OpenAI's October 26 launch of GPT-6.1 Terra included head-to-head comparisons with Claude Opus 5.5, highlighting substantial cost advantages. According to OpenAI's estimates, Terra's computer-use benchmark costs roughly one-ninth of Opus 5.5's. In three synthetic tests run at maximum effort, Terra consistently delivered lower expenses and faster turnaround, tying Opus on one task, outscoring it on another, and completing a third that Opus could not finish within its response limits.

OpenAI claims Terra achieves more than three times the token efficiency of Opus at max effort, while costing one-quarter the price for standard input and output tokens. The company reports an estimated cost per task of approximately one-ninth. On Agents' Last Exam, a benchmark measuring multistep computer-use operations, Terra scored 60.0% at $1.01 per task compared to Opus's 63.2% at $8.82—a 3.2-percentage-point accuracy 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, versus Opus 5.5's $4 and $20. Using identical input and output token ratios without additional pricing adjustments, Terra would need four times as many tokens to match Opus's total cost. Real-world task expenses depend on actual token consumption, cache operations, and any vendor-specific pricing modifications.

Prior testing of Terra against OpenAI's GPT-6 Astra revealed accuracy issues, including uncalculated numerical totals and incorrect answers, particularly on large-dataset analysis tasks. Since OpenAI's Opus comparisons emphasize computer use and scientific coding, two new tests targeted those domains. A third test recreated the large-dataset scenario where Terra had struggled, to assess whether Opus faced similar limitations.

Test methodology

Each model ran through OpenAI's and Anthropic's APIs with effort settings at maximum. The tests underwent multiple iterations when initial scores proved too similar; results reported here come from final versions only, each run once per model. This approach does not establish repeatability or average performance across different workloads.

No turn limits were imposed. Both models operate under a 128,000-token output ceiling per response, including reasoning tokens; multiturn tasks can exceed this total. Token counts aggregate across all turns in each final run, with reasoning tokens included in output totals. Cost estimates use standard uncached rates, not actual invoices. Terra's pricing increases for individual requests exceeding 272,000 input tokens; aggregate task totals alone do not reveal whether this threshold was crossed. Caching-adjusted totals appear separately. Figures exclude earlier test iterations and separately billed tools or execution infrastructure.

The three tests

  • Release triage: A simulated GitHub-style environment with a failed software release. Each model had to resolve it via a screenshot-based browser interface using clicks and typed URLs. The final version contained six root causes across code, a feature flag, and a secret, plus decoy pull requests and three planted instructions attempting to bypass safety settings.
  • Cell lineage: A noisy 72-frame microscope video of dividing cells. Models wrote and executed their own Python code to track divisions, logging each one and tracing every cell in the final frame back to its origin.
  • CI history audit: Approximately 1,200 rows of CI test results in a single prompt without tools, followed by 50 questions. The dataset mixed three time zones, three duration formats, renamed tests, duplicate rows, and a mid-month price change.

Release triage results

This test produced a 50–50 tie that persisted through all iterations. Scoring used 50 pass/fail checks evaluating whether models filed and closed correct issues, fixed each problem to allow release passage, and avoided unintended changes to decoy pull requests and settings while ignoring planted instructions.

In the hardest version, both models identified all six root causes, reverted appropriate pull requests on the release branch, and reused the oldest open issue rather than creating a duplicate. Neither followed any planted instructions. The path to success differed: Opus typed URLs 102 times while Terra did so 23 times, relying mostly on clicks. Opus also generated approximately five times more output tokens.

For this browser-based task, Terra achieved identical scores at roughly 17.1% of Opus's estimated uncached token cost, a 5.83-to-one cost advantage. This was a separate synthetic exercise, not a replication of OpenAI's published benchmark. Prompt caching can reduce costs for repeated screenshots and conversation inputs; cached and uncached totals appear below.

  • 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

Cell lineage results

This test most closely resembles OpenAI's Terminal-Bench-Science comparison between Terra and Opus 5.5 with fallbacks, though this synthetic cell-tracking exercise does not replicate that exact benchmark. Terra scored 48 out of 50 while Opus scored 43.5 out of 50, with Terra using approximately one twenty-fourth of Opus's estimated uncached token cost.

Scoring allocated 20 points for identifying cell divisions, minus half a point per incorrectly reported division. Another 20 points came from tracing each final-frame cell to its origin. The remaining 10 points derived from summary questions such as total division count and cells leaving the frame.

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

Despite its lower score, Opus required more turns and code executions. It ran code 96 times across 129 turns over approximately 36 minutes. Terra executed code 30 times across 37 turns, completing 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

CI history audit results

This test did not replicate OpenAI's computer-use or scientific-coding benchmarks; it resembles large-dataset exercises run against Astra. The initial version contained 3,920 rows, and neither model produced a final answer before hitting the 128,000-token output limit.

The second version reduced the file to approximately 1,200 rows and introduced a mid-month price change as an additional complication. Terra finished with 47 out of 50 correct answers. It miscounted two status totals by one and answered $0 for a GPU runner question when the data did not address that category. Opus again produced no final answer before reaching its 128,000-token-per-response ceiling—a failure to complete at max effort within that constraint, not proof that Opus could not succeed under different configurations.

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

  • 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

Overall results

Cost figures represent estimates derived by multiplying reported token counts by each vendor's standard rates; cached-token rows apply vendor-specific lower cache rates. Terra tied Opus on release triage, outscored it on cell lineage, and completed the CI audit while Opus returned no answer. Across these three final runs, Terra's estimated uncached token cost totaled $25.99 versus Opus's $214.96—approximately 12.1% of Opus's cost, 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 figures exclude earlier test versions.

Analysis

Terra performed competitively when it could click, execute code, and verify results. On the final table task without tools, Terra returned an answer while Opus did not. This outcome applies to these specific runs and does not constitute a general model ranking.

These tests employed max effort, whereas earlier Astra comparisons used high effort. Task differences prevent isolating the effort-setting effect. Anthropic recommends beginning Opus 5.5 at its default medium effort and escalating to higher settings only when testing demonstrates quality improvements. That configuration was not evaluated here.

For browser and code-execution workloads, running Terra at max effort in a trial before committing to Opus would be prudent. Comparing each model at the effort level optimal for the specific workload, then measuring accuracy, completion rates, time, and cost across repeated runs, provides better guidance. Numerical answers on large tables warrant additional verification, as Terra produced errors in this test and in the earlier Astra comparison.