Docs

Everything you need to start comparing AI models in 30 seconds.

The 30-second tour

  1. Open the app.
  2. Type a question.
  3. Pick the models you want to compare (we've picked sensible defaults).
  4. Hit submit. Each model streams its answer in parallel; once they're in, you get a single rewritten “aggrai's answer,” a ranked list of final scores, and a per-model Aggr-Score radar.

That's it. The rest of this page is just background.

What you get back

Per-model answers

One card per model with the full response. Each card header shows runtime and token usage so you can see which model was fastest and which was most verbose.

aggrai's answer

When you select more than one model, a summary pass writes a single answer that draws on the strongest content across all the models. Above it, the final scores rank every model at a glance. Read just this section and you have a complete answer; the per-model cards are there if you want to see who said what. The summary comes back in the same language as your question.

Aggr-Score

A second pass scores each answer on five dimensions — accuracy, completeness, calibration (epistemic honesty), clarity, and insight — each the average of three specific rubric checks. Shown as a radar chart (each axis out of 10) with that model's overall Aggr-Score (out of 10) in the centre. Accuracy carries the most weight (30%); a confidently-wrong answer is capped at 4.0 regardless of how well-written it is. Each model's strongest dimensions are highlighted on its own radar, and you can click any radar (or the “+” button) to expand a plain-English breakdown of that answer's strengths and weaknesses. Scoring is done by an AI judge that never sees which model wrote which answer, but it is still a judgement — a useful tiebreaker, not gospel.

Factual & time-sensitive questions

Not every question needs a five-model debate. Simple factual questions (“2+2”, “capital of France”) get a single direct answer — no point burning the compute on a comparison everyone would agree on. And when a question depends on recent information, we usually run a live web search first and give the models the results, with the sources listed under the answer. When nothing was searched and the question looks time-sensitive, the answers open with a “This looks time-sensitive” note: the models are answering from their training data, so double-check anything that moves fast.

Tiers and limits

  • Free — our fast, lightweight models (Claude Haiku 5.5, GPT-4o Mini, GPT-6 Luna, Gemini 2.5 Flash, Mistral Small, and more). Up to 3 per comparison. No account needed.
  • Standard — every flagship model (Claude Sonnet 5.5, GPT-4o, GPT-6.1 Sol, Gemini Pro, Grok 4.7, Kimi K3, Qwen3.8 Max, the Codex + Devstral coding models, and more). Up to 3 per comparison.
  • Premium — adds GPT-6 Astra and the deep-research models (Claude Opus 5.5, Claude Fable 5.1, GPT-6 Astra Pro, DeepSeek v4 Pro, Kimi K2 Thinking, GLM-5.2, MiniMax M2.5, Qwen3.8 Max Prime, Grok 4.20 Multi-Agent, Nemotron 3 Ultra). Up to 5 per comparison.

The line-up changes as new models ship — see the full model catalog for what's live right now, and Pricing for plan details.

How questions are routed

We send your question in parallel to each selected model. We don't modify your question. The summary and the quality scores are worked out in a separate pass once the models have answered.

Caching

Common example questions from the landing page have their answers and summary pre-warmed in a shared cache, so they return instantly. Your own questions are never shared across users; asking the identical question with the same models again in the same tab replays your earlier answer.

Sharing comparisons

Each conversation gets its own short link (e.g. /app/c/8kQ2xz). Signed in, your conversations are saved to your account, so you can reopen them from Chat history on any device. To show one to someone else, press Share: they get a link to a saved copy with your exact answers and scores, no re-run needed. For a comparison, they can carry it on with whichever of its models their own plan includes; a Free viewer can't continue one that only used paid models.

API

We don't expose a public API yet. If you need one, get in touch and tell us your use case.

Roadmap

  • Public API
  • More interface languages (the models already answer in the language you ask in)

Still stuck?

Check Help for FAQs, or contact us.