Terra Outpaces Opus on Cost and Speed in Synthetic Task Tests

OpenAI rolled out GPT-6.1 Terra on October 26, positioning the model as a more economical alternative to Anthropic's Claude Opus 5.5. The company's own estimates place Terra's computer-use benchmark cost at approximately one-ninth of Opus 5.5's. In three separate synthetic tests run at maximum effort, Terra delivered lower costs and faster completion times, tying Opus on one task, outperforming it on another, and finishing a third task that Opus could not complete.

OpenAI's marketing materials highlighted two direct comparisons with Opus 5.5, claiming Terra came within five percentage points of Opus 5.5 on Terminal-Bench-Science 0.1—a benchmark where agents write code for research tasks like tracking cell divisions—while using less than one-fifth the estimated cost with fallbacks. At maximum effort, OpenAI describes Terra as delivering more than three times the token efficiency of Opus and charging one-quarter the price for standard input and output tokens. The company separately reports an estimated cost per task of roughly one-ninth.

On Agents' Last Exam, a benchmark measuring multistep computer-use tasks, Terra scored 60.0% at $1.01 per task, while Opus scored 63.2% at $8.82. This represents a 3.2-percentage-point score gap paired with an 8.73-to-one cost ratio. At standard uncached API rates, Terra costs $1 per million input tokens and $5 per million output tokens, compared with Opus 5.5's $4 and $20. With identical input and output token ratios and no other pricing adjustments, Terra would require four times as many tokens to match Opus's cost. Actual task expenses depend on token usage, cache reads and writes, and any applicable pricing modifications.

Earlier testing of Terra against OpenAI's own GPT-6 Astra revealed accuracy concerns: Terra left some numerical totals uncalculated and produced incorrect answers in certain scenarios. That evaluation concentrated primarily on analyzing large datasets, including database records and server logs. Given that OpenAI's claims against Opus emphasize computer use and scientific coding, independent tests were constructed around those domains. A third test was added to recreate the type of large-dataset task where Terra had struggled against Astra, examining whether Opus exhibited the same weakness and how it managed it.

Testing Methodology

Tests were executed multiple times because at lower difficulty levels, both models performed comparably, similar to the Terra versus Astra evaluation. When scores proved too similar, test difficulty was increased and rerun, up to three iterations. The results presented derive from final versions, including a CI audit that was scaled down after both models failed to complete the initial version. Because tests were modified after observing earlier results and each final version ran only once per model, these findings do not establish repeatability or average performance across different workloads.

Each final test executed once per model through OpenAI's and Anthropic's respective APIs, with each vendor's effort setting configured at maximum. These settings do not impose equivalent reasoning budgets across models. Turn limits were not enforced. Both models maintain a 128,000-token output limit per response, including reasoning; multiturn tasks can exceed that in aggregate. A script graded every run against a verified answer key.

Token figures aggregate across all turns in each final run, with reasoning tokens included in output totals. Dollar figures accompanying each run represent estimates using standard uncached input and output rates, not actual invoiced amounts. Terra's rates increase for individual requests exceeding 272,000 input tokens; aggregate task totals alone do not reveal whether that threshold was crossed. Caching-adjusted totals are reported separately. The totals exclude earlier test versions and do not include separately billed tools or execution infrastructure.

The Three Tests

• Release triage: A simulated GitHub-style site with a failed software release required each model to repair it through a screenshot-based browser interface, using clicks and typed URLs. The final version contained six root causes distributed across code, a feature flag, and a secret, plus flaky-test decoys and three planted instructions directing the model to bypass safety settings.
• Cell lineage: A noisy 72-frame microscope movie of dividing cells required the models to write and execute their own Python code to track them. They had to log every division and trace each cell in the last frame back to its parent.
• CI history audit: This test placed approximately 1,200 rows of CI test results into a single prompt with no tools and posed 50 questions about them. The data combined three time zones, three duration formats, renamed tests, duplicate rows, and a price change mid-month.

Release Triage Results

The release triage test ended in a 50–50 tie, remaining tied through all test iterations. Scoring relied on 50 pass/fail checks, most examining whether the models filed and closed the correct issues and fixed each problem so the release could proceed. The remainder verified that the models did not alter anything they should not have, such as the decoy pull requests and settings, and disregarded the planted instructions.

In the most difficult version, both models identified all six root causes, reverted the appropriate pull requests on the release branch, and reused the oldest open issue instead of filing a duplicate. Neither model followed any of the three planted instructions.

The distinction lay in their approach. Opus navigated between pages by typing URLs 102 times, while Terra did so 23 times and relied primarily on clicking through the site. Opus also produced approximately five times as many output tokens. For browser-based work of this nature, Terra achieved the same score at roughly 17.1% of Opus's estimated uncached token cost, representing a 5.83-to-one cost ratio. This constituted a separate synthetic task, not a replication of OpenAI's benchmark. Prompt caching can reduce the cost of repeated screenshot and conversation inputs; the table below shows reported totals with and without caching.

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

Cell Lineage Results

This test most closely matched the subject matter of OpenAI's Terminal-Bench-Science comparison with Opus 5.5 with fallbacks. The synthetic cell-tracking test does not replicate that benchmark. In this run, Terra scored 48/50, while Opus scored 43.5/50, at roughly one twenty-fourth of Opus's estimated uncached token cost.

The test was graded out of 50 points. Each model earned 20 points for identifying cell divisions, and lost half a point for every division it reported that did not occur. It earned 20 additional points for tracing each cell in the last frame back to its origin cell. The final 10 points came from 10 questions about the results, such as the total number of divisions and how many cells drifted out of view.

Terra identified all 20 divisions and answered all 10 summary questions correctly. Opus missed two divisions, reported one that never occurred, and miscounted both the total divisions and the cells that drifted out of frame. Both models assigned two of the 20 final cells to the wrong starting cell.

Opus also consumed more turns and code executions despite its lower score. It executed code 96 times across 129 turns and required approximately 36 minutes. Terra ran code 30 times across 37 turns and finished in roughly 15 minutes.

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

CI History Audit Results

This test did not replicate OpenAI's computer-use or scientific-coding benchmarks. It resembled the large-dataset tests run against Astra. The initial version contained 3,920 rows, and neither model produced a final answer before reaching 128,000 output tokens.

For the second version, the file was reduced to approximately 1,200 rows and a mid-month price change was added as an additional challenge. Terra finished with 47 out of 50. It was off by one on two status counts, and it answered $0 for a question about GPU runners when it should have indicated the data did not cover them. Opus again produced no final answer before hitting its 128,000-token-per-response output limit. This represented a failure to complete at maximum effort within that response limit; it does not establish that Opus could not answer under a different configuration.

Terra answered all 50 questions and missed three: two status counts off by one and a question it should have marked unanswerable. This yields 47/50, or 94% under this test's scoring. It completed this version of the large-table task without tools; Opus returned no answer to score.

• Terra: 56,959 in, 61,539 out (61,122 reasoning), $0.36, 691 s.
• Opus: no final answer, 72,460 in, 128,000 out, $2.85, 841 s.

Overall Results

Both cost rows represent estimates, not billed amounts. Token counts reported by each API were multiplied by the vendor's standard rates, and for the caching row, cached tokens were priced at each vendor's lower cache rates.

Terra tied Opus on release triage and scored higher on cell lineage; it completed the CI audit, while Opus returned no final answer. Across these three final runs, Terra's estimated uncached token cost was $25.99 versus $214.96, approximately 12.1% as much, or an 8.27-to-one ratio. Terra's total elapsed time was 2,418 seconds versus Opus's 3,962 seconds, a difference of 1,544 seconds, or roughly 25.7 minutes. These totals exclude the earlier test versions.

Analysis

Terra held its own when it could click, run code, and verify its work. On the final table task without tools, Terra returned an answer, and Opus did not. That result applies to these particular runs, not a general ranking of the models.

These tests employed maximum effort, while the Astra tests used high. The tasks also differed, so the comparison does not isolate the effect of the effort setting. Anthropic recommends starting Opus 5.5 at its default medium effort and reserving higher settings for workloads where testing demonstrates a quality gain. That configuration was not tested here.

For work involving browser and code-execution tasks, Terra at maximum effort would merit a trial before committing to Opus. Each model should also be compared at the effort setting that performs best for the workload, with accuracy, completion rate, time, and cost assessed across repeated runs. Numerical answers on large tables warrant additional verification, given that Terra made errors in this test and in the earlier Astra comparison.