What happened
The ExtractBench repository added Codex GPT-6.1 Terra to its Pro leaderboard in commit #99. The change covers two reasoning configurations, labelled Medium and Low, and reports their results across all six benchmark challenges. The repository says both the time-limited and no-time-limit views are available; the README leaderboard presents the time-limited ranking while the CSV contains the fuller set of scores.
In the README’s ranked top ten, Codex GPT-6.1 Terra Medium enters at number seven among 15 systems. Its aggregate row covers 100 cases, with 81 completed runs. It records a 27.00% perfect rate, 74.19 Value F1, 60.16 cost-related score, 64.80 in the final reported aggregate column, and a listed cost of 4.06 cents. The leaderboard is ordered first by percentage perfect, then by Value F1, then by lower cost. That ordering matters: Terra Medium’s placement is not a claim that it is the highest-scoring system on every measure.
The added per-challenge rows show a mixed profile. At medium effort, Terra reaches 87.68 Value F1 on charts inside documents and 74.38 on drawn relations. Its long-chart-sheets row is much weaker, at 34.57 Value F1 and 0.89 in the adjacent cost-related measure. The results therefore describe capability under a particular set of extraction tasks rather than a uniform property of the model.
The same commit adds low-effort results. Across all tasks, Codex GPT-6.1 Terra Low has a 11.00% perfect rate, 51.30 Value F1, 52.72 in the cost-related column, 56.53 in the final aggregate column, and a listed cost of 2.62 cents. Its results vary sharply by challenge: charts inside documents reaches 65.46 Value F1, while long chart sheets reaches 16.20. The benchmark does not explain why those task-level differences occur.
For context within the table, Terra Medium’s 74.19 Value F1 is close to Codex GPT-6.1 Sol’s 74.06, while its displayed cost is 4.06 cents versus Sol’s 13.86 cents. But Sol retains a higher listed cost-related score, 63.96 versus Terra Medium’s 60.16, and the two systems have different perfect-run rates. Those figures are benchmark observations, not a universal recommendation to replace one configuration with another.
What the archive adds
This is a new benchmark fact about a known model family, rather than a restatement of an earlier product-access story. Netlify adds GPT-6.1 Terra to its AI Gateway and Agent Runners concerned availability through Netlify’s platform; this update measures a Codex configuration on document extraction.
Trend signal
This modestly supports Agent economics moves to completed work. ExtractBench exposes a more operational comparison than token price alone by putting completed cases, perfect outcomes, extraction quality, and displayed cost in the same table. It does not establish production economics, because the benchmark’s tasks and scoring are narrower than real workflows.
Why it matters
Teams using coding agents for document-heavy work often face a configuration decision rather than a simple model-selection decision: how much reasoning effort is worth purchasing for a given class of extraction task? This update gives one public data point. On this benchmark, medium effort substantially improves aggregate Value F1 and perfect-run rate relative to low effort, while increasing the displayed cost from 2.62 to 4.06 cents.
That can inform a practical evaluation plan. Use low effort for work where a downstream validator, human review, or retry loop can cheaply catch errors; reserve medium effort for tasks where structured extraction errors are costly or hard to detect. The task-level spread is equally important. A single aggregate score can conceal difficult document forms, particularly long chart sheets, where neither setting is reliably strong enough to run without safeguards.
What happens next
- More agent configurations will be added to the leaderboard. (3 months, 65%). Watch for: new rows using the same six challenge categories and both time-limit views.
- Builders will test effort routing rather than choose one default setting. (6 months, 60%). Watch for: evaluations that assign effort by document type and verification cost.
Limits and uncertainty
The source is a repository commit and leaderboard data, not a peer-reviewed evaluation report or a production case study. It does not describe the prompts, tool configuration, hardware, sampling settings, retry policy, or the exact interpretation of every reported CSV column. It also does not establish statistical uncertainty, representativeness beyond the six challenges, or comparative performance on ordinary coding work. The displayed costs are benchmark-specific measurements and should not be treated as published API pricing or a forecast of a team’s operating cost.
Skill to build now
Create a small internal extraction test set with at least three document shapes that matter to your team, such as tables, diagrams, and long spreadsheets. Run the same job at two reasoning settings, record exact-field accuracy, human correction time, retry rate, and per-run cost. Then route work by the cost of an undetected error, not by a model’s aggregate leaderboard position.
Reader takeaway
Codex GPT-6.1 Terra Medium performs materially better than its low-effort setting on ExtractBench Pro’s aggregate document-extraction tasks, but the benchmark also exposes substantial weakness on some document types. Treat the result as evidence for workload-specific effort routing and local evaluation, not as a blanket model ranking.
Open question
Will the benchmark publish enough configuration and methodology detail for teams to reproduce the cost-quality trade-off on their own extraction pipelines?