On 26 October 2026, OpenAI released GPT-6.1 Terra, the successor to GPT-5.6 Terra. It is the company's model for high-volume agent work: running many steps quickly, using tools, and operating software through its screen, known as computer use. It costs half as much per token as GPT-6.1 Sol, and on OpenAI's own agent benchmark it scores slightly higher. It is already available on Databricks and in GitHub Copilot. Here is what it is, how it compares and where it fits.
What Terra is for
OpenAI now has three tiers. GPT-6 Astra is the flagship for the hardest problems, GPT-6.1 Sol is the balanced workhorse for reasoning and coding, and Terra is tuned for speed and cost on agent loops: filling in forms, moving data between apps, triaging tickets, or running a test suite and fixing what fails. OpenAI says it is more than three times faster than Astra and Sol.
Benchmarks
These are OpenAI's launch figures at maximum effort, so treat them as the vendor's view. Independent tests will follow.
OpenAI's published results at maximum effort, October 2026. Cost per task: Terra $1.01, Sol $2.00, Astra $7.24, Claude Opus 5.5 $8.82, GPT-5.6 Terra $3.58.
| Benchmark | GPT-6.1 Terra | Comparison |
|---|---|---|
| Agents' Last Exam | 60.0% at $1.01 a task | Sol 58.6% at $2.00; Claude Opus 5.5 63.2% at $8.82 |
| OSWorld 2.1 (desktop computer use) | 74.2% at $1.19 a task | GPT-5.6 Terra 57.1%; Claude Sonnet 5.5 83.9% at $5.74 |
| MMMU-Pro (visual reasoning) | 85.5% | Sol 86.3%; Astra 86.7% |
| Terminal-Bench 4.0 | 50.3% | GPT-5.6 Terra 21.5% |
| Hallucination rate, OpenAI test | 16.7% at low effort, 4.3% at max | GPT-5.6 Terra 53.6% at low effort |
Artificial Analysis, which tests models independently, measured an average cost of $0.40 per task on its Intelligence Index at maximum effort, against $1.40 for GPT-5.6 Terra, and a Coding Agent Index score six points below Sol.
Price and availability
- API
- gpt-6.1-terra: $1 input, $5 output, $0.05 cached input per million tokens. Batch and Flex are half price; fast mode costs double.
- Long prompts
- Requests with more than 272K input tokens are billed at twice the input rate and 1.5 times the output rate for the whole request.
- ChatGPT and Codex
- In ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users; not in regular ChatGPT chat, Free or Go at launch.
- Other platforms
- GitHub Copilot (generally available), and Databricks Unity Gateway from 27 October.
- Effort levels
- None, low, medium (the default), high, xhigh and max.
- Not supported
- Fine-tuning, and image or audio output.
The safety findings to read first
OpenAI's system card rates Terra as critical capability in cybersecurity and high in biology and chemistry, and it uses the same safeguards as Astra and Sol. Two results matter for anyone building agents:
Should you use it?
- 1
Use Terra for volume
Routing, extraction, form filling and repetitive tool calls, where speed and price matter more than the last few points of quality.
- 2
Keep Sol or Claude for hard reasoning
Complex analysis, research and difficult code still favour Sol, Opus 5.5 or Sonnet 5.5.
- 3
Test computer use on your own apps
74% on a benchmark still means failures. Start with read-only tasks and human approval for anything that writes.
- 4
Watch long prompts
Keep inputs under 272K tokens where you can, or the whole request costs more.
- 5
Pin behaviour with evaluations
OpenAI lists a moving model alias, so rerun a small test set when the model updates.
Check your understanding
1/3How does GPT-6.1 Terra's price compare with GPT-6.1 Sol?
Terra is $1 input and $5 output per million tokens; Sol is $2 and $10.
2/3Which model led OSWorld 2.1 in OpenAI's comparison?
Sonnet 5.5 scored 83.9% against Terra's 74.2%, at a higher cost per task.
3/3What did the system card show about peer-agent requests?
11 of 80 for Terra against 1 of 80 for Sol; enforce permissions outside the model.
You got 0 out of 3.
Compare it with everything else in the best AI models in October 2026, and read GPT-6.1 Sol and OpenAI DevDay 2026 and long-running AI agents.
Sources: InfoWorld, "OpenAI bets on AI economics with lower-cost GPT-6.1 Terra" (27 October 2026); OpenAI's GPT-6.1 Terra announcement and system card as summarised by Learnetto and TrueFoundry (26 to 27 October 2026); Artificial Analysis; Databricks release notes (27 October 2026). Benchmark figures are OpenAI's unless stated.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 28 October 2026. About me