Coding
Greenfield prototype · L
Feature or refactor touching many files/modules; design decisions involved.
No community data yet · no published cell yet — be the first to report
upper bound — not a typical estimate
$1,500–$3,900per full unattended campaign run· 8M–57M tokens
This is the ceiling an unattended agent fleet can reach on work this size — plan against it, do not expect it. These are entire eval campaigns. A supervised multi-day task usually lands far below this ceiling — but an overnight fleet can genuinely hit it.
Typical for this shape: no direct lab per-task data exists yet. No adjacent per-task reference exists either — community reports are the only path to a real number here.
Based on Complete autonomous eval runs: OSWorld 2 full-run costs and GBAEval total token consumption. Full basis and evidence.
No community reports for this cell yet — yours would be the first.