Every AI answer on this site comes with a receipt.

Model races, a daily guessing game, and a tokenizer you can poke — priced to six decimal places. Your visit costs $0.00: the defaults replay recordings, and live runs are opt-in and capped.

Model Arena — one prompt, three or four models, every answer priced to six decimals; the winner is arithmetic (correct → cheapest → fastest). No voting, no vibes: a deterministic, priced race over recorded challenges. Opens on letter-count, the race the priciest lane got right. Guess Which AI — one odd prompt, three anonymous answers, some days a fourth written by me. Name the model; the reveal prints each answer’s model, capture date, and exact price.

What does text cost?

Change this number to the reply length you expect. It is an assumption, not a prediction.

Receipt uses 500 output tokens.

TOKEN RECEIPT2026-08-18
GPT-4.1 nano
input — 0 tok × $0.10/MTok $0.000000
assumed reply — 500 tok × $0.40/MTok $0.000200
GPT-5.4 nano
input — 0 tok × $0.20/MTok $0.000000
assumed reply — 500 tok × $1.25/MTok $0.000625
Claude Haiku 4.5
input ~ 0 tok (est.) × $1.00/MTok $0.000000
assumed reply — 500 tok × $5.00/MTok $0.002500
Claude Sonnet 5
input ~ 0 tok (est.) × $2.00/MTok $0.000000
assumed reply — 500 tok × $10.00/MTok $0.005000
TOTAL for one request to all four$0.008325
input: your text · output: your visible 500-token starting assumption · OpenAI input counts exact (o200k) · Claude input counts est. ±20% · prices: provider list, verified 2026-07-09
How these numbers are made

Tokens are the chunks a model actually reads — about three-quarters of an English word each, so 100 words is roughly 130–140 tokens.

OpenAI rows are exact. Your text is encoded in your browser by the gpt-tokenizer library using o200k_base, the same encoding OpenAI’s models use. Same input, same count, every time.

Claude rows are estimates. Anthropic doesn’t publish its tokenizer, so there’s nothing exact to run. I multiply the o200k count by 1.15 and mark it “est.” — read ±20% into it. Where this site has real Claude numbers — Arena races and Guess reveals — they come from Anthropic’s own usage reports, and those are exact.

The math on every line: input tokens × input rate ÷ 1,000,000, plus assumed output tokens × output rate ÷ 1,000,000. Each line is rounded to six decimals at display — recompute it by hand. Rates per million tokens (MTok), 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.

Reply length is your assumption. The field starts at 500 output tokens and stays visible beside the receipt. Change it to zero for an input-only price, or use the reply length you expect. The site does not predict it.

Nothing you type leaves the page — the counting runs in your browser. The headline’s own token boundaries are computed once, when the site is built. Your text is tokenized live, on your device.

That tokenizer is the live demo — exact OpenAI counts in your browser, Claude counts labeled as estimates. All projects — seven builds, this tokenizer included; receipts where the facts have landed.