What should a task cost?

Cents per task — $0.01 to $0.10 for a frontier model on a self-contained change.

Based on Aider polyglot suite totals ÷ 225 exercises: $0.006–$0.08 per exercise across frontier models. Exercises arrive with tests and no repository context; real XS–S tasks in a live repo run higher with harness overhead.

Budget a few dollars per attempt; $0.10 with cheap models, $30+ at max reasoning effort.

Based on SWE-bench per-task costs on the Verified and Multilingual splits ($0.05–$0.36 on efficient models, higher with reasoning effort) and HAL's cross-agent range of $0.08–$32 per SWE-bench Verified task. Benchmark bugs come pre-scoped with a failing test; finding and scoping your bug is extra — and failed attempts cost the same as successes.

Roughly $10–$20 and 35k–65k tokens for one serious agentic pass at high effort.

Based on CursorBench per-task cost and token counts at three reasoning-effort levels. A single benchmark under agent-run conditions — one calibrated anchor, not a market rate.

Unattended agent campaigns reach $1,500–$3,900 and 8M–57M tokens — the ceiling to plan against, not the typical case.

Based on Complete autonomous eval runs: OSWorld 2 full-run costs and GBAEval total token consumption. These are entire eval campaigns. A supervised multi-day task usually lands far below this ceiling — but an overnight fleet can genuinely hit it.

Thirty cents to a few dollars for one genuinely hard single-shot problem.

Based on ARC-AGI and WeirdML per-problem costs across frontier models. Puzzle-style tasks — closest to one tough chat question, not a workflow.