About and methodology

The daily process

1. We write one challenge, usually coding or web building, with success criteria decided before any model runs.
2. Every participating model receives the exact same prompt and starting materials. The prompt is published exactly as given — whitespace and indentation included — with a SHA-256 fingerprint of its UTF-8 text (line breaks recorded as LF).
3. Models are run outside this website by the editor. For each run we record the model identifier requested, the resolved version when the provider exposes one, the tool or harness used, settings, limits, attempts and any human intervention.
4. We publish the outputs, screenshots, videos and measurements. You vote, and the community decides the result.

How the result is decided

The community decides. Each battle's poll is open for 48 hours; while it is open, the side with more votes is shown as leading. When it closes, the side with more votes wins. Nobody at AI Battle picks the winner, and we do not use an AI judge or a weighted score.

• Winner: more votes than every other entry; the others are recorded as losses.
• Draw: the top entries have exactly the same number of votes.
• No winner: nobody voted. Every entry is unscored — this does not mean everyone lost.

Every entry is also checked against the same fixed rules (does it load, animate, follow the format), and its time and cost are reported. These are facts shown next to the vote, not a score. The editor writes the rules, runs the battles, and removes abusive votes; if an infrastructure problem makes a comparison unfair, the battle is withdrawn.

Leaderboard and statistics

The leaderboard has one row per model version, ordered by wins, then win rate, then scored battles. Win rate is wins ÷ (wins + losses + draws). Equal records share a rank. Versions with fewer than 5 scored battles are marked provisional.

Averages for cost and time include only runs where the value is known — including failed runs — and always show how many runs were measured. Unknown values are shown as “Not available”, never as zero. Model pages combine versions of the same model line; a battle is counted once even if two versions took part.

Versions and measurements

When a provider exposes only a changeable alias or a product label, we record that label and the date of the run and mark it as an alias. We never invent a pinned version. Costs are attributable run costs marked as reported or estimated; subscription fees are not presented as precise per-run costs. Token counts follow each provider's own accounting and may not be directly comparable. Elapsed time is wall-clock time from submitting the task to the final output, including tool execution and retries.

Community voting

Each battle has an anonymous poll — “Which result do you think is best?” — open for 48 hours after publication. One vote per browser per battle, protected by a bot check and rate limits. Clearing cookies or switching devices can bypass the one-vote limit, so the poll measures public preference, not a verified or blind evaluation. The poll decides the battle's result and its place in the standings. Votes identified as abusive may be invalidated; this is logged, and the result is recalculated.

Limitations

A battle is one observed run under stated conditions. It does not prove that one model is better than another in general. Results depend on the prompt, the tools, the settings and chance. Read the evidence and draw your own conclusions.

Corrections and withdrawals

If we get something wrong we correct it in the open: corrected battles show the date and a summary of every correction, and the previous version is retained in our records. We never silently re-run a model and replace its original result. If an error invalidates the comparison, the battle is withdrawn: its page explains why and it stops counting toward any ranking.

Usage-limit resets

Reset times come from two independent public trackers, with thanks: codex-resets.com for Codex and clauderesets.com for Claude. We show their records unedited and compute the same statistics for both. The “beg for a reset” button is inspired by codex-resets.com. AI Battle is not affiliated with either tracker, OpenAI or Anthropic.

Every 10,000 different people who beg between two resets, we post one public message to the company, tagging only its official company accounts, and link it from the resets page. Each browser counts once per reset, after a bot check. Nothing is sent to the companies automatically or privately.

Independence

Any sponsorship on AI Battle is clearly labelled. Sponsors buy placement only. They have no access to results before publication, no say in challenges or judging, and no influence on rankings. If a sponsor is also a battle participant, we disclose it.