Tools · Cost of a finished task
What does a finished task cost?
Every AI pricing page on the internet prices the token. None prices the job. Between the two sits the harness — the scaffolding around the model that decides how often a task lands, and therefore how many attempts you pay for. Put a success rate in and the bill changes shape.
One working day at a time. An agent working in your codebase most of the working day.
Best generally-available intelligence — the hard call, the planner.
Compressed context with retrieval, structured retry with backoff, basic feedback into the loop.
Success rate per attempt — use your own
The two boxes below are the whole argument. They open on a borrowed number from a 100-task coding benchmark. If you have run these models on your own work and know how often each one lands first time, type that in — everything above is scaffolding for these two figures.
Left at zero, so the figures below are tokens only. Set both to price the working days that never land.
Cheapest per finished working day
£8.52
Gemini 3.1 Pro · £4.94 of tokens per attempt, 58.0% of which land
The other one
£37.48
Claude Fable 5 · £21.74 per attempt
What a pricing page would say
Gemini 3.1 Pro at £4.94 per working day — the cost of one attempt, quoted as though every attempt lands.
Where the cheaper model stops being cheaper
Holding Claude Fable 5 at 58.0%, Gemini 3.1 Pro stays the cheaper choice per finished working day all the way down to a 13.2% success rate per attempt. Below that line the dearer model is the cheaper one, and no price sheet will tell you so.
Copper line: Gemini 3.1 Pro, cost per finished working day as its success rate moves. Dashed line: Claude Fable 5 held at 58.0% — £37.48. Top of the scale is £49.43.
How this was worked out
One working day is 2.0000M tokens in and 0.4000M out, with 60% of the input served from cache. That is the month’s stated assumption — 20 working days · heavy context re-read each turn · ~40M tokens in, ~8M out per month, much of it cached — divided by the 20 working days it contains.
At a success rate of p, a finished working day takes 1/p attempts on average — and that figure does not move when you change the retry cap. Raising the cap does not cut your token bill; it cuts how many working days get abandoned to a person. Human time is the only thing “attempts before you give up” touches.
Default success rates are Zhang et al., arXiv:2605.23950, Table 2 — the mean of the three models tested across 3 models × 3 harnesses, 100-task SWE-bench Verified subset, 2 runs per cell. They are coding-task results, and MenFem did not measure them. If you have your own measured rate, use it: it beats a borrowed one every time.
In that experiment, changing the harness moved results 7.80× as much as changing the model, and flipped which model won in 6 of 9 pairs. The direction and the reversals are the robust findings. The 7.80× itself is a within-study measurement on 100 coding tasks over 2 runs, and it is not a general constant — the authors decline to treat it as one, and so does this tool. Do not carry it to other kinds of work.
The tested cells behind the Improved default: GPT-5.4 58.5% · Kimi K2.6 59.0% · GLM-5.1 56.5%. Read the paper before you lean on any of it.
One limit worth naming: that experiment measured success rates, not token consumption. A fuller harness plausibly spends more tokens per attempt — verification and self-checking are not free — and this instrument does not model that, because the source does not measure it. The Full-harness figures here are, if anything, flattering.
Prices are each provider’s own published list rates, read from their pricing pages. Converted at £0.7445 per $1 as of 2026-08-04. Your real bill will differ — this is the assumptions above applied to today’s list prices, not a quote.
What this changes
A price per token is not a price. It is the price of one attempt, quoted as though attempts always work. The number you actually pay is the price of one attempt divided by the share of attempts that land — and nobody selling you tokens has any reason to mention the divisor.
Raising the retry limit does not cut your token bill. It cuts how many jobs end up on a person’s desk. The tokens per finished job are 1/p whatever the cap, because the extra attempts you buy are exactly the attempts you were already going to spend on the jobs that eventually land. What the cap buys is fewer abandonments, and abandonments are priced in human minutes.
The ranking can flip. A model that is ten times cheaper per token survives a landing rate ten times worse before it loses. Push it past that line and the expensive model is the cheap one. This is not a hypothetical: in the controlled experiment this tool’s defaults come from, changing the harness flipped which model won in six of nine comparisons.
The number you want is your own. The defaults here are borrowed from a 100-task coding benchmark and they are the weakest thing on the page. Run your own work through both models, count what lands, and type those two figures in. Everything else is scaffolding for that measurement.
Where the defaults come from
One paper: a controlled 3×3 factorial crossing three models of comparable capability with three harness configurations, over a 100-task SWE-bench Verified subset, two runs per cell. It is a position paper, not a peer-reviewed result; the sample is small; and every task in it is a coding task. Its authors decline to treat their headline variance ratio as a general constant, and neither does this instrument. MenFem measured none of it — we are reading the paper, and we say so on the calculator itself, beside every figure it lends us.
What travels is the direction and the ranking reversals. What does not travel is the coefficient, and what certainly does not travel is any of it to work that is not coding. The full read, including the limits, is in the harnesses knowledge base.
Embed this tool
Paste this snippet into any page to run the calculator there.
Prices are each provider’s own published list rates, read from their pricing pages and kept as a dated series (latest reading 2026-08-03). This is list pricing, not your negotiated rate. For the monthly bill without the harness layer, see Cost of the Job; for the investor-facing view of the same dataset, the Token Price Index.