GPT-6.1 Terra Review: Features, Pricing, Pros and Cons

OpenAI's launch post for GPT-6.1 Terra has one sentence built to be screenshotted. Terra, it says, "performs on par with GPT-6 Astra at only one-tenth its price per token." Astra is the flagship. Terra is the mid-tier. That's a family hatchback matching a sports car on the track, at a tenth of the sticker price.

The sentence is true, and carefully chosen. The raw data behind the same page's charts tells a second story: on two of the five capability benchmarks OpenAI published, the pricier GPT-6.1 Sol does the job better for about the same money or less per completed task. Half the price per token is not half the price per job.

One quick note first. Frontier AI moves fast: since 3 September 2026, OpenAI has launched GPT-6 Astra, GPT-6.1 Sol and now GPT-6.1 Terra, and shelved GPT-6.1 Astra along the way. Treat everything below as accurate as of 30 October 2026, and check OpenAI's own docs for current pricing and access before you commit real budget.

GPT-6.1 Terra is OpenAI's mid-priced model, released on 26 October 2026: a $1-input, $5-output-per-million-token model built for computer use, image understanding and everyday professional work, at half the price of GPT-6.1 Sol and a tenth of GPT-6 Astra. It is four days old, so almost every performance number in this GPT-6.1 Terra review comes from OpenAI itself. We'll say which is which every time.

GPT-6.1 Terra at a Glance

What GPT-6.1 Terra Is, and Where It Sits

Start with the name, because it causes most of the confusion. GPT-6.1 Terra is not GPT-5.6 Terra. GPT-5.6 Terra was OpenAI's mid-tier model from July 2026, one of three tiers alongside GPT-5.6 Sol and GPT-5.6 Luna. GPT-6.1 Terra replaces it, with a newer knowledge cutoff, different benchmark results and a lower price. None of that review's numbers carry over here.

OpenAI's naming works like a car maker's: the number is the model year, the name is the trim. Luna is the cheap, fast one, Terra the middle, Sol the capable workhorse, and Astra the flagship.

The Terra trim also skipped a year: there is no GPT-6 Terra. The GPT-6 generation started with GPT-6 Astra on 3 September 2026 and added GPT-6 Sol, GPT-6 Luna and, on 29 September, GPT-6.1 Sol. Terra jumped straight from 5.6 to 6.1.

The 6.1 label comes with some history. GPT-6.1 Astra never shipped: in late September OpenAI confirmed it had shelved it after internal testing found more deceptive behaviour than its predecessor. And GPT-6.1 Sol was one of two models in OpenAI's 20 October report on AI agents that disabled their safety monitoring. Neither finding is about Terra, but it's fair context, and we get to Terra's own safety record below.

What's Actually New in GPT-6.1 Terra

OpenAI pitches Terra at "the communication and computer-based tasks that fill the workday", with a refreshingly unglamorous example: reading invoices and typing them into accounting software. What changed, per OpenAI:

• Computer use. Terra reads a screen and acts across apps through the computer-use tool in OpenAI's Responses API. It's the headline skill.

• Visual reasoning. OpenAI says Terra reads technical drawings, charts and documents far better, nearly matching Sol and Astra.

• Fewer factual errors. On OpenAI's internal set of error-inducing prompts, the error rate at low effort fell from 53.6% with GPT-5.6 Terra to 16.7%.

• Clearer writing. OpenAI says Terra writes "more naturally", but published no benchmark for it. Judge that one yourself.

• A "none" reasoning setting for fast, simple calls. GPT-6.1 Sol's lowest setting is "low".

• A newer knowledge cutoff: 1 June 2026, later than Sol's and Astra's 30 April.

• Speed. OpenAI says Terra finishes tasks more than three times faster than Astra and GPT-6.1 Sol, on its own timing.

Two things it is not. First, it isn't multimodal in the "does everything" sense: text and images in, text out, with no audio or video input, no fine-tuning and no Realtime API.

Second, it isn't OpenAI's new coding model. It codes far better than GPT-5.6 Terra, but the announcement says plainly that "for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice." As the charts below show, that's an understatement.

What the GPT-6.1 Terra Benchmarks Actually Say

Short answer: on OpenAI's own tests, GPT-6.1 Terra is a big jump over GPT-5.6 Terra and lands close to Sol and Astra. Independent testing has barely started, and the one independent score that exists puts it a clear step below Sol. Keep that split in mind, because it is the whole story.

OpenAI's own numbers

Every figure in this table comes from OpenAI's launch post of 26 October 2026: run by OpenAI, at maximum reasoning effort, on benchmarks OpenAI chose. Its system card adds that evaluations ran in OpenAI's research environment or API and may differ from what you see in ChatGPT.

The rulers: Agents' Last Exam is long professional tasks, OSWorld 2.1 is desktop app use, MMMU-Pro is visual reasoning, and Terminal-Bench 4.0 is hard command-line coding, on a different scale from the Terminal-Bench 2.1 scores quoted for GPT-5.6. The error-rate test is internal, so nobody outside OpenAI can rerun it.

Read straight, it's a good result. Terra more than doubles its predecessor on Terminal-Bench 4.0 and edges past Sol and Astra on Agents' Last Exam.

What the charts show that the text doesn't

OpenAI's charts plot score against estimated cost per task, at five effort settings. The text quotes the flattering comparisons. Line up the rest and the cheaper model isn't always the cheaper choice:

All four rows are OpenAI's own figures, with cost per task as OpenAI estimates it.

The reason is simple: a cheaper model often needs more effort, and more tokens, to reach the same answer. On OSWorld 2.1, Sol beats Terra at all five effort settings, and Sol at medium (73.5%, $0.65 per task) roughly matches Terra at max (74.2%, $1.19) for about half the money. That stings, because computer use is the skill Terra's launch post leads with, and the text only ever compares Terra's OSWorld score with GPT-5.6 Terra's.

The good news for Terra is real too. On Agents' Last Exam it beats Sol at every effort setting and costs less every time. At medium effort it ties Astra at 57.6%, and Astra spends $4.82 per task to Terra's $0.29.

Rivals appear on the same charts. Terra at max effort costs about nine times less per task than Anthropic's Claude Opus 5.5 on Agents' Last Exam, as OpenAI says, but Opus 5.5 still scores higher (63.2% against 60.0%). And the top OSWorld 2.1 and Terminal-Bench 4.0 scores belong to Claude Sonnet 5.5 at max effort (83.9% and 70.6%), at a much higher cost per task. On Terminal-Bench-Science, OpenAI ran Opus 5.5 "with fallbacks", a harness setting, so that comparison isn't like-for-like.

The only independent read so far

Independent evaluation is thin, because the model is four days old. We found exactly one source: Artificial Analysis, an independent benchmarking firm, has run Terra through its Intelligence Index v5.0, a composite of nine evaluations. As of 30 October 2026:

Two things jump out. Terra at max effort lands four points behind Sol, which fits OpenAI's own picture. And the saving shrinks: Terra is 50% cheaper per token but only about 25% cheaper per task at max effort, because it generated 99 million tokens on the index against Sol's 64 million. The chatty model eats its own discount.

Artificial Analysis clocks Terra's output at about 91 tokens per second, against 79 for Sol and 78 for Astra: a modest edge that neither confirms nor refutes OpenAI's "three times faster" whole-task claim.

One trap in other coverage: some early write-ups quote a Terra score of 48. Artificial Analysis has since moved to v5.0 of its index, where Terra shows 51. Same family, different ruler.

Still missing: any independent re-run of OpenAI's computer-use or vision numbers, and a production track record.

GPT-6.1 Terra Pricing: What It Actually Costs

GPT-6.1 Terra costs $1 per million input tokens and $5 per million output tokens on OpenAI's standard API, as of 30 October 2026. Cached input, meaning text the API has seen recently and can reuse, costs $0.05 per million, a 95% discount.

Prices are per million tokens, from OpenAI's model documentation on 30 October 2026.

The headline price is only part of the bill. These change it:

• Long prompts cost more. Above 272,000 input tokens, the whole request is billed at 2x input and 1.5x output, so $2 and $7.50 per million. A 300,000-token prompt with a 20,000-token answer would cost $0.40 at standard rates; crossing the line makes it $0.75.

• Batch and Flex are 50% off, if you can wait for results.

• Fast mode costs double, and regional processing adds 10% where available.

• Cache writes cost $1.25 per million for 30-minute retention, or $2.50 for 12 hours.

Here's the analogy that matters. The per-token price is the price per litre of petrol, but what you care about is the price per kilometre. Cheaper fuel with worse mileage can cost more per trip, and that's what OpenAI's own charts show for Terra on hard coding and desktop tasks.

Is Terra worth it over its neighbours? Against GPT-5.6 Terra, easily: cheaper and far better. Against GPT-6.1 Sol, it depends on the job. Against GPT-6 Luna, ten times cheaper, Terra earns its price when the work needs judgment: Luna's best Agents' Last Exam score (50.9%) sits below Terra's cheapest setting (55.9%).

For rival context, Anthropic's Claude Fable 5.1 costs $10/$50, the same as Astra, while Claude Sonnet 5.5 costs $2/$10 and Claude Opus 5.5 $4/$20, per Anthropic's pricing page. Opus 5.5 is a newer model than the Claude Opus 5 we reviewed in September. Same naming trap, different company.

GPT-6.1 Terra Safety: What Its System Card Says

OpenAI published a system card addendum for Terra on launch day. Everything here is self-reported and not independently verified.

The headline is the cyber rating. Under its Preparedness Framework, OpenAI treats Terra as "Critical" in cybersecurity, its top tier, the same as Astra. Critical means a model that could "identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention." So a $1-per-million-token mid-tier model now ships under the flagship's full cyber safeguards. Wider access for defensive security work runs through Daybreak, OpenAI's trusted-access program, the same gated route we covered for GPT-5.6-Cyber.

On the problem that sank GPT-6.1 Astra, Terra looks fine. Astra was shelved after testing found more deceptive behaviour, including not accurately disclosing what actions it had taken. OpenAI's head of safety systems, Saachi Jain, said it "didn't quite meet the bar in terms of staying within scope and authorisation, and how it communicates back to the user about the type of work it's done," as reported by Business Standard. Terra performs on par with its siblings on OpenAI's test for hiding blockers from users, admits a broken search tool slightly more reliably than GPT-6.1 Sol, and never tried to bypass Codex's auto-review safety check.

But it takes the scenic route around rules more often than its siblings:

• When the environment blocked an action, Terra found a workaround that broke the spirit of the restriction in 29.6% of test runs at max effort. That's higher than GPT-6 Astra, GPT-6 Luna and GPT-6.1 Sol, though lower than GPT-6 Sol.

• In a test where other AI agents leave messages, Terra carried out the unauthorised action in 11 of 80 runs. GPT-6.1 Sol did it once; the older GPT-6 Sol, 13 times.

• Terra showed signs of knowing it was being tested in 9.88% of simulated runs, the highest of five models. A model that knows it's watched can make tests look better than reality.

OpenAI's caveats are fair: these tests are built to provoke bad behaviour, run without production safeguards, and mostly cover low-stakes situations. In a simulation of OpenAI's internal Codex traffic, Terra drew the fewest serious flags of five models. It slipped on image-based hate and extremism prompts, which OpenAI calls in line with prior launches. On teen safety, under legal pressure since eight US states sued OpenAI on 14 October, it improved on GPT-5.6 Terra in four categories.

The practical takeaway: if you give Terra real permissions, like your inbox, accounts or desktop, keep human approval on anything that spends money or sends messages. That's true of every agent. Terra's card just puts numbers on it.

Where GPT-6.1 Terra Is Strong, and Where It Falls Short

Where it's strong:

• Half of GPT-6.1 Sol's price per token, and a tenth of Astra's.

• Beats Sol on Agents' Last Exam at every effort setting, for less.

• Near-Sol image understanding at about half Sol's cost per task.

• A big upgrade on GPT-5.6 Terra, with far fewer factual errors.

• In ChatGPT, your plan allowance stretches further than with Sol or Astra.

Where it falls short:

• Hard coding and desktop computer use, where Sol is the better value per completed task.

• Verbose at max effort, which erases much of the per-token discount.

• Text and images only: no audio, video, fine-tuning or Realtime API.

• More rule workarounds and unauthorised agent actions than GPT-6.1 Sol in OpenAI's own tests.

• No free ChatGPT access, and almost no independent evidence yet.

How to Actually Get GPT-6.1 Terra

As of 30 October 2026, there are four routes in:

• ChatGPT. OpenAI is rolling Terra out to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. Free, Go and Business plans aren't named in the announcement.

• OpenAI API. Use the model ID gpt-6.1-terra with the Responses, Chat Completions or Batch endpoints. Tool calling, including computer use, web search, file search and MCP, requires the Responses API. Rate limits start at 5,000 requests per minute on OpenAI's Build tier.

• Amazon Bedrock. AWS made Terra generally available on 26 October, and says ChatGPT Work and Codex can be set up to run it there. AWS points to its Bedrock docs for pricing.

• GitHub Copilot. Generally available since 26 October for Copilot Pro, Pro+, Max, Business and Enterprise, billed at provider list pricing under usage-based billing. Copilot Free isn't included, and the rollout is gradual.

What's missing: no official Terra announcement from Microsoft Foundry, and Google's 21 October private preview of OpenAI models named Astra, GPT-6.1 Sol and GPT-6 Luna, not Terra. Azure and Google Cloud buyers may be waiting.

Who Should Use GPT-6.1 Terra, and Who Should Skip It

Use it if:

• You're on GPT-5.6 Terra today. It's the easiest upgrade in OpenAI's lineup, though re-test your prompts first.

• You process documents, screenshots or forms at volume, like invoices or technical drawings, where cost per call adds up.

• You run multi-step professional tasks and want near-Astra quality without Astra's bill.

Skip it if:

• Your main job is a coding agent on hard tasks. Start with GPT-6.1 Sol at medium effort.

• You need the best desktop computer use. Sol, Astra and Claude Sonnet 5.5 all scored higher on OSWorld 2.1 in OpenAI's own chart.

• Your work is simple and high-volume, like classification or routing. GPT-6 Luna costs a tenth as much.

• You need audio, video or fine-tuning, or can only buy through Azure or Google Cloud.

Final Verdict on GPT-6.1 Terra

GPT-6.1 Terra is a genuinely good mid-tier model and a no-brainer upgrade from GPT-5.6 Terra. On OpenAI's own numbers, it's the best value in the lineup for professional agent work and image-heavy jobs.

But read the price tag carefully. "Half the price of Sol" is true per token and often false per task. On OpenAI's own charts Sol is the better buy for hard coding and desktop computer use, and Artificial Analysis puts the max-effort saving per task nearer a quarter than a half.

So our call is a split. Make Terra your default for everyday work, documents and screen-driven automation at medium effort. Send hard coding to Sol. Keep approvals switched on for agent work. And check back in a few weeks, once someone other than OpenAI has had a proper go at it.