What does your agent really cost?

Price one agent task end to end: the attempt, the retries, the runs that fail. Then see what the same workload costs on every model we track, and whether it beats doing the work by hand.

Official list prices · — Retries included Failures included No sign-up

Your workload

Primary model Loading list prices…

Tokens per task Input plus output tokens for one attempt, before retries. A 250K budget is a medium coding task; chat and support agents usually run 20–40K.

Start from a published scenario

One click loads tokens per task, success rate and retry overhead from a published project. Published dollar costs usually assume heavy prompt caching — tune the cache slider to match them.

Input / output split

75% in · 25% out

Retrieved context and tool results are input tokens. Generated text and code are output, and output usually costs 3–5× more.

Prompt-cache hit rate

Share of input tokens served from the provider cache — system prompt, tool schemas, repo context. Billed at the cached rate.

Retry overhead

10%

Extra attempts caused by tool errors, loops and failed validations. Each one is billed in full.

Tasks per day Monthly figures assume 30 days at this volume.

Agent success rate

85%

Attempts that finish without human rescue. Failed runs still burn tokens, so they raise the cost of every solved task.

Human baseline used for ROI

Minutes per task

Loaded cost / hour Salary plus overhead, USD.

Fixed monthly agent cost Evals, observability and integration upkeep, amortised per month. Used for break-even and net savings — set 0 to ignore.

Cost per solved task

—

all-in, including retries and failures

—

Cost per attempt

—

Monthly spend (30d)

—

Solved tasks / month

—

Tokens / month

—

Blended $ / 1M tokens

—

Monthly savings vs human

—

ROI vs a human baseline

45 min @ $65/h

Human cost per task—

Agent cost per solved task—

Savings—

Break-even volume—

Break-even is the number of solved tasks per month needed to repay the fixed monthly cost at the current saving per task. Failures are already priced into the solved-task cost.

Same workload, every model

Provider comparison

Meter your own runs

#	Model	Provider	Tier	Blended $/1M	$/attempt	$/solved task	30d spend
Loading list prices…

— List prices per 1M tokens, snapshot —; cached input is billed at the provider's cached rate where one is published. Ranked by all-in cost per solved task for the workload above.

Method

The arithmetic, in the open

Benchmark costs

Step 1 — attempt cost

—

Every input token is billed at list price unless it hits the prompt cache.

Step 2 — all-in task cost

—

Retries and loops are priced as extra full attempts.

Step 3 — solved-task cost

—

Failures don't vanish: dividing by the success rate spreads their spend over the runs that worked.

—

Not included: sandbox and VM time, evaluation harnesses, human review, latency and outages — real line items, none of them token-priced. Meter your own runs with the open-source SDK → instead of modelling them.

Your agent's real numbers beat our model

Submit a project with a public source and we will price it, verify it and put it on the board — or meter it automatically with the SDK.

Submit your agent Read the SDK docs