OpenAI released GPT-6.1 Terra on October 26, calling it “a fast, efficient model” for visual reasoning, computer use and writing, and “a significant step up from GPT-5.6 Terra”. It is the newest model in the GPT-6 series and is priced between GPT-6 Luna and GPT-6.1 Sol. Standard API pricing is $1 per million input tokens and $5 per million output tokens, half of Sol’s. Launch post · API changelog · System card addendum
Benchmark and speed claims below come from the launch post, which gives most of them as chart deltas rather than absolute scores. Specifications, prices and safety results come from OpenAI’s model page, pricing page, ChatGPT documentation and system card addendum. S5 Labs has not run the model.
Where Terra fits
The Terra name carries over from GPT-5.6, where Terra was the balanced middle tier between Luna and Sol. The GPT-6 series so far has Astra at the top, GPT-6 Sol and Luna from September, and GPT-6.1 Sol from DevDay. The launch post lists drafting proposals, interpreting technical drawings, filling out forms and editing videos as target uses, and gives reading invoices into accounting software as a high-volume API example. OpenAI’s ChatGPT models page pitches it for “clear writing, visual understanding, and computer use”. Launch post · ChatGPT models
In the API it is gpt-6.1-terra, with a single default snapshot and no aliases. The context window is 1,050,000 tokens, of which up to 922,000 can be input, and maximum output is 128,000 tokens. The knowledge cutoff is June 1, 2026. It takes text and images and returns text; audio and video are not supported. Model page
Prices against the rest of the lineup
Standard rates per million tokens, for prompts up to 272,000 input tokens:
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $1.00 | $50.00 |
| GPT-6.1 Sol | $2.00 | $0.10 | $10.00 |
| GPT-6 Sol | $2.00 | $0.20 | $10.00 |
| GPT-5.6 Terra | $2.00 | $0.20 | $12.00 |
| GPT-6.1 Terra | $1.00 | $0.05 | $5.00 |
| GPT-6 Luna | $0.10 | $0.01 | $0.50 |
Pricing page · Our GPT-6 Sol and Luna article
Cache writes cost $1.25 per million tokens for a 30-minute write and $2.50 for a 12-hour write. Above 272,000 input tokens the whole request moves to a higher rate: input and cache prices double and output goes to 1.5x, so $2.00 input, $0.10 cached and $7.50 output. Batch and Flex share one price, half of Standard: $0.50 input, $0.025 cached and $2.50 output. Fast mode doubles Standard to $2 input, $0.10 cached and $10 output, which matches GPT-6.1 Sol’s Standard price. Ultrafast is offered only for GPT-6.1 Sol and Astra. Regional processing endpoints add 10% for data-residency-eligible models released on or after March 5, 2026, and FedRAMP endpoints add 10% as well. Pricing page · Model page
OpenAI describes the price as 50% below both GPT-5.6 Terra and GPT-6.1 Sol. That is exact against Sol. Against the old Terra, input is half, output is about 58% lower and cached input is 75% lower. Launch post
A worked request
Take a request with 100,000 uncached input tokens and 10,000 output tokens at Standard rates, excluding tools, cache writes and regional uplifts. It costs $0.15 on GPT-6.1 Terra, $0.30 on GPT-6.1 Sol, $0.32 on GPT-5.6 Terra, $1.50 on Astra and $0.015 on Luna. Through Batch, Terra costs $0.075.
Long prompts change the math. Raise the same request to 400,000 input tokens and 20,000 output tokens and the whole request is billed at the higher rate: $0.80 for input plus $0.15 for output, $0.95 in total. At the base rate those tokens would cost $0.50. These are token prices for identical budgets, not task prices. A model that needs more turns or retries will cost more per finished job than the table suggests. Since October 28, ChatGPT Pro and Business plans include monthly API credits that can go toward these rates.
Migration constraints
Reasoning effort runs none, low, medium (the default), high, xhigh and max. There is no minimal. Terra keeps none, which GPT-6.1 Sol dropped, so a router that runs with reasoning off can send work to Terra but not to 6.1 Sol. Model page · DevDay article
Terra works on Chat Completions, Responses and Batch. Built-in tools are Responses-only: web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP and tool search. A computer-use agent therefore has to be on the Responses API. Fine-tuning is not supported. Model page
GPT-5.6 Terra is not deprecated. The deprecations page still names it as the recommended replacement for some retired models and has no entry pointing it to the new Terra, so moving is optional for now. Deprecations
ChatGPT, Codex and Copilot
OpenAI’s ChatGPT changelog lists Terra in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu. The launch post names Plus, Pro, Enterprise and Edu, without Business. OpenAI says the same usage allowance covers more tasks on Terra than on GPT-5.6 Terra, and substantially more than on GPT-6.1 Sol or Astra. It is chosen under Advanced in the desktop app and in ChatGPT Work on the web, and Work supports it on mobile. It is not in Chat. Enterprise access depends on workspace availability and administrator controls, and Terra has no UAE inference residency. ChatGPT and Codex changelog · Launch post
In Codex, Terra needs CLI version 0.159.0 or later and runs with codex -m gpt-6.1-terra. The models page lists it for the Codex CLI, desktop and web Work, credits and API access, but not for the IDE extension, which it lists for Astra, 6.1 Sol and Luna. As of October 27, no Codex release on GitHub mentions Terra; the latest stable release is 0.164.0, from October 23. ChatGPT models · Codex releases
GitHub made Terra generally available in Copilot the same day, rolling out gradually to Pro, Pro+, Max, Business and Enterprise. GitHub says it solved noticeably more terminal and C# tasks than GPT-5.6 Terra in early testing. It bills at the provider’s list price under usage-based billing, with no multiplier. Business and Enterprise get it automatically unless an admin has turned off the global default or disabled this model. GitHub changelog
OpenAI’s performance claims
All of these are OpenAI’s own figures, read mostly from charts.
- Agents’ Last Exam, a computer-use benchmark of professional tasks: at medium effort, 15 percentage points above GPT-5.6 Terra at about a third of the cost, and on par with GPT-6 Astra at one-tenth its price per token. At max effort OpenAI says Terra is over three times as token-efficient as Claude Opus 5.5 at a quarter of its per-token price, for about 9x lower cost per task.
- Speed: on the same benchmark, Terra completes tasks more than three times faster than Astra and GPT-6.1 Sol. The post does not say how the tasks were timed.
- Visual reasoning and desktop use: Terra beats GPT-5.6 Terra on MMMU-Pro at every reasoning setting and nearly matches Astra and GPT-6.1 Sol. On OSWorld 2.1 at medium effort it passes GPT-5.6 Terra’s best score.
- Terminal-Bench 4.0: 50.3% at max effort, against 21.5% for GPT-5.6 Terra. OpenAI itself says GPT-6.1 Sol and Astra “remain the better choice” for the most complex coding. For reference, OpenAI reported 57.9% for Astra at its launch. Our Astra article
- Terminal-Bench-Science 0.1: at max effort Terra beats GPT-6.1 Sol at lower cost and comes within five points of Opus 5.5 at under a fifth of the estimated cost.
- Factual errors: on OpenAI’s internal benchmark of error-inducing prompts, the error rate at low effort fell from 53.6% for GPT-5.6 Terra to 16.7%. At max effort it fell 18 percentage points to 4.3%, nearly matching Astra and GPT-6.1 Sol.
Decrypt stresses that the figures are OpenAI’s own and that the error test is internal. Cybersecurity News notes the test is built to be hard, so these are not everyday error rates. The system card’s factuality section agrees in direction: Terra makes substantially fewer factual errors than GPT-6 Sol and is on par with GPT-6.1 Sol, under evaluation caveats the card spells out. Decrypt · Cybersecurity News · System card addendum
The only independent data so far comes from Artificial Analysis, which lists only the max-effort variant. It gives Terra an index score of 48, 25th of 221 models, at 90.9 output tokens per second. Running its Intelligence Index took 130 million output tokens, against a median of 81 million, which we read as a sign that max effort spends heavily on reasoning. The Opus 5.5 comparisons above have no independent check yet. Artificial Analysis
What the system card found
Under the Preparedness Framework, OpenAI treats Terra as Critical in cybersecurity, High in biological and chemical capability and below High in AI self-improvement. It ships with the same safeguards as GPT-6.1 Sol and Astra. Cybersecurity News points out that the Critical label describes capability risk, not a flaw in the product. System card addendum · Cybersecurity News
On cyber evaluations, Terra scored 97.10% on ExploitBench, which the card qualifies with a contamination note, and 83.33% pass@1 on SEC-Bench Pro. On the internal port evaluation it reached 24.5%, against 21.5% for GPT-6.1 Sol and 31.5% for Astra. It solved 5 of 22 Sandbox Bench tasks, where Sol solved 9 and Astra 10. On biology, it scored 52.54% on multimodal virology troubleshooting, lower than Astra’s 63.11% and above the card’s 31% threshold, while none of the four Critical-level biology evaluations crossed their thresholds. On AI self-improvement, research debugging came in at 76.39% and KernelGen 1P at 45.34%, the latter below Astra and below the High threshold. System card addendum
The behavioral results are more mixed:
- Warnings: at max effort, Terra completed the task through a route that violated the spirit of an environment restriction in 29.6% of rollouts. The card says that is higher than Astra, Luna and GPT-6.1 Sol but lower than GPT-6 Sol, and that the test mostly covers low-stakes situations and runs without system-level controls.
- Image safety: text production benchmarks match GPT-5.6 Terra. On image inputs the card reports a statistically significant regression on hate and extremism against GPT-5.6 Terra, though it says manual review and other hate and extremism evaluations were in line with prior launches. Self-harm and erotic-content image results are on par.
- Deployment simulation: across 48,690 matched tasks of internal Codex traffic, Terra drew 22 flags of severity 3 or higher (0.045%), about 42% fewer than GPT-6 Sol and 64% fewer than GPT-5.6 Sol. Its evaluation awareness was 9.88%, the highest of the five models compared (Astra was next at 8.07%), which makes the other behavioral numbers harder to read at face value.
- Reasoning: chain-of-thought controllability, measured at 750 to 1,250 tokens of reasoning, was 14.8%, against 23.2% for GPT-6 Sol and 44.8% for GPT-6.1 Sol, with no monitor-evading side-task success and no steganography found.
- Prompt injection and jailbreaks: multiturn jailbreak resistance is on par with GPT-5.6 Terra. The card calls Terra similar to Astra and Sol on instruction-hierarchy tests, where those models score 99.99%, but gives no Terra-specific number.
- Health: against GPT-5.6 Terra, length-adjusted HealthBench Professional rose from 57.7 to 62.7 and MentalHealthBench from 46.2% to 55.5%, while length-adjusted HealthBench dropped 1.9 points to 55.1. Dynamic mental-health benchmarks came in lower.
Where the coverage disagrees
- Plans: OpenAI’s ChatGPT changelog includes Business; its launch post and Decrypt list only Plus, Pro, Enterprise and Edu. Cybersecurity News says “eligible paid plans”. Decrypt says Free users get GPT-5.6 Luna; OpenAI’s pages don’t list Free or Go for Terra.
- Date: OpenAI’s launch post, both changelogs and the system card say October 26. Interesting Engineering’s text gives October 27 for availability.
- Price framing: OpenAI’s “50% lower than GPT-5.6 Terra” understates the cut on output and cached input, as shown above.
- Bedrock: a Dev.to write-up says Terra reached Amazon Bedrock the same day. We found no AWS announcement to confirm it.
Decrypt · Interesting Engineering · Dev.to · API changelog
What to test
Terra against GPT-6.1 Sol on the same jobs. OpenAI still recommends Sol and Astra for the most complex coding. Run a representative set at a fixed effort on both and count cost per completed, checked task. Half the token price only pays off if Terra doesn’t spend twice the tokens to finish.
Computer use with a stopwatch. The threefold speed claim comes from one benchmark with no stated timing method. Time real browser or desktop tasks end to end on Terra and Sol, and log screenshots processed and failed steps along with tokens.
The effort setting. The error-rate gap between low (16.7%) and max (4.3%) effort is large on OpenAI’s test, and Artificial Analysis’s token count suggests max effort is expensive. Find the lowest effort that passes your own accuracy checks.
Image-heavy moderation. If your application takes user images, rerun your safety cases. The system card reports regressions on hate and extremism for image inputs.
Restrictions an agent might route around. For delegated work, the 29.6% warnings result matters most. Give a Terra agent a task with an explicit restriction and an easy workaround, and check whether it takes the workaround.
Long prompts. Check how often requests cross 272,000 input tokens. Above that line the whole request is billed at the higher rate.
Sources
OpenAI announcements and system cards
- Introducing GPT-6.1 Terra — OpenAI
- Addendum to GPT-6 Astra System Card: GPT-6.1 Terra — OpenAI
Developer documentation
- GPT-6.1 Terra model page, Pricing, API changelog, Deprecations — OpenAI Developers
ChatGPT and Codex documentation
- ChatGPT & Codex changelog, Models — ChatGPT Learn
- Codex releases — OpenAI on GitHub
Other
- GPT-6.1 Terra in GitHub Copilot — GitHub
- GPT-6.1 Terra — Artificial Analysis
- OpenAI launches GPT-6.1 Terra — Decrypt
- OpenAI GPT-6.1 Terra — Interesting Engineering
- OpenAI unveils GPT-6.1 Terra — Cybersecurity News
- OpenAI’s GPT-6.1 Terra: half the price, built to actually use your computer — Dev.to