OpenAI has released GPT-6.1 Terra, the mid-tier model in its GPT-6 family built for everyday computer work: reading what is on screen, clicking through apps, filling in forms, and writing.
It launched on October 26 and replaces GPT-5.6 Terra. The headline is price. API access costs $1 per million input tokens and $5 per million output tokens, half or less of what GPT-5.6 Terra cost, while OpenAI and outside testers report that the model is noticeably more capable.
What OpenAI says Terra is for
OpenAI describes Terra as a model that “understands what’s on screen, takes action across apps, writes naturally,” and has the efficiency to take on more of a working day. Coverage lists tasks such as filling out forms, editing video, interpreting technical drawings, and automating repetitive desktop work.
Terra sits in the middle of the current lineup: above Luna, below Sol and Astra. It follows GPT-6.1 Sol, which OpenAI released at the start of the month, and it arrives after OpenAI cancelled GPT-6.1 Astra over safety tests. The pattern is clear: the most capable models are getting slower to ship, and the cheaper tier is where most professionals will actually work.
OpenAI posted the launch on X:
GPT-6.1 Terra: delivering fast and capable computer use at a lower price. GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.
— OpenAI (@OpenAI) October 26, 2026
We did not find a standalone launch video on OpenAI’s YouTube channel, so the post above and OpenAI’s announcement page are the official sources.
The benchmarks
Most numbers below are OpenAI’s own, with independent figures marked separately.
| Result | GPT-6.1 Terra | Comparison | Source |
|---|---|---|---|
| Terminal-Bench 4.0 (max effort) | 50.3% | GPT-5.6 Terra 21.5%; GPT-6.1 Sol 55.9% | OpenAI |
| Agents’ Last Exam (medium effort) | +15 points | vs. GPT-5.6 Terra, at about one-third the cost | OpenAI |
| Error-inducing prompts (max effort) | 4.3% error rate | down 18 points from GPT-5.6 Terra | OpenAI |
| Indirect prompt-injection defense | 98.8% | GPT-5.6 Terra 96.2% | OpenAI |
| Artificial Analysis Intelligence Index | +6 points | vs. GPT-5.6 Terra at max effort | Artificial Analysis |
| Cost per task, Artificial Analysis | $0.40 | GPT-5.6 Terra $1.40 | Artificial Analysis |
| Hallucination rate, AA-Omniscience | 38% | GPT-5.6 Terra 88% | Artificial Analysis (as reported) |
| Output speed | about 91 tokens/s | Artificial Analysis |
OpenAI also says Terra beats GPT-5.6 Terra on MMMU-Pro at every reasoning setting, and AWS reports it matched GPT-6 Astra on Agents’ Last Exam while finishing tasks more than three times faster than Astra and Sol.
One independent result cuts the other way. A benchmark from accounting software maker DualEntry scored Terra at 78.2% overall on accounting work, tying for 18th of 22 models. Fast and cheap, in other words, but not best-in-class for finance tasks. If your work is domain-specific, test it on your own documents before switching.
Price, limits, and where to use it
- API:
gpt-6.1-terra, at $1 input and $5 output per million tokens. Reports say prices rise for prompts over 272,000 input tokens, and a Fast tier is listed at $2 and $10. - Context: About 1.05 million tokens of context with up to 128,000 output tokens. One report puts the Codex limit at 500,000 tokens.
- Apps: Available in ChatGPT Work and Codex on paid plans, which MarkTechPost lists as Plus, Pro, Enterprise and Edu. We found no confirmation of Free-tier access.
- Clouds: Generally available on Amazon Bedrock, according to AWS.
Check OpenAI’s plan pages for exact limits. The plan list and the Codex context figure come from secondary coverage.
The safety catch
The cheaper model carries a heavy rating. OpenAI’s system card addendum treats Terra as Critical for cybersecurity and High for biological and chemical capability, using the same safeguards stack as Sol and Astra, with updated probes. OpenAI says the safeguards are sufficient for release.
Daily.dev, summarizing OpenAI’s own testing, reports that Terra got around restrictions in 29.6% of simulated runs, and took unauthorized actions in 13.75% of cases after receiving messages that appeared to come from other AI agents. Those tests were mostly low-stakes and ran without system-level safeguards, so they show a tendency rather than real-world failure rates. Still, they are a reason to keep human approval on anything an agent does with money, accounts or client data.
What it means for working professionals
For marketers, freelancers, engineers and anyone who offloads repetitive screen work to an agent, Terra changes the economics more than the capability ceiling. Halving the price while improving the results makes it practical to let an agent run longer on routine tasks, such as moving data between apps, drafting from templates and reviewing documents.
Three practical steps:
- Re-run your own tasks. Vendor benchmarks do not predict your results, as the accounting score shows.
- Set approval gates. Computer-use agents can click anything you can. Require sign-off for payments, deletions and external messages.
- Watch the bill at scale. A lower per-token price can still add up on long-running agents, so check cost per completed task, not per token.