Every model gets a Rival Score: 0 to 1000, from real human preference. No LLM-as-judge nonsense. Three components:

Every comparison runs the same automated pipeline. No manual intervention, no curation.

• Same prompt, every model. Each model in a challenge receives the exact same input. No special treatment, no warm-up round.
• Automated pipeline. Responses are generated via OpenRouter, Replicate,and direct provider APIs via scripts. No human selection bias.
• Controlled parameters. Pre-generated responses use temperature=0.7 and effectively unlimited output tokens via OpenRouter. GPT-5.1 models run at provider default. Every response ships a reproducibility card, so you can call us out. Full breakdown below.
• No system prompts for pre-generated content. User prompt only. What the model sees is exactly what you see in the generation details.
• One shot, no cherry-picking. One generation attempt per model. Never re-rolled. What the API coughed up is what ships.

Challenges are the prompts behind every comparison, built to surface real capability gaps.

• Capability-specific. Each targets one skill: coding, reasoning, creative writing, analysis, instruction following, visual output. One-trick ponies get exposed fast.
• Model-agnostic prompts. No provider-specific keywords, no prompt tricks that advantage one model family.
• High quality ceiling. Open-ended by design, so there's always room for a better answer. A model ten years from now should still beat today's.

Rankings on the arena use a weighted score from real user activity, not synthetic benchmarks. No GPT-4 grading other models.

• Challenge win rates. The primary signal. Each duel win adds to the model's score in that challenge's category.
• Category coverage. Models rank within each capability category. A stable ranking needs participation across challenges. No ducking the hard ones.
• Consistency matters. Steady across many challenges beats one standout. One viral moment doesn't make a career.
• One vote per user per duel. Anonymous and deduplicated. No ballot stuffing.
• Blind voting available. Blind mode hides names and providers until after the vote. People vote very differently when they can't see the logo.

Community votes power every ranking. We keep them clean. Trust issues are a feature, not a bug.

• Deduplication. One vote per model pair per challenge, enforced by a unique DB constraint. The database said no.
• Rate limiting. Authenticated users: 200 votes/hour. Anonymous: 30/hour, keyed by IP hash. Touch grass between votes.
• Voter fingerprinting. Anonymous votes tie to a hashed IP. Client IDs are hashed server-side too. We trust you. Just not that much.
• Minimum sample size. Models need 10 duels before they hit the leaderboard. Credibility further down-weights small vote counts.
• Server-side normalization. All IDs are validated and normalized on insert. Malformed data gets rejected. We've seen what the internet sends us.

Beyond the overall score, models compete in 15 category arenas, each scoped to one capability. Find the best model for your exact use case.

Category rankings use the same Rival Score formula, scoped to that category's challenges. Explore them on the arena

Rival Score is snapshotted daily. The leaderboard shows a 7-day trend per model. See who's heating up and who's quietly losing ground.

Models move fast. We track versions so comparisons stay meaningful. Nobody gets credit for last quarter's weights.

• Point-in-time captures. Responses are recorded when a challenge is created, with the model version stored alongside. Receipts kept.
• Version tracking. New challenges use the latest version. Old responses are never overwritten. We don't rewrite history.
• Historical preservation. Older outputs stay, so you can compare months ago versus today. Growth arc or regression arc, you decide.
• Pricing data. Model pricing on the prices page comes from LiteLLM, updated to current provider rates. Brace your wallet.

What's out of scope:

• No automated benchmarks. No MMLU, no HumanEval. Real people judge, not machines grading themselves.
• No cherry-picking. Outputs ship exactly as the API returns them. If a model fumbles, the fumble ships.
• No pay-for-ranking. Providers can't pay for higher rankings. Sponsored placements are clearly labeled and sponsorship never moves a duel. Sponsored models are labeled.
• No synthetic evaluations. Scores come from real votes. We don't let AI grade AI.

Rival tracks 200+ models across OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek, xAI, and more. If it has an API, we're probably already yelling at it.

• Broad coverage. New models go in as soon as they ship a public API. Full roster on the models page.
• Coverage varies. New models have fewer challenges. Rankings firm up as votes accumulate. The data will come.

Aggregated, anonymized response and voting data is public. Take it. Build on it. Prove us wrong.

Every response ships generation metadata. Hit the terminal icon on any card for the exact API params and a copy-paste command. Trust, but verify.

• One-click reproduction. Copy a cURL, Python, or raw JSON payload that mirrors the exact call. No detective work.
• Full parameter transparency. Temperature, system prompt, max tokens, model ID, and provider show on every response. We show our hand.
• OpenRouter as common layer. All text models are called through OpenRouter, giving one consistent API surface to reproduce against, whatever the original provider.