Different AI models are good at different things — one is faster and cheaper for an everyday question, another reasons better through a hard problem, another writes better code, another can search the live web. Most chat products make you pick a model yourself and hope you picked well. arbiter doesn't — you just type your message, and arbiter reads it and routes it to whichever model actually fits, every time. You're talking to software, not a person, at every point in the product.
Most messages are ordinary — a quick question, a rewrite, a bit of brainstorming — and go straight to a fast, inexpensive default model. arbiter only reaches for something more capable (and slower) when your message actually calls for it: real code, math, multi-step reasoning, a long document, a specialized domain question, or something that depends on current events and needs a live web search. You never choose any of this yourself — it happens before the reply starts streaming back.
11 models across 6 providers, each with a defined job. This list is generated from the routing table itself, so it is what is actually answering today.
claude-haiku-4-5-20251001Everyday questions — the default laneclaude-sonnet-5Code · Rewriting image prompts · Fallback for math, multi-step reasoning, long documents, structured output, specialist domain questions, uploaded filesclaude-opus-5Math · Multi-step reasoning · Specialist domain questions · Uploaded files · Fallback for code, long documents, structured output · Fallback for the frontier tiergemini-3.5-flashLong documents beyond about 200k tokens · Fallback for the frontier tiergemini-3.1-flash-image-previewFallback for image generationgpt-5-miniStructured output · Fallback for everyday questionsgpt-image-2Image generation and editingsonar-proCurrent events, with live web searchdeepseek-v4-pro"Delve further" middle rung for code, math, multi-step reasoning, specialist domain questionsdeepseek-v4-flashFallback for everyday questionskimi-k2.6Long documents up to about 200k tokensEvery reply has a Delve further button. If the default answer feels shallow, this re-runs your message on a more capable model — useful for anything that turned out harder than it first looked.
Describe what you want in plain language — arbiter rewrites your request into a proper, structured image prompt first (shown in the reply, and editable — you can tweak it and re-run before committing to a generation) — then actually generates it. Once you have an image, you can ask for a follow-up change in the same conversational way — “make it darker,” “make the dog bigger” — and arbiter regenerates with that change applied, without you having to repeat the whole description.
Image generation requires a free account (one real generation costs meaningfully more than an ordinary text reply, which is also why it's the one place with a visible daily limit — 3 a day on the free plan, 30 on paid). Everything else in arbiter works for signed-in and anonymous visitors alike.
A new visitor gets a handful of real messages (5, with the same automatic routing everyone else gets) before being asked to create a free account. Signing up carries over whatever you were already talking about — nothing is lost.
Each conversation is its own thread, listed in the sidebar. Start a new one anytime, switch back to an old one, or delete one you don't need — the sidebar can also be collapsed out of the way if you'd rather have the extra width.
Every message is screened before it reaches a model, including a tighter, more cautious pass for image requests specifically — image results are more visible and more shareable than a text reply, so that lane errs harder toward blocking. A message that reads as a safety concern gets crisis resources alongside arbiter's own response, not instead of one.
See also Terms and Privacy Policy.