🌏 中文版
Model Details
| Field | Value |
|---|---|
| Model ID | gpt-6.1-terra |
| Vendor | OpenAI |
| Parameters | Undisclosed |
| Context Window | 1,050,000 tokens |
| Input pricing (USD/1M tokens) | $1.00 (cached $0.05) |
| Output pricing (USD/1M tokens) | $5.00 |
| Open source | No |
| Release date | 2026-10-26 |
| Official announcement | OpenAI: Introducing GPT-6.1 Terra |
| Family | GPT-6.1 (sibling GPT-6.1 Sol released 2026-09-29, $2 input/$10 output; flagship GPT-6 Astra at $10/$50; predecessor GPT-5.6 Terra at $2/$12) |
Highlights
- On Agents' Last Exam (a hands-on computer-use benchmark) at medium effort, it scores 15 points above predecessor GPT-5.6 Terra at about 1/3 the cost, matching flagship GPT-6 Astra's score while costing just 1/10 as much per token
- At max effort, it's over 3x more token-efficient than Claude Opus 5.5 and costs 1/4 as much per token, pushing per-task cost down to roughly 1/9 of Opus 5.5's; it also completes tasks over 3x faster than GPT-6 Astra and GPT-6.1 Sol
- On Terminal-Bench 4.0 (command-line coding agent tasks) at max effort, it reaches 50.3%, more than double the predecessor's published 21.5%
- Internal factual-error rate at low effort drops from the predecessor's 53.6% to 16.7% (about a 37-point drop); at max effort it falls to 4.3%, now close to flagship GPT-6 Astra and GPT-6.1 Sol
Benchmark Results
| Benchmark | Score | Predecessor (GPT-5.6 Terra) | Best competitor |
|---|---|---|---|
| Agents' Last Exam (medium effort) | +15pp over predecessor, at ~1/3 the cost | Baseline (absolute score not disclosed) | Matches GPT-6 Astra at 1/10 the cost |
| Terminal-Bench 4.0 (max effort) | 50.3% | 21.5% | No other model's score disclosed in this announcement |
| Internal factual-error rate (low effort, lower is better) | 16.7% | 53.6% | Competitor scores at this setting not disclosed |
| Internal factual-error rate (max effort, lower is better) | 4.3% | 22.3% (back-calculated from OpenAI's stated "18-point drop") | Close to GPT-6 Astra and GPT-6.1 Sol |
| Terminal-Bench-Science 0.1 (max effort) | Beats GPT-6.1 Sol at lower cost; within 5 points of Claude Opus 5.5 | Not provided in this announcement | Trails Claude Opus 5.5 by ~5 points, at under 1/5 the cost |
⚠️ All figures are OpenAI's own benchmarking under vendor-chosen reasoning-effort settings, pending independent reproduction. Most rows are the relative gaps OpenAI published rather than both models' absolute scores — exact numbers for GPT-5.6 Terra and competitors weren't fully disclosed in this announcement.
Source: OpenAI's own benchmark figures from its 2026-10-26 announcement. The max-effort predecessor baseline is back-calculated from the disclosed drop: 4.3% + 18 points = 22.3%. Coordinate conversion: baseline y=300, 4 units per percentage point (y = 300 − value × 4), axis starts at 0, domain 0–60.
Versus the Predecessor and Competitors
Compared with GPT-5.6 Terra, standard pricing is cut in half outright: input drops from $2.00 to $1.00, output from $12.00 to $5.00 — a pricing strategy, not just a score bump. Scores didn't suffer for it: Terminal-Bench 4.0 more than doubled from 21.5% to 50.3%, and the low-effort factual-error rate dropped by roughly two-thirds. This generation of Terra gets both cheaper and stronger at the same time, rather than trading one for the other.
Against flagship GPT-6 Astra, Terra already matches Astra's score on Agents' Last Exam at just 1/10 the per-token price, and nearly matches both Astra and GPT-6.1 Sol on MMMU-Pro visual reasoning. That pushes OpenAI's own product line downward in price: computer-use precision that used to require the flagship tier is now within reach of a mid-priced model.
Against Claude Opus 5.5, Terra is 3x more token-efficient at max effort and costs 1/4 as much per token — multiplying out to roughly a 9x gap in per-task cost. But on Terminal-Bench-Science 0.1, the benchmark that leans hardest on scientific reasoning, Terra still trails Opus 5.5 by about 5 points. OpenAI itself isn't claiming this generation of Terra sets a new ceiling; the pitch is "good enough score, dramatically lower cost."
Implications for Agent Development
This release centers on two capabilities — computer use and visual reasoning — aimed at agent scenarios that run high volumes of repetitive work without needing top-tier reasoning on any single task.
- If you're building agents that read screens, operate desktop apps, fill out forms, or edit video: Terra already beats the predecessor's best score on OSWorld 2.1 (multi-step desktop tasks) at medium effort, at lower cost — currently the best price-performance option for this kind of workload
- If you're running high-volume, budget-constrained pipelines (e.g., reading invoices and entering them into accounting software): standard pricing halved to $1 input / $5 output, plus $0.05 cached input, meaningfully lowers the cost of repetitive batch work
- If you're building coding agents: the more-than-doubled Terminal-Bench 4.0 score is worth re-evaluating against mid-complexity tasks currently routed to Sol or Astra — though OpenAI explicitly says the most complex coding tasks still call for GPT-6.1 Sol or GPT-6 Astra
- Not a fit: workloads that need top-tier scientific reasoning — Terra still trails Claude Opus 5.5 by about 5 points on Terminal-Bench-Science 0.1, a gap this release didn't close
Today's Takeaway
This release doesn't challenge any score ceiling, yet it halves standard pricing and cuts the factual-error rate by more than two-thirds — suggesting OpenAI's current priority for its mid-priced tier isn't chasing a new record, but delivering flagship-grade computer-use capability at a tenth of the price. For agent developers, a release that holds scores flat while shifting the entire cost structure downward is often more worth recalculating your pipeline's unit economics over than a few extra points on a leaderboard.