GPT-6.1 Terra beats Claude Opus 5.5 on cost and speed
OpenAI compares GPT-6.1 Terra to Claude Opus 5.5 on three synthetic tasks. Mistral’s model performed better on two tests, completed the third, and cost significantly less.
OpenAI pitted GPT-6.1 Terra against Claude Opus 5.5 on three synthetic tests of development and agentic work, published on October 31, 2026. In these trials, Terra matched Opus on a software triage task, outperformed it on a cell lineage analysis, and completed a data audit that Claude did not finish. Most importantly, the estimated cost strongly favored Terra, which matters directly for teams already paying for models to execute work, not just to chat.
Measured gaps between two high-end models
Jessica Wachtel specifies that comparisons were made at maximum effort mode, using native APIs from both providers, without dialogue turn limits. This detail matters because effort settings do not guarantee the same reasoning budget from one model to another. The results therefore do not say who “wins” absolutely, but what each model delivers in a specific framework with precise constraints.
On the “release triage” software triage test, both models scored 50 out of 50. However, Terra was more economical in how it reached this score: 23 round trips via the address bar versus 102 for Opus, and about five times fewer output tokens. As a result, Terra’s estimated cost was 17.1% of Opus’s on this task. A saving that is far from trivial when multiplied by hundreds of runs.
When code and data make the difference
The second test, focused on tracking cell lineage from a noisy 72-frame microscopic video, gave Terra 48 out of 50 versus 43.5 out of 50 for Opus. Terra tracked the expected 20 divisions and correctly answered 10 summary questions, while Opus missed two divisions, reported a nonexistent event, and misattributed two final cells. Again, Terra consumed far fewer resources, with 37 turns versus 129 and about 15 minutes versus 36 minutes.
The third test involved auditing a CI history of about 1,200 lines, with timezone changes, heterogeneous duration formats, and a mid-month price change. Terra scored 47 out of 50. Opus did not produce a final answer before hitting its 128,000 token output limit per response. This is not proof of general incapacity, but a clear signal on this type of task when data volume grows. Models like tables… until they start talking to themselves.
What the numbers say, without marketing gloss
Across all three final runs, Terra produced an estimated uncached cost of $25.99 versus $214.96 for Opus, about 12.1% of its rival’s cost. Total time was also reduced: 2,418 seconds for Terra, versus 3,962 seconds for Opus. In tokens, Terra totaled 25,292,559 input and 138,611 output, while Opus reached 51,895,447 input and 369,032 output. The gaps are wide enough to interest product managers, SREs, or AI teams seeking to industrialize agentic workflows without blowing budgets.
One useful nuance remains: these tests were built by Jessica Wachtel and do not reproduce a single official benchmark or full production conditions. However, they show a trend already visible in other model comparisons: useful quality no longer depends solely on raw accuracy, but on the ratio between cost, speed, completion, and ability to remain reliable as context lengthens. That’s where trade-offs become concrete and less glamorous than demo promises.
By the numbers
- 50/50 — Terra and Opus score on software triage.
- 48/50 — Terra’s score on cell lineage.
- 47/50 — Terra’s score on CI audit.
- $25.99 — Terra’s estimated total uncached cost.
- $214.96 — Opus’s estimated total uncached cost.