GPT-6.1 Terra: BenchLM Public Benchmark Ledger and Comparison Method

One-sentence takeaway

This snapshot from 2026-10-26 brings GPT-6.1 Terra's public results, source labels, and coverage gaps together in one profile. Itemized results are labeled as provider-reported values; the overall score and ranking are labeled as estimates, and the page explicitly warns that the available evidence is insufficient to establish a verified overall position.

Use cases

• Suitable for: Reviewing Terra's benchmark entries, category positions, public coverage, and the evidence status of its overall ranking as listed by BenchLM through 2026-10-26.

• Not suitable for extrapolating to: Treating the platform's overall score as an independent measurement, taking the ranking as a definitive ordering of model quality, or inferring performance on unlisted tasks.

• Applicable model version: GPT-6.1 Terra (the profile identifies OpenAI as the provider and gives a release date of 2026-10-26).

• Test environment or client: BenchLM model profile and benchmark ledger; this is not a unified, independent testing environment.

• Reasoning effort and parameters: The profile note says canonical ledger rows use Max effort; the itemized table does not show run parameters for each result. Other reasoning-effort levels, prompts, tools, and repetition counts are not specified.

Evaluation method

• The profile is dated 2026-10-26; the data interface identifies the BenchAlign methodology version as bench-align-v5.8-2026-09-30.

• The page presents raw scores separately from the platform's aggregated rankings. All 12 displayed benchmark rows are labeled Provider exact and point to OpenAI release pages or the Terra system card supplement; this means the source is exact, not that BenchLM independently ran these tests.

• The profile says Terra has displayable evidence for 12/768 tracked benchmark slots. The Agentic category has 4 rows, Multimodal has 1, and Knowledge has 5; 2 additional rows are external signals.

• The page presents scores, rankings, percentiles, weights, row counts, and evidence status by category. Agentic, Multimodal, and Knowledge show category rankings; Coding, Reasoning, Multilingual, Instruction Following, and Math are marked unmeasured.

• The methodology page says the overall score is based on independent external leaderboard families combined after uncertainty calibration; direct benchmark evidence can adjust it by at most 3 points. Without independent external sources, the overall position is Estimated. The estimator uses model signals, calibrated benchmark estimates, lineage priors, older policy estimates, and qualifying peer models; missing benchmarks remain blank rather than receiving fabricated scores.

• The methodology page also discloses that several 95th-percentile error checks worsened in the 2026-09-30 retrospective validation. These estimated intervals do not represent confidence in a rank or certainty about the ordering of the entire leaderboard.

Key results

• Overall leaderboard: 75.87/100, #10/229, evidence status Estimated. The profile also cautions that public ranking coverage is not yet sufficient to establish a verified position.

• Categories: Agentic 72.4, #8/133; Multimodal 91.6, #4/58; Knowledge 74.9, #12/188. Coding, Reasoning, Multilingual, Instruction Following, and Math are all unmeasured.

• Agentic results: Terminal-Bench 4.0: 50.30%; Agents' Last Exam: 60.0%; Terminal-Bench-Science 0.1: 58.3%; OSWorld 2.1 Offline: 74.2%, identified as an offline subset of 82 tasks.

• Multimodal result: MMMU-Pro: 85.5%.

• Knowledge results: HealthBench raw: 51.1%; length-adjusted: 55.1%; HealthBench Professional: 62.7%; Professional raw: 64.2%; HealthBench Hard: 31.4%.

• External-signal results: ExploitBench: 97.1%; SEC-Bench Pro: 83.33%. The profile lists both as display-only; they should not be counted as direct ranking benchmarks in this page's aggregate ranking.

• Speed: Unmeasured. The page provides no independently run speed or time-to-first-token data.

Raw data

The profile also shows input/output API prices of $1/$5 per million tokens, cached input at $0.05, and a blended price of $3 per million tokens. This is pricing comparison information, not an evaluation result.

Conclusions and limitations

• This profile is a snapshot indexing and comparing public benchmarks and their evidence status. It does not say that BenchLM ran a unified evaluation of Terra, and it does not disclose per-item run logs, prompts, sample counts, or complete tool configurations.

• The itemized data in all 12 records comes from OpenAI, not independent benchmark runs. The profile's distinctive value is that it consolidates scores, sources, coverage, peer positions, and the limitations of estimated rankings.

• The overall score and rank cannot be recomputed by simply combining the itemized percentages. BenchLM's estimation method depends on current inputs, calibrated estimates, model lineage, and other factors; the methodology page explicitly describes uncertainty in estimated positions.

• Benchmarks with the same name may not be directly comparable when versions, subsets, tools, or evaluation configurations differ. In particular, the OSWorld row is an 82-task offline subset of OSWorld 2.1 Offline, not a result for a complete OSWorld evaluation.

• This dynamic profile reflects a snapshot from 2026-10-26; the leaderboard, source coverage, and estimates may change.

Reproduction notes

On 2026-10-27, the model profile, its public benchmark data interface, and the BenchLM methodology page were checked through Tabbit. The data refresh date, BenchAlign version, 12 itemized results, source labels, category summaries, and methodology limitations were cross-checked. No model inference or independent retesting was performed; the page does not disclose the complete underlying run configuration.