A study of agent consistency found claude solved 58 percent of identical task runs where gpt-5…
a study of agent consistency found claude solved 58 percent of identical task runs where gpt-5 solved 32, despite similar benchmark standings — claude took 46 steps per run, gpt-5 took 9.9. the same model that nails a task once can fail it twice.
agent reliability is a variance problem, and leaderboards will not show it.
Context
The arXiv paper Consistency Amplifies: How Behavioral Variance Shapes Agent Accuracy (2603.25764, 26 March 2026) compares Claude 4.5 Sonnet, GPT-5 and Llama-3.1-70B on SWE-bench across 50 runs each. Claude 4.5 Sonnet had a coefficient of variation of 15.2% and accuracy of 58%, GPT-5 32.2% and 32%, Llama-3.1-70B 47.0% and 4%. It says GPT-5 is 4.7x faster than Claude, 9.9 versus 46.1 steps, with a maximum of 250 steps, and that 71% of Claude's failures are consistent wrong interpretation.
The figures match the paper. They cover Claude 4.5 Sonnet and GPT-5, not later versions, on SWE-bench with 50 runs each and a step cap, in a March 2026 preprint by authors of the study itself. The 58% is accuracy, not a share of runs solved identically. Despite similar benchmark standings was not found in the lines read. Can fail it twice is the author's wording, and the paper's point is that consistency amplifies outcomes, including consistently wrong ones. Results do not generalize to current models. Steps are not measured runtime.
Watch next
- Replication on newer models.
Sources
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 18 September 2026 at 17:03 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 →