Skip to content
SOLVENCY
Sign in

Harness economics · OpenBench

Same model. Different harness bill.

Four coding harnesses ran GPT-5.6 Sol on the same 15-task benchmark. Solvency reprices each harness’s proxy-measured token usage at the model’s current verified API rates.

OpenBench-only basis · never mixed with the general model leaderboard

Matched pass rate · different usage

Codex costs 3.77× Pi per solved task at the same 72.7% pass rate.

$1.37 vs $0.363 per solved task · point estimates, not a claim of statistical superiority

Lowest point estimate

Pi

$0.264 / attempt · $0.363 / solved

Price basis

Current API

$5 input · $0.50 cache read · $30 output per million tokens

Source: OpenBench harness benchmark (MIT) · verified 2026-08-23 · GPT-5.6 Sol prices verified 2026-08-21; source usage repriced, not an observed dollar bill

One model · one benchmark

Harness cost comparison

Ranked by cost per solved task. Only the harness changes; lower is better. Infrastructure and rate-limited failures are excluded exactly as OpenBench specifies.

GPT-5.6 Sol harness comparison: version, outcomes, pass rate, source-usage-repriced cost per attempt and per solved task
#HarnessVersionSolvedPass$ / attempt$ / solvedvs lowest
1Pi0.80.1032/4472.7%$0.264$0.3631.00×
2Claude Code2.1.21435/4577.8%$0.327$0.4201.16×
3Grok Build0.2.103 (89c3d36fb6f1)37/4582.2%$0.491$0.5971.64×
4Codex0.144.532/4472.7%$0.997$1.373.77×

Source: OpenBench harness benchmark (MIT) · verified 2026-08-23 · source usage repriced: measured tokens × current verified model prices; no Solvency loop or tier assumption

Cost basis

What this comparison means

This is neither a provider invoice nor Solvency’s modelled task tier. It is an independently reproducible reprice of source-observed usage.

Calculation

cost_per_attempt =
  mean_uncached_input × current_input_price
  + mean_cache_read × current_cached_input_price
  + mean_output × current_output_price

cost_per_solved = cost_per_attempt / solve_rate

Task tier, assumed loops, cache-hit controls, retry cap and frontier-efficiency assumptions cannot move these rows. Tests enforce that invariant.

Read with care

  • Each arm has only 44 or 45 countable attempts, so the correctness intervals are wide.
  • OpenBench reports subscription-backed access; Solvency applies current API token rates for comparability, not the subscription fee.
  • Cursor and OpenCode are excluded because their usage is CLI-self-reported; Devin is excluded because its split usage is incomplete.
  • The source flags Pi for a rerun after a surprising per-task swing. Treat the point estimate as provisional.