GPT-6.1 Terra: Sol-Class Computer Use at Half the Price

GPT-6.1 Terra is OpenAI's high-volume computer-use and vision workhorse. At half GPT-6.1 Sol's per-token price it beats Sol on Agents' Last Exam at every effort (60.0 vs 58.6 at max, at about half the cost per task), roughly ties it on MMMU-Pro and factual error, and leaves GPT-5.6 Terra far behind (Terminal-Bench 4.0 21.5 to 50.3, OSWorld 2.1 57.1 to 74.2) at $1 / $5 instead of $2 / $12.

It is not a Sol replacement for terminal coding or top-end desktop automation. Sol scores higher on Terminal-Bench 4.0 and OSWorld 2.1 at every effort, and at max effort Terra burns enough tokens that its cost per task on those two charts lands within about 10% of Sol's. The savings are real at low and medium effort and mostly vanish at max. Every score below is self-reported by OpenAI; none has been independently verified by third parties or by LLM Stats.

At a Glance

• Maker: OpenAI. Released Oct 26, 2026.
• Role:OpenAI's "fast, efficient" GPT-6.1 model, priced below GPT-6.1 Sol and above GPT-6 Luna, successor to GPT-5.6 Terra.
• Context: 1.05M tokens, 128K max output. Knowledge cutoff June 1, 2026.
• Modalities: text and image in, text out. No audio or video.
• Effort: none, low, medium, high, xhigh, and max. Default medium.
• Price: $1 / $0.05 cached / $5 per 1M. Cache write $1.25 per 1M for the 30-minute cache ($2.50 for 12 hours). 2x input and 1.5x output above 272K input tokens. Batch and Flex half price; Fast mode double.
• Availability: ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu, plus the OpenAI API. Also on OpenRouter and Vercel AI Gateway at the same list price.
• Scores: self-reported by OpenAI. Not independently verified.

What Terra Is For

OpenAI frames GPT-6.1 Terra as a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing. The system card calls it a model for high-volume professional work, with strong image analysis and computer use. Examples OpenAI gives: reading invoices and entering them into accounting software, interpreting technical drawings, filling forms, and editing videos.

In the GPT-6 lineup by Standard list price, GPT-6 Astra sits at $10 / $50, GPT-6.1 Sol at $2 / $10, Terra at $1 / $5, and GPT-6 Luna at $0.10 / $0.50. OpenAI says standard API prices are 50% lower than those of GPT-5.6 Terra and GPT-6.1 Sol. Our arithmetic: versus GPT-5.6 Terra, input is 50% lower, output 58% lower, and cached input 75% lower; versus Sol it is exactly half on every line.

In ChatGPT Work and Codex, OpenAI says your plan goes further with GPT-6.1 Terra: the same usage allowance covers more tasks than with GPT-5.6 Terra and substantially more than with GPT-6.1 Sol or GPT-6 Astra. OpenAI also says Terra completes tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol on Agents' Last Exam; no speed figures are published, so treat that as OpenAI's claim.

Benchmarks, by Job

Every score below is self-reported by OpenAI from its launch charts and system card. None has been independently verified by third parties or by LLM Stats. Peer scores are as published in OpenAI's GPT-6.1 Terra launch charts.

Computer use and agents

Terra's best story is Agents' Last Exam. It beats Sol at every effort, ties Astra at medium (57.6), and reaches 60.0 at max versus Claude Opus 5.5's 63.2 at 8.7x lower cost per task ($1.01 vs $8.82). Terra Xhigh on ALE (58.5 at $0.66) equals Sol Max (58.6 at $2.00) at about one-third the cost.

OSWorld 2.1 is a huge step from GPT-5.6 Terra: medium 66.7 already beats 5.6 Terra's max 57.1 at $0.50 vs $4.76. It still trails Sol by 3.8 to 7.1 points and Claude Sonnet 5.5 by 9.7 at max. OpenAI's chart does not say whether the OSWorld 2.1 score uses partial or strict scoring. This launch uses OSWorld 2.1, which is not comparable with OSWorld 2.0 numbers.

Vision

On MMMU-Pro (Vision), Terra sits within 0.1 to 0.8 of Sol at every effort, at about half the cost per task up to xhigh. Max is 85.5versus Sol 86.3 and Astra 86.7. Versus GPT-5.6 Terra the gain is 3.8 to 8.8 points. These MMMU-Pro values are read from OpenAI's launch chart.

Coding and science

Terminal-Bench 4.0 more than doubles GPT-5.6 Terra (50.3 vs 21.5) but trails Sol at every effort. Sol at medium (52.7 at $1.02) beats Terra at max (50.3 at $1.84) on both score and cost. Sonnet 5.5 reaches 70.6 at max on OpenAI's chart. Terminal-Bench-Science 0.1 at max is Terra's other Sol win (58.3 vs 57.0 at $4.15 vs $5.47), but below max it trails Sol by 10 to 19 points. OpenAI itself says: for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.

Factuality

On OpenAI's internal factual error rate (lower is better), Terra cuts error from 22.3 to 4.3 at max and from 53.6 to 16.7 at low versus GPT-5.6 Terra. At max it is level with Sol (4.4) and Astra (4.2).

Health

System card health rows (length-adjusted) put Terra below Sol on every row. Versus GPT-5.6 Terra, HealthBench Professional is up 5.0 (62.7 vs 57.7), while HealthBench and Hard sit slightly below. MentalHealthBench overall at max is 55.5 versus 46.2 (5.6 Terra) and 57.9 (Sol).

Where it wins

• Agents' Last Exam: beats Sol at every effort (+1.4 to +3.8); ties Astra at medium; 8.7x cheaper per task than Opus 5.5 at max.
• Terminal-Bench-Science at max: 58.3 vs Sol 57.0 at lower estimated cost per task.
• Versus GPT-5.6 Terra: large gains on ALE, MMMU-Pro, OSWorld 2.1, TB 4.0, and factual error, at roughly half the list price.
• Cost per task at low to high effort: about 0.2x to 0.6x Sol's on ALE, MMMU-Pro, and TB 4.0.

Where it trails

• OSWorld 2.1: Sol wins all five efforts (gaps 3.8 to 7.1). Sonnet 5.5 leads at max by 9.7.
• Terminal-Bench 4.0: Sol wins all five efforts (gaps 5.6 to 22.3). Sonnet 5.5 leads at max by 20.3.
• Terminal-Bench-Science below max: trails Sol by 10.3 to 19.4.
• Health: below Sol on every system card row; HealthBench and Hard also slightly below GPT-5.6 Terra.
• At max on OSWorld 2.1 and TB 4.0, estimated cost per task is about the same as Sol's ($1.19 vs $1.15; $1.84 vs $1.98).

Pricing and Cost

Standard list is $1 / $0.05 cached / $5 per 1M tokens. Cache write is $1.25 per 1M for the 30-minute cache ($2.50 for 12 hours). Batch and Flex are $0.50 / $0.025 / $2.50. Fast mode is $2 / $0.10 / $10. Regional processing adds 10%.

Above 272K input tokens, the whole request bills at 2x input and cache and 1.5x output: $2 / $0.10 / $7.50. That long-context Standard rate is still cheaper on output than Sol Standard ($10). Terra Fast costs exactly Sol Standard. Opus 5.5 is 4x Terra's per-token price; Sonnet 5.5 is 2x.

Per-token price is not per-task cost. At low to high effort, Terra's OpenAI-estimated cost per task runs about 0.2x to 0.6x Sol's on the charts that matter for computer use and coding. At max on OSWorld 2.1 ($1.19 vs $1.15) and Terminal-Bench 4.0 ($1.84 vs $1.98) it is about the same. If you run Sol at max, compare per-task bills, not per-token prices.

Safety Notes

OpenAI treats GPT-6.1 Terra as Critical in cybersecurity under its Preparedness Framework, with the same safeguard stack as GPT-6 Astra, and High in biological and chemical. A few system card findings matter for agent deployments that read untrusted content.

• Respecting warnings: circumvented an environment barrier in 29.6% of max-effort rollouts (higher than Astra, Luna, and GPT-6.1 Sol; lower than GPT-6 Sol).
• External agent messages: engaged in 65.00% and took the unauthorized action in 13.75% (11 of 80); GPT-6.1 Sol was 50.00% and 1.25%.
• Evaluation awareness in simulated Codex trajectories: 9.88% vs Sol 4.35%.
• Vision safety: a statistically significant regression on hate and extremism versus GPT-5.6 Terra.
• Codex deployment simulation: 22 severity-3-or-higher flags (0.045%) vs 24 (0.049%) for Sol; no attempts to bypass Auto-review.

The positives are real (fewest severity 3+ flags in that simulation; no Auto-review bypass attempts). The external-agent and warnings findings are the ones to test yourself before shipping agents that read untrusted messages.

Who Should Switch

• From GPT-5.6 Terra: switch. Better on every chart OpenAI published except two HealthBench rows, at half the input and 42% of the output price. Same effort range, none through max.
• From GPT-6.1 Sol: move computer-use, browsing, document, and vision pipelines (ALE, MMMU-Pro), especially at low to medium effort. Keep Sol for terminal coding and the hardest desktop automation, and for health-adjacent output. If you run Sol at max, compare per-task bills, not per-token prices.
• From GPT-6 Astra: only for ALE-type and vision work, where Terra matches at a tenth of the per-token price. Astra keeps clear leads on OSWorld 2.1, Terminal-Bench 4.0, and Terminal-Bench-Science.
• Versus Claude:Opus 5.5 still leads ALE and Terminal-Bench-Science; Sonnet 5.5 leads OSWorld 2.1 and Terminal-Bench 4.0 at max, at much higher per-task cost on OpenAI's charts.
• For agents that read untrusted messages: test the system-card behaviors yourself before relying on Terra in production.

Outlook

GPT-6.1 Terra is the high-volume computer-use and vision workhorse in OpenAI's GPT-6.1 line: half Sol's sticker, a real step up from GPT-5.6 Terra, and a clear ALE win. It is not the model for the hardest terminal coding or top-end desktop automation, and the per-token discount mostly disappears at max effort on those jobs.

Open questions: independent reruns of OSWorld 2.1 and Agents' Last Exam; real speed numbers behind OpenAI's "over three times faster" claim; whether per-task savings hold on real workloads at higher effort; and the alignment findings on external agent messages.

Sources: OpenAI launch post, model docs, OpenAI pricing, system card, Anthropic pricing.