About

How this site uses AI

What runs live: nothing, unless you ask. Every default page — the races, the daily game, this page — serves recorded fixtures, so a default visit costs $0.00 in AI spend. Each recording is stamped with its capture date, and replayed numbers are the numbers from that recording, not simulations.

What runs when you opt in: the Arena’s “run it live” button streams real answers from real models. Before offering its live consent control, the page requires the live switch, budget ledger, nonblank session secret, both Turnstile keys, and positive effective daily budget and run limits; it does not preflight every downstream provider credential. The request then verifies a single-use Turnstile bot token. Every admitted request counts against both a signed browser-session quota and a hashed-IP quota; the code default is five admissions per UTC day on each. Before a provider call, Arena reserves the run’s conservative worst-case cost under a code-default shared $0.50 UTC-day ceiling. It sends each lane once through AI Gateway, with Gateway retries disabled. A Gateway or transport failure ends that lane as an error instead of buying a second attempt. An admission refusal closes that stream with its terminal event; a model declining one lane is a paid lane verdict and the rest of the race continues. Neither starts replay: the replay control stays available for you to choose. In the 2026-07-12 recorded pass a single race cost between $0.0001 and $0.0052 depending on the task, and the two apps’ captures together billed just under three cents. Those captures do not predict a later live run.

The models and their list prices (per million tokens, as listed by OpenAI and Anthropic, verified 2026-07-09): GPT-4.1 nano $0.10 in / $0.40 out · GPT-5.4 nano $0.20 in / $1.25 out · Claude Haiku 4.5 $1.00 in / $5.00 out · Claude Sonnet 5 $2.00 in / $10.00 out.

What’s exact and what’s estimated: OpenAI token counts on this page are exact — computed in your browser with the same encoding OpenAI uses (o200k). Claude counts are estimates, marked “est.” wherever they appear, because Anthropic doesn’t publish its tokenizer; I use a ×1.15 heuristic and you should read ±20% into it. Arena and game costs come from the providers’ own usage numbers, so those are exact.

What the scoring can’t see: Arena checks constraints — exact strings, patterns, schemas. It can tell you an answer followed the rules and what it cost. It cannot tell you an answer was good; on the haiku task, that’s deliberately your job.

How it was built: in three days, July 10–12, 2026, by one person who is not a career engineer, directing three models — one for planning, one for writing most of the code, one for review. I made the calls and I own the mistakes. The letter-counting race where every cheap model gets it wrong stays up because it keeps me honest too.

What I collect: no accounts, no cookies for tracking. Your game streak lives in your browser’s localStorage. Visit counts come from Cloudflare Web Analytics, which runs without cookies and doesn’t identify you.