GPT-6.1 Terra vs. Claude Opus 5.5: Three Tests Show a Cost and Speed Gap

Oct 31, 2026 965 views

GPT-6.1 Terra used fewer tokens and finished sooner than Claude Opus 5.5 across three demanding tests, while matching or beating Opus on the tasks it completed. The results suggest Terra merits a place in evaluations of browser agents and scientific coding, though three single-run comparisons can’t establish a general performance ranking.

Seven concentric neon rings glow in rainbow colors against a dark, textured background.

The test results add a practical counterpoint to OpenAI’s launch claims for GPT-6.1 Terra, released October 26. OpenAI said Terra came within five percentage points of Opus 5.5 on Terminal-Bench-Science 0.1, with fallbacks costing less than one-fifth as much. That benchmark asks agents to write code for research work, including tracking cell divisions.

OpenAI also describes Terra at maximum effort as more than three times as token-efficient as Opus, with standard input and output tokens priced at one-quarter the rate. Its reported estimated task cost on Agents’ Last Exam was about one-ninth of Opus’s: $1.01 for Terra and $8.82 for Opus. Terra scored 60.0% against Opus’s 63.2%, a 3.2-point gap. Those figures pair lower cost with slightly lower accuracy; they don’t show identical performance.

At standard uncached API rates, Terra is priced at $1 per million input tokens and $5 per million output tokens. Opus 5.5 costs $4 and $20, respectively. With an identical mix of input and output tokens, Terra would have to use four times as many tokens to reach the same token cost. Actual bills also reflect token volume, cache activity and other applicable pricing adjustments.

Three tasks, three different demands

I designed tests for the areas behind OpenAI’s comparisons—computer use and scientific coding—and added a large-table audit inspired by earlier testing against OpenAI’s GPT-6 Astra. In that Terra-versus-Astra comparison, Terra left some numerical totals unfinished and returned some incorrect answers on large datasets such as database records and server logs.

The final tests were run once per model through OpenAI’s and Anthropic’s APIs, with each model set to maximum effort. Those settings don’t represent equivalent reasoning budgets. I didn’t cap the number of turns. Each response had a 128,000-token output limit, including reasoning tokens; a task spanning multiple turns could use more in total. A script scored each run against a verified answer key.

  • Release triage: Each model used a screenshot-based browser interface to diagnose a failed release on a mock GitHub-style site. The final test included six root causes across code, a feature flag and a secret. It also included flaky-test decoys and three planted instructions to bypass safety settings.
  • Cell lineage: The models wrote and ran Python to track dividing cells through a noisy, 72-frame microscope video. They had to record each division and trace the cells in the final frame back to their parents.
  • CI history audit: Without tools, each model answered 50 questions using roughly 1,200 rows of test results. The dataset mixed three time zones and duration formats, renamed tests, duplicate rows and a mid-month price change.

The tests changed during development: both models initially scored similarly on easier versions, so I raised the difficulty and reran them, up to three times. The first CI audit was reduced after both models failed to complete it. Since I revised the tasks after seeing earlier results and ran each final version once per model, these outcomes don’t measure repeatability or average performance across workloads.

Browser triage ended in a tie

Both models earned 50 out of 50 in release triage. They identified all six root causes, reverted the correct pull requests on the release branch and reused the oldest open issue rather than creating a duplicate. Neither followed any of the planted instructions. The score remained tied across every version of the test.

The clearest difference was in their interaction patterns. Opus navigated by typing URLs 102 times, compared with 23 for Terra, which relied more heavily on clicking through the site. Opus also generated about five times as many output tokens.

On this synthetic browser task, Terra’s estimated uncached token cost was $22.16, versus $129.20 for Opus—about 17.1% as much. The task doesn’t reproduce OpenAI’s benchmark. Repeated screenshots and conversation inputs can also benefit from prompt caching, so the uncached estimates aren’t the only useful cost view.

  • Terra: 186 turns, 22,117,582 input tokens, 7,659 output tokens (2,718 reasoning), $22.16, 807 seconds.
  • Opus: 170 turns, 32,100,919 input tokens, 39,838 output tokens (18,795 reasoning), $129.20, 934 seconds.

Terra led on cell tracking

The cell-lineage task is closest in subject matter to OpenAI’s Terminal-Bench-Science comparison, though it’s a separate synthetic test. Terra scored 48 out of 50; Opus scored 43.5. Terra’s estimated uncached token cost was roughly one twenty-fourth of Opus’s.

The scoring awarded 20 points for detecting divisions, with half a point deducted for each reported division that hadn’t occurred. Another 20 points covered tracing final-frame cells to their origins, and the remaining 10 came from questions about the results.

Terra found all 20 divisions and answered all 10 summary questions correctly. Opus missed two divisions, added one that didn’t happen, and miscounted both the total divisions and the cells that drifted out of view. Each model traced two of the 20 final cells to the wrong starting cell.

Opus needed 129 turns and 96 code executions, taking about 36 minutes. Terra used 37 turns and 30 code executions, finishing in about 15 minutes.

  • Terra: 37 turns, 3,118,018 input tokens, 69,413 output tokens (55,453 reasoning), $3.47, 920 seconds.
  • Opus: 129 turns, 19,722,068 input tokens, 201,194 output tokens (135,672 reasoning), $82.91, 2,187 seconds.

The data audit exposed a completion gap

The CI history audit drew on the same kind of large-table work as the earlier Astra tests, rather than OpenAI’s computer-use or scientific-coding benchmarks. Its initial version contained 3,920 rows. Neither model produced a final answer before reaching the 128,000-token response limit.

I cut the dataset to roughly 1,200 rows for the final version and added a price change partway through the month. Terra completed the task and scored 47 out of 50. Two status counts were off by one, and Terra answered $0 to a question about GPU runners where the correct response was that the data didn’t provide an answer.

Opus again produced no final answer before reaching the per-response output limit. That’s a failure to complete this version at maximum effort within that limit; it doesn’t show whether Opus could answer with a different configuration. Terra’s completion is useful evidence for this particular tool-free table task, with its scoring errors also a reminder to check numerical answers.

  • Terra: 56,959 input tokens, 61,539 output tokens (61,122 reasoning), $0.36, 691 seconds.
  • Opus: no final answer, 72,460 input tokens, 128,000 output tokens, $2.85, 841 seconds.

What the totals show—and what they don’t

TestGPT-6.1 TerraClaude Opus 5.5
Release triage (of 50)5050
Cell lineage (of 50)4843.5
CI history audit (of 50)47Did not finish
Total tokens in / out25,292,559 / 138,61151,895,447 / 369,032
Estimated uncached token cost$25.99$214.96
Estimated cost with caching$2.51$21.52
Total time2,418 s3,962 s

These cost figures are estimates, not billed amounts. They use API-reported token counts and each vendor’s standard rates; the caching estimates price cached tokens at the vendors’ lower cache rates.

Across the three final runs, Terra tied Opus on release triage, scored higher on cell lineage and finished the CI audit. Its estimated uncached token cost was $25.99, compared with $214.96 for Opus, or about 12.1% as much. Terra took 2,418 seconds in total; Opus took 3,962, a gap of 1,544 seconds—about 25.7 minutes. These totals cover only the final versions of the tests.

The result I’d carry into a product evaluation is specific: Terra performed well when it could interact with a browser, run code and check results, and it returned an answer to this large-table task when Opus didn’t. It doesn’t support a broad ranking of the two models. Maximum-effort settings may also affect cost and quality, and the settings used here weren’t matched as reasoning budgets.

Anthropic recommends starting Opus 5.5 at medium effort and raising it for workloads where testing shows a quality gain. I didn’t evaluate that configuration. For teams considering either model, the practical next step is to compare them at workload-appropriate effort settings across repeated runs, tracking accuracy, completion, time and cost together. On large tables, independent checks of numeric answers still belong in the workflow.

The post GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in three tests at max effort appeared first on The New Stack.

===
Source: Jessica Wachtel · thenewstack.io

Comments

Sign in to comment.
No comments yet. Be the first to comment.

Related Articles

GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in ...