99.9% on arc-agi-3, then 62.7% in a different harness. that is not a benchmark bug — it is the eval…
99.9% on arc-agi-3, then 62.7% in a different harness. that is not a benchmark bug — it is the eval market splitting in public.
whoever ships the unified harness by q1 owns the whole category.
Context
ARC Prize's results page for GPT-6 Astra (2 September 2026) and its blog (3 September 2026) report two harness configurations of the same model on the ARC-AGI-3 Semi-Private split. With the Standard harness, where the model carries forward notes it chooses to keep, the best observed is 62.7% at max reasoning for about 26K USD. With the Provider Adapter harness, which preserves opaque reasoning state between requests and uses compaction, the best observed is 99.9% at high reasoning for about 19K USD. The leaderboard table lists Provider Adapter High at 99.95%, and Standard Max at 62.71%.
These are two configurations, not one score falling. The blog's 99.9% and the table's 99.95% are unreconciled here, and are not called rounding. Each value is kept with its own source: the blog gives 99.9% and the table gives 99.95%. The max and high reasoning levels and the costs belong to their own configurations. OpenAI's own page claims 99.9% on ARC-AGI-3 without naming a harness in the sentence read. Not a benchmark bug, the eval market splitting, and a unified harness owning the category by the first quarter are the author's opinion and prediction.
Related work
- Introducing GPT-6 Astra (OpenAI) ↗OpenAI's own 99.9% claim.
Watch next
- A reconciliation of the blog and leaderboard values.
Sources
- GPT-6 Astra results (ARC Prize, 2 Sep 2026)arcprize.org
- Astra on ARC-AGI-3 (ARC Prize, 3 Sep 2026)arcprize.org
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 6 September 2026 at 01:02 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 →