on 2.5 (the same on Flare and Sunburst), and 16:9 costs less ($0.004 at low).

#Video — the whole set, with billing truth

Kling 3.0 (std/pro) and Kling O1 (5 or 10 s, $0.112/s), Veo 3.1 / Fast / Lite, Seedance 2.0 and Seedance 2.5 (480p/720p only, 4–30 s, first + last frame, no reference images there — the 30-reference and 1080p features need the BytePlus key) — image-to-video with end frames, text-to-video, reference images where OpenRouter offers them. Per-second rates, audio included in the estimate:

The listed rates hold: a 4 s Veo 3.1 clip bills $0.80, and $1.60 with audio. Seedance comes out close to the BytePlus price. Kling O1 renders at 1924 × 1076. Kling 3.0 renders at 720p on OpenRouter, pro included, with no 4K: routed nodes offer only std and pro. And OpenRouter reports the exact billed cost on every job, which lands on the ledger row as is.

#Get a key

Create a key at openrouter.ai/keys (sk-or-…), buy a few dollars of credits, and paste it into ⚙ Settings → OpenRouter.

#The fine print

• Your native key always wins. Add a Gemini/OpenAI/Kling/… key later and those models bill direct again automatically.
• Marcel works through OpenRouter only (among aggregators).
• Kling's 4K master tier, Seedance's Smart Length, Seedance 2.5's 1080p and reference images aren't on OpenRouter; Kling O1 takes only 5 or 10 s there. A missing setting always stops the run with a clear message, never silently.
• Every routed call — text, image, or video — lands in the spend ledger with provider-reported figures (token counts, cache activity, billed cost), marked "via openrouter".