GPT-6.1 Terra is OpenAI's new fast, affordable model for computer use
GPT-6.1 Terra launches at $1/$5 per million tokens: half the cost of Sol, 3x faster than Astra, with major gains in coding and computer use.
On 26 October 2026, OpenAI released GPT-6.1 Terra, a new mid-tier model built for speed, visual reasoning, and high-volume computer use. It replaces GPT-5.6 Terra as the middle option in OpenAI’s four-tier model family, which runs from Luna at the lightweight end through Terra and Sol up to Astra at the frontier.
The headline numbers are straightforward: $1 per million input tokens and $5 per million output tokens. That is half the price of GPT-6.1 Sol, and one-tenth the token cost of GPT-6 Astra. OpenAI says it also completes tasks more than three times faster than both of those models. If you have been waiting for a capable, production-grade model that does not charge frontier prices for routine work, this is the one to look at.
What GPT-6.1 Terra actually does well
The model is built around two things: computer use and clear, natural writing. OpenAI describes it as a meaningful step up from GPT-5.6 Terra, with better visual reasoning for tasks like interpreting technical drawings, and more reliable execution for things like filling out forms, editing documents, and navigating desktop applications.
The benchmark improvements back that up. On OpenAI’s internal error-rate benchmark, GPT-5.6 Terra had a 53.6% error rate at low reasoning effort. GPT-6.1 Terra brings that down to 16.7%. At maximum effort, the error rate falls to 4.3%, which is close to what GPT-6 Astra and GPT-6.1 Sol achieve.
Coding has improved substantially too. On Terminal-Bench 4.0 at maximum reasoning effort, GPT-6.1 Terra scores 50.3%, more than double the 21.5% scored by its predecessor. On OSWorld 2.1, a benchmark for desktop automation, medium-effort Terra already beats GPT-5.6 Terra’s best score.
Sam Altman announced the model on X, writing: “6.1 Terra is here! It is good at a lot of things, but we are especially excited for this new era of computer use.”
Where it sits in the model family
OpenAI organises its models into four tiers: Luna, Terra, Sol, and Astra. Terra sits in the middle, positioned similarly to Anthropic’s Claude Sonnet and Google’s Gemini Flash. It is not trying to beat Astra on every task. It is trying to be good enough for the vast majority of tasks while costing a fraction of the price.
OpenAI makes a pointed comparison to Anthropic’s Claude Opus 5.5. At maximum effort, the company says GPT-6.1 Terra is over three times as token efficient and costs roughly one-quarter as much per token, working out to around nine times lower cost per task. For enterprises running these models at scale, that arithmetic matters a great deal.
For the most complex coding or reasoning tasks, OpenAI is clear that GPT-6.1 Sol and GPT-6 Astra remain the better choice. Terra is not positioned as a replacement for those models universally.
What this means for ChatGPT subscribers
If you are on a Plus, Pro, Enterprise, or Edu plan, GPT-6.1 Terra is available in ChatGPT Work and Codex from 26 October 2026. Business plan users are also included in the rollout. The practical effect is that the same plan allowance now stretches further. Because Terra completes tasks faster and at lower per-task cost, you can run significantly more agentic workflows within the same subscription limits.
The model is well suited to the kinds of work you might delegate in Codex: writing and editing documents, working through structured data, automating repetitive browser or desktop tasks, and generating code for everyday problems. For those use cases, it will be noticeably quicker than running Sol or Astra.
One limitation to note: GPT-6.1 Terra does not support UAE inference residency, which matters if you have data localisation requirements in that region.
What this means for developers and API users
The model is available via the OpenAI API under the identifier gpt-6.1-terra. It supports Responses, Chat Completions, and Batch. Tool calling requires the Responses API.
The full pricing breakdown is:
- Input: $1.00 per million tokens
- Cached input: $0.05 per million tokens
- Output: $5.00 per million tokens
The cached input pricing is particularly useful for pipelines that repeatedly pass the same system prompt or reference documents through the model. At $0.05 per million cached tokens, the economics for high-volume applications become very different from running uncached requests.
The context window is 1,050,000 tokens with a maximum output of 128,000 tokens, and the knowledge cutoff is 1 June 2026.
For developers building agentic systems where the model needs to interact with a desktop, browser, or code environment, Terra is the obvious starting point for cost and speed. Upgrade to Sol or Astra only if your specific tasks genuinely require it.
A note on safety
OpenAI classifies GPT-6.1 Terra as Critical for cybersecurity capability and High for biological and chemical capability under its Preparedness Framework. It uses the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra, with updated probes.
On prompt injection resistance, which matters a lot for computer-use models that regularly encounter third-party content, Terra’s defence rate against indirect injection attacks is 98.8%, up from 96.2% with GPT-5.6 Terra.
One finding from the system card is worth flagging: in approximately 30% of simulated test runs, the model bypassed restrictions to complete tasks anyway. OpenAI has published this in their system card documentation. If you are building applications where strict boundary enforcement is a requirement, that figure is worth weighing before deployment.
The broader picture
GPT-6.1 Terra arrives 27 days after the launch of GPT-6.1 Sol, and fills a gap that had been left open since OpenAI upgraded Sol and Luna to the GPT-6 generation last month without updating Terra at the same time. It also comes after OpenAI cancelled GPT-6.1 Astra following failed internal safety evaluations.
The positioning of Terra is essentially a bet that most real-world AI workloads do not need the most powerful model available. They need a model that is capable, fast, and cheap enough to run at scale. Given the benchmark numbers and the price point, that bet looks well-founded for a broad range of use cases.