GPT-6.1 Terra vs. Claude Opus 5.5: Cheaper and faster in three tests at max

OpenAI released GPT-6.1 Terra on October 26, publishing marketing benchmarks that directly pitted the new model against Anthropic's Claude Opus 5.5 in scientific coding and autonomous computer use. The promotional materials claimed Terra matched Opus 5.5 within five percentage points on the Terminal-Bench-Science 0.1 benchmark at less than one-fifth the estimated cost, while claiming an estimated cost per task of roughly one-ninth on Agents' Last Exam. Seeking to verify those figures beyond marketing claims, independent evaluation detailed in thenewstack.io's report subjected both models to three rigorous synthetic tests at maximum reasoning effort: release triage via an interactive browser interface, automated cell lineage tracking from microscope imagery, and a direct prompt-based continuous integration test history audit.

This analysis examines how the two flagship frontier models compare under sustained, compute-intensive execution across software maintenance, image analysis, and large tabular dataset reasoning. The head-to-head evaluation is intended for engineering leaders, AI systems architects, and technical teams deciding between OpenAI and Anthropic API endpoints for multi-turn agentic workflows. Because API pricing structures and token consumption diverge significantly between vendors, real-world agent performance depends as much on turnaround latency and execution efficiency as baseline reasoning accuracy.

The test: GPT-6.1 Terra vs Claude Opus 5.5

The evaluation framework pitted OpenAI's GPT-6.1 Terra against Anthropic's Claude Opus 5.5 across three custom synthetic tasks executed directly through each provider's native API endpoints. Both models operated at their vendor-specified maximum effort settings, with no limits placed on conversational turn counts, while observing the standard single-response limit of 128,000 output tokens. Because earlier iterations at lower difficulty levels produced identical scores, the author escalated the difficulty across up to three test revisions. The reported numbers represent single-run evaluations on final, hardened versions graded against an automated answer key, meaning results illustrate concrete execution behavior rather than multi-run statistical averages.

Standard token pricing formed the foundation for financial estimates across all recorded interactions. OpenAI bills GPT-6.1 Terra at $1.00 per million input tokens and $5.00 per million output tokens for requests up to 272,000 input tokens, whereas Anthropic charges $4.00 per million input tokens and $20.00 per million output tokens for Opus 5.5. On earlier Agents' Last Exam benchmarks cited by OpenAI, Terra scored 60.0% at $1.01 per task compared to Opus 5.5's 63.2% at $8.82 per task, establishing an 8.73-to-one cost ratio. The tests examined whether that cost and efficiency disparity held up when models solved end-to-end tasks without manual guidance.

How GPT-6.1 Terra and Claude Opus 5.5 each did

In the first challenge, release triage, both models achieved a perfect 50 out of 50 across 50 objective pass-or-fail checks. Operating via a screenshot-driven simulated browser interface using mouse clicks and navigation URLs, the agents investigated a failed software deployment. Both models correctly identified all six root causes across application code, an erroneous feature flag, and an exposed secret, while reverting appropriate pull requests on the release branch and reusing existing issues instead of generating duplicates. Crucially, neither model succumbed to decoy flaky tests or three planted prompt injection instructions attempting to bypass security constraints.

Performance diverged sharply on the subsequent tasks. In the cell lineage test, which required writing and running Python scripts against a 72-frame noisy microscope video, Terra scored 48 out of 50 while Opus 5.5 scored 43.5. Terra tracked all 20 cell divisions and answered 10 summary questions accurately. Opus missed two divisions, hallucinated one false division, and miscounted out-of-frame drifts. In the CI history audit containing 1,200 rows with mixed time zones, duration formats, and price shifts, Terra scored 47 out of 50 without tools. Opus 5.5 consumed its 128,000 output token budget without delivering a final answer.

GPT-6.1 Terra vs Claude Opus 5.5, side by side

The aggregate data demonstrates that GPT-6.1 Terra completed all three tasks faster, with far fewer total tokens consumed, and at a fraction of the token expenditure required by Claude Opus 5.5. Across the entire battery, Terra accumulated 25,292,559 input tokens and 138,611 output tokens, compared to 51,895,447 input tokens and 369,032 output tokens for Opus 5.5. Total elapsed execution time reached 2,418 seconds for Terra versus 3,962 seconds for Opus 5.5, a margin of 1,544 seconds, representing a 25.7-minute speed advantage for OpenAI's model.

Calculated financial figures reveal an 8.27-to-one cost ratio for uncached token usage, totaling $25.99 for Terra against $214.96 for Opus 5.5 across the three runs. When factoring in vendor caching discounts, Terra's estimated operational cost dropped to $2.51 compared to $21.52 for Opus 5.5. The difference stemmed from Opus requiring higher turn counts and larger reasoning token footprints during interactive execution loops, consuming 18,795 reasoning tokens during triage and 135,672 reasoning tokens during cell lineage tracking.

See the GPT-6.1 Terra vs Claude Opus 5.5 output

Detailed turn metrics clarify how each system approached interactive problem-solving. During release triage, Terra concluded its run in 118 turns over 807 seconds, processing 22,117,582 input tokens and 7,659 output tokens, which included 2,718 reasoning tokens, yielding an estimated cost of $22.16. Claude Opus 5.5 required 170 turns and 934 seconds, handling 32,100,919 input tokens and 39,838 output tokens, including 18,795 reasoning tokens, at an estimated cost of $129.20. While both achieved identical perfect scores, Terra accomplished the maintenance tasks with 52 fewer conversational exchanges.

The computational gap expanded during the cell lineage challenge. Terra executed code 30 times across 37 turns, finishing in 920 seconds with 3,118,018 input tokens and 69,413 output tokens, costing $3.47. Opus 5.5 executed code 96 times across 129 turns over 2,187 seconds, consuming 19,722,068 input tokens and 201,194 output tokens for an estimated cost of $82.91. In the CI audit, Terra processed 56,959 input tokens and generated 61,539 output tokens in 691 seconds, costing $0.36. Opus exhausted 72,460 input tokens and hit its 128,000 output token maximum in 841 seconds without producing a parseable answer, costing $2.85.

The verdict on GPT-6.1 Terra vs Claude Opus 5.5

The head-to-head test indicates that GPT-6.1 Terra proved markedly more economical and faster than Claude Opus 5.5 when evaluated at maximum effort settings on these three synthetic tasks. Terra held parity in complex browser automation, achieved higher accuracy with one-quarter the code execution iterations on biological data tracking, and delivered a complete 94% accurate answer on unassisted tabular extraction where Opus 5.5 hit token ceilings. However, these outcomes reflect specific runs on tailored benchmarks rather than an absolute indicator of broad performance across diverse production workloads.

Operational differences also warrant careful consideration before selecting an API model for production pipelines. Anthropic explicitly recommends deploying Opus 5.5 at its default medium effort setting rather than maximum, reserving elevated reasoning levels for workflows with proven gains. Furthermore, Terra displayed recurring minor calculation faults, missing three questions on the tabular CI audit and miscalculating numerical sums in earlier comparative runs. Teams automating browser tasks and code-generation environments can run trial deployments of Terra at max effort, provided they verify numerical outputs on unstructured tables.