Closed beta: top-ups come with 30% extra credits, subscriptions with 50–70% extra. Feedback is welcome on the support page.

Pricing

Quoted first.
Charged after.

Priced on the result. A free estimate comes first — a static scan of seconds, in which none of your code runs. Only continuing takes the 2-credit analysis deposit, credited in full against the patch. You decide whether to pay only once it is confirmed faster end to end in your own program.

Trial pack
$3
10 credits · $0.30 / credit
  • The best rate we offer, once per account
  • ≈ 2–3 full runs on your own repository (a run takes about 2–5 credits)
  • Credits valid 24 months
  • Private code welcome (zip upload)
Usage pack
$5 +
$0.77 / credit · +30% in beta · pick an amount
  • Credits valid for 24 months
  • Private repos, results kept private
  • Buy more as needed, no commitment
  • A run takes about 2–5 credits ≈ $1.5–4, depending on code size, complexity, what the run finds and the speedup
  • Cards and WeChat Pay, any amount; WeChat is a desktop QR scan
Pro +50% credits · beta
$28 / mo
42 credits · $0.67 / credit
  • 50% bonus credits per dollar during the beta
  • ≈ 10–15 full runs per month
  • Private repos · 5 concurrent
Max +60% credits · beta
$99 / mo
158.4 credits · $0.63 / credit
  • ≈ 40–60 full runs per month
  • Private repos · priority queue · 5 concurrent
Max+ +70% · beta
$199 / mo
338.3 credits · $0.59 / credit
  • ≈ 85–130 full runs per month
  • Top queue priority · 5 concurrent
  • For heavy individual use: big repos, many runs in flight
Beta bonus · top-ups +30% · Pro +50% · Max +60% · Max+ +70% Subscription credits reset each period · packs last 24 months Packs: WeChat (desktop QR) / card · subscriptions: card only · WeChat buyers take the 30-day pack, same price, same bonus

What a run costs

Compute always at 50%, tokens at 5–10% by speedup

Metered on the compute and tokens actually burned, against public list prices: the GPUs at Beam's or Modal's list, the tokens at Anthropic's own API price list. The machine is hardware we rent and costs the same whatever the run finds, so it is always 50%; a run that never touches a GPU (submitted as CPU only) has no machine line at all; the tokens are the agent's work, and the speedup says what that work was worth. Below: one real RTX 4090 run, 2.8 hours, 1.61× end to end, whose tokens land in the 8% tier.

ComputeGPU + sandbox, by the second
5.5 × 50% = 2.8
Model tokensreading, changing, judging
23 × 8% = 1.8
At list ≈ 28.7 creditsYou pay ≈ 4.6 credits
listyou pay
5%speedup under 20%
6%20%–50%
8%50%–100%
10%100% and up

What you pay for tokens, by the end-to-end speedup the run delivered — the run above is 1.61x and lands in the highlighted tier. It is the number at the top of your report, which you can check yourself. Under 1.10x the patch is not sold: the deposit comes back and nothing is charged.

When money moves

1 · Free estimate
0
A static scan of seconds; none of your code runs, no GPU is held. You get a projected speedup and its range, then decide.
2 · Continue
2 credits
The analysis deposit: reading the repo, building the harness, timing the baseline. Credited in full against the patch; returned when no speedup is found.
3 · Take the patch
quote − 2
Only after the speedup is confirmed end to end in your own program. The quote is actual usage, and you never pay more than it.
Target missed (< 1.10×)No patch is sold, the report is yours, and the deposit comes back.
The failure is oursInfrastructure, timeout, redeploy, an environment that will not build, an entry we cannot find, VRAM — the deposit comes back.
Skipped the estimateThe deposit is taken at submit; cancelling within 5 minutes still returns it.
SubscribersPro waives 1 deposit per period, Max / Max+ 3. Stopping at the estimate is always free.

When 4.6 credits pay for themselves

The patch is bought once; the speedup keeps running. 1.61× = 62% of the machine time, so every hour keeps 38% of what it used to cost you. At public cloud rates:

RTX 4090$0.7 / h · saves $0.27
17 hours to pay back
RTX 5090$1.1 / h · saves $0.42
11 hours
H100$6.9 / h (AWS) · saves $2.61
2 hours

A job that retrains 20 hours a week saves more H100 time in its first month than the patch cost.

Billing rules

When you're charged, and when you're not

EstimateFree; none of your code runs.
Continue2-credit deposit, credited in full against the patch.
UnlockThe quote, which is actual usage with no markup; never more than it.
Target missedNo patch sold, report still yours, deposit returned.
Our failureThe deposit returns to your balance.
OrderThis period's subscription credits first, then packs, oldest first.

Speedup = the wall clock of your own command end to end, optimised against the frozen baseline; the report breaks it into startup, the key loop and what follows.

Start with the numbers. Decide after.

Questions? Email us