GPT-6.1 Terra and Claude Opus 5.5 Compared in Three Coding Tasks

OpenAI launched GPT-6.1 Terra on October 26 and compared it with Claude Opus 5.5 in marketing materials. In three separate synthetic tests, Terra tied Opus on release triage, scored higher on cell lineage, and completed a CI history audit that Opus did not finish, while estimated uncached token costs were far lower.

OpenAI's claims

OpenAI said Terra came within five percentage points of Opus 5.5, with fallbacks at less than one-fifth the estimated cost, on Terminal-Bench-Science 0.1. At max effort, it described Terra as more than three times as token-efficient as Opus and one-quarter the price for standard input and output tokens, with an estimated cost per task of about one-ninth. On Agents' Last Exam, Terra scored 60.0% at $1.01 per task and Opus scored 63.2% at $8.82, a 3.2-point gap and an 8.73 to 1 cost ratio. Standard uncached API rates are $1 per million input tokens and $5 per million output tokens for Terra, compared with $4 and $20 for Opus 5.5.

Release triage

In a 50-point release triage test with six root causes, flaky-test decoys, and planted instructions to bypass safety settings, both models scored 50. Both found all six root causes, reverted the right pull requests, reused the oldest open issue, and ignored the planted instructions. Opus typed URLs 102 times while Terra did so 23 times and mostly clicked. Opus wrote about five times as many output tokens. Terra's estimated uncached token cost was about 17.1% of Opus's, a 5.83 to 1 ratio.

Cell lineage and CI audit

In a 50-point cell-tracking test, Terra scored 48 and Opus scored 43.5. Terra matched all 20 divisions and answered all 10 summary questions; Opus missed two divisions, reported one that did not happen, and miscounted total divisions and cells that drifted out of frame. Both assigned two of 20 final cells to the wrong starting cell. Terra ran code 30 times over 37 turns in about 15 minutes; Opus ran code 96 times over 129 turns in about 36 minutes. Terra's estimated uncached token cost was about one twenty-fourth of Opus's. In a 50-question audit of about 1,200 rows of CI results with no tools, Terra scored 47. It was off by one on two status counts and answered $0 for a GPU-runner question when the data did not cover them. Opus produced no final answer before reaching its 128,000-token output limit. Terra used 56,959 input and 61,539 output tokens, costing an estimated $0.36; Opus used 72,460 input and 128,000 output tokens, costing an estimated $2.85. An earlier 3,920-row version produced no final answer from either model before that limit.

Costs and caveats

Across the three final runs, Terra used 25,292,559 input and 138,611 output tokens for an estimated uncached cost of $25.99; Opus used 51,895,447 input and 369,032 output tokens for $214.96. That is about 12.1% as much, or an 8.27 to 1 ratio. Estimated costs with caching were $2.51 and $21.52. Terra's total elapsed time was 2,418 seconds versus 3,962 seconds. The author ran each final test once per model at max effort through the vendors' APIs, changed tests after earlier results, and said the results do not establish repeatability or average performance. They also do not isolate the effect of effort settings, and Anthropic recommends starting Opus 5.5 at default medium effort.

Sources: The New Stack