Performance

What an AI jury costs, and how fast it is

The cost of each step, with and without proof of which model answered. Every number is either measured on this site or modeled from published figures, and labeled.

One ruling

$0.0059

5 jurors, model calls modeled · measured $0.0020 to $0.010

Time to a ruling

20 to 60 s

measured · plain code is under a second

Proving it

+$0.049

and minutes of delay modeled design only

Certify a repo

$0.269

12 rules, without proofs modeled

Running an ASC

Step	Assumption	Cost	Time	Note
Compile an ASC (once)	2 calls on anthropic/claude-opus-5.5	$0.084	not in the model	one time per contract modeled
Inscribe it with IQ Labs code-in	1,909 bytes, 3 Solana transactions	$0.0013 + IQ fee	not in the model	Solana base fee modeled. IQ's Solana fee is not published.
One ruling, budget jury of 5	1,500 tokens in, 300 out per juror	$0.0059	20 to 60 s	measured range on the site: $0.0020 to $0.010
Pay out on Solana	1 transaction	$0.0004	not in the model	base fee at the SOL price in the model modeled

Trade-off: a ruling is the modeled cost above and takes 20 to 60 seconds in the measured runs. Plain contract code is cheaper and faster. Wrap only the actions where a wrong answer is expensive.

Proving which model answered

Signatures prove which wallet voted, not which model ran. Anthropic's FlashProve prototype produces a cryptographic proof that a specific model produced a specific answer, without revealing the weights. Today it works for one open model, Llama 4 Maverick, on contexts up to 1,024 tokens. design only Proofs are not live on this site.

Step	Cost	Time	Note
Ruling by 5 Llama 4 Maverick jurors	$0.0024	20 to 60 s	the only model FlashProve has been shown on modeled
Prove each answer (5 proofs, 1,024-token contexts)	$0.026 to $0.080	about 9 GPU-seconds each; a batch of 24 takes about 223 s	H100 at $1.99 to $6.16 an hour
Verify the 5 proofs	$0.024	about 17 s each on 24 CPU threads	anyone can verify; cost shown at $1.00 an hour of CPU
Store the proofs	off chain		about 18.78 MB each, about 94 MB per ruling; only hashes go on chain

Overhead

About 72 times the GPU work of serving the same model normally, up to 94 times at the heaviest load tested.

Hardware per juror

one 80 GB GPU and about 290 GB host memory. Kept on all month: $1,453 to $4,497.

The catch

A proven jury today is five copies of one open model. Proof costs you the diversity of labs that makes the jury neutral.

Anthropic: Progress on provable inference · FlashProve technical report · FlashProve code

Certifying code

Review	Cost	Time	Note
Certify a repo, 12 rules, budget jury of 5	$0.269	not in the model	8,000 tokens of evidence per rule modeled
Same review, with proofs	$2.78 to $8.60 + $2.55 to verify	about 84 GPU-minutes	540 proofs: evidence split into 1,024-token checks, Maverick jurors only design only

A proven review is possible but costly today, because each proof covers at most 1,024 tokens and the evidence for a rule is about 8,800 tokens of input plus output. It gets cheaper as the context limit grows.

Try your own numbers

Jurors

Model tierTokens per juror

1,800

Prove answersGPU price per hour

$1.99

SOL price

$85.00

Rulings per month

10,000

Rules to certify

Cost per ruling

$0.0063

Per month

$63.29

Time per ruling

20 to 60 s

Certification

$0.269 model and proof cost

Flagship prices every juror as openai/gpt-5-mini. That jury was not recorded. Maverick is five copies of one open model. 10,000 rulings a month is a calculator input, not a measured volume. Proofs are design only.

All inputs and formulas are in data/performance-model.json. Model prices: OpenRouter catalog snapshot, 2026-10-11. SOL at $85.00 in the file. FlashProve figures from Anthropic's research post (Nov 2, 2026). As of 2026-11-04.