GPT-6.1 Terra scores 42.3% on Artificial Analysis' FrontierCode 2.0 run, 0.1 behind GPT-6 Astra at a tenth of the list price

Artificial Analysis' own run of Cognition's FrontierCode 2.0, read at 15:00 UTC on October 28, puts GPT-6.1 Terra at 42.3% at max effort. That is third of 28 models and 0.1 point behind GPT-6 Astra (42.4%). OpenAI lists Terra at $1 / $5 per million input / output tokens and Astra at $10 / $50. Claude Opus 5.5 leads at 46.4%.

THE TOP OF THE BOARD

Claude Opus 5.5 46.4%, GPT-6 Astra 42.4%, GPT-6.1 Terra 42.3%, Claude Sonnet 5.5 41.7%, GPT-6.1 Sol 40.6%, Muse Wonder Full 37.8%, Claude Haiku 5.5 34.9% and Claude Fable 5.1 31.6%. The best open-weights model is DeepSeek V4.1 Flash 1017 at 28.5%, tenth overall.

HOW ARTIFICIAL ANALYSIS RAN IT

FrontierCode 2.0 is Cognition's private benchmark of 132 tasks in real open-source repositories, graded on whether a maintainer would merge the agent's change. Artificial Analysis runs every task three times in its own Stirrup agent harness, capped at 500 turns, and says the benchmark is now part of its Intelligence Index.

TWO RUNS, DIFFERENT NUMBERS

Cognition publishes its own results, and Terra is not in them. Where both test a model, the scores differ: Opus 5.5 45.5% on Cognition's run against 46.4% on Artificial Analysis', Sonnet 5.5 41.0% against 41.7%, GPT-6 Astra 42.8% against 42.4%, Haiku 5.5 27.7% against 34.9%. Cognition's row labelled 'DeepSeek V4.1 Flash' scores 19.3%, while Artificial Analysis' run of the October 17 build scores 28.5%; Cognition's file does not say which build it tested. @scaling01 flagged the DeepSeek gap on X.

WHAT WE DO NOT KNOW

What each model spent per task on this run, so the list prices above are not the cost of a task: a model that thinks longer can cost more per task. Neither page explains why the two runs differ. The hidden 'Obsidian 3' row that briefly sat in Cognition's file is not on Artificial Analysis' board.