Where does an AI agent spend its money?
Seven cost buckets across 33 recorded SWE-bench attempts. Stage shares, missing buckets and rounding, with list-price estimates rather than invoices.
Crewshal recorded $92.64 of notional model cost across 33 SWE-bench Verified attempts. Editing and running code took $44.05; research took $20.56.
Together, those two buckets account for 69.7% of the recorded total, a calculation on the raw costs. They are the largest buckets in this run. This does not measure the saving from changing either stage.
These were subscription CLI calls through Claude Sonnet 5.5. The platform priced the recorded tokens at list prices. Actual charges remain unknown here; these figures are not an invoice. They also do not measure customer delivery quality.
Seven recorded cost buckets
The chart covers all 33 attempts across the two published campaigns, including failed attempts and empty patches. It does not describe one chosen task.
The raw attempt records contain these buckets:
- Act (edit and run): $44.05, present in 32 attempt records.
- Research: $20.56, present in all 33 records.
- Verify: $8.72, present in 27 records.
- Other: $6.24, present in all 33 records.
- Review: $5.58, present in 32 records.
- Context compaction: $5.41, present in 32 records.
- Onboarding notes: $2.09, present in all 33 records.
These are counts of records with a cost bucket. They are not counts of passed checks or completed stages. A missing bucket does not prove zero spend or a completed check.
The published study and method name the campaigns and platform builds. Campaign 1 covered 25 attempts on build f0ac3a8a; campaign 2 covered eight on build 236c0d3f. The records cover October 4–5, 2026. This is historical evidence, not a new run of today's software.
Use the full denominator before rounding
The seven displayed costs add to $92.65. The chart's total is $92.64. Both follow from the same records: each bar rounds on its own.
The raw stage costs sum to $92.637510, which rounds to $92.64. Their sums also match each attempt's recorded cost before rounding. The stage call counts add to 1,632, the same number as the attempt records.
For a share, use the raw bucket cost divided by the full raw total. Act accounts for 47.5%; research accounts for 22.2%. Verify and review together account for 15.4%. Each percentage is a calculation, rounded to one decimal place.
Those shares describe this run. They do not show that removing verification or review would preserve quality. No matched experiment here isolates the effect of either change.
What these buckets cannot tell us
There is no separate Plan bucket in this chart. Do not assign it a cost of $0 or spread Other across missing stages. The chart cannot recover those amounts.
The extract lists calls, estimated cost and output tokens for each recorded bucket. It does not give input, cache-read or cache-write tokens by bucket. It also gives no token-flow edges between stages. A token-flow diagram would require evidence that this extract does not hold.
Cost and time have different units. A cost bucket does not establish a stage's active time, elapsed time or order. The cost chart cannot establish whether a check passed. The benchmark receipt keeps the selected attempt and its outcome separate from the all-attempt stage chart.
More generally, a recorded subtotal with unknown prices remains incomplete. A missing price does not mean a free call. A cost estimate and a source receipt do not verify an actual charge or establish customer acceptance.
Keep the interim comparison separate
The Opus and Sonnet stage chart uses a different scope. Its published October 6 record has three of eight declared pairs graded. It uses six run-mode stage labels, rather than the seven buckets above.
That chart excludes onboarding and calls outside a stage, so its bars do not sum to each arm's attempt total. The arms also ran on different platform builds. Do not merge that interim comparison into the 33-attempt total or treat it as a model ranking.
For other questions, read cost per resolved task or claims versus recorded costs. The agent harness explainer describes the pipeline. This article stays with the recorded cost buckets and their limits.