Arena's harness tax study
arena's harness tax study: running the same model through different agent harnesses barely moves success rates but can double the bill per task. the orchestration layer is now a 2x tax with no quality upside.
the wrapper is the new legacy code.
Context
Arena's post of 16 September 2026 tested 21 model-harness pairs, seven models and three harnesses (Claude Code, Codex CLI and Pi), on the same 30 randomly sampled tasks from each of SWE-bench Lite and Terminal-Bench 2.0, each run three times per task at high effort with a 100-turn cap, with costs from a fixed direct-API price list dated 1 September 2026 and 95% intervals by 10,000 bootstrap resamples. It reports that harness choice has little effect on success but can significantly affect cost, up to 5x for the same model, with Claude Code about 2.0x Pi and 1.6x Codex on SWE-bench Lite and 1.5x Pi on Terminal-Bench 2.0 for shared models.
Double is one pair's case; Arena's range is 1.5x to 5x. The samples are small, 30 tasks per benchmark, on two public benchmarks models may have seen, and the post is not peer reviewed and says results may differ on other workloads. No quality upside overstates it: Arena reports slightly higher success for Fable 5 in Claude Code than in Pi at about twice the cost, and 1.1% higher success for Claude Code in one case. The orchestration layer being a 2x tax is the author's conclusion.
Related work
- Harness tax (igniting.github.io) ↗Snippet only.
- Harness tax review (Tom Rochette) ↗Snippet only.
Watch next
- Larger replications on other benchmarks.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 18 September 2026 at 08:16 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →