OpenAI did not launch its newest model with a bigger number on a leaderboard. It launched it with a smaller number on the price list. GPT-6.1 Terra arrived on October 26 and 27 at $1 per million input tokens and $5 per million output tokens, which is half what both GPT-5.6 Terra and GPT-6.1 Sol cost. It is available in ChatGPT for Plus, Pro, Enterprise and Edu subscribers, and the company paired it with a 50% price cut for agentic computer use.
The pitch is simple: make AI agents cheap enough to run all day. Whether Terra is good enough to deserve that job is the more interesting question, and the answer depends on whose benchmarks you trust.
The short version
- API pricing is $1 input, $0.05 cached input and $5 output per million tokens, half the rate of GPT-5.6 Terra and GPT-6.1 Sol
- OpenAI says Terra at max effort scores 60.0% on Agents’ Last Exam at about $1.01 per task, against 50.7% at $3.58 for its predecessor
- Computer-use scores jump on OpenAI’s own charts: 74.2% on OSWorld 2.1, up from 57.1%
- Independent indexes put Terra near 48, while Claude Opus 5.5 sits in the mid to high 50s
- Most of the headline claims are from OpenAI, and the computer-use results have not been independently verified
What the new price list actually looks like
The standard rate is only part of the story. Batch and Flex processing run another 50% below standard, which pushes the cost of non-urgent agent work very low. Fast mode costs double the standard rate, and prompts over 272K input tokens are billed at a higher tier. Cached input at five cents per million is the quiet lever here: agents resend the same instructions and tool definitions constantly, so a cheap cache can matter more than the headline rate.
| Mode | Input (per 1M) | Output (per 1M) |
|---|---|---|
| Standard | $1.00 | $5.00 |
| Cached input | $0.05 | n/a |
| Batch and Flex | 50% below standard | 50% below standard |
| Fast mode | 2x standard | 2x standard |
The benchmark claims, in one picture
OpenAI’s case rests on cost per finished task rather than raw intelligence. On Agents’ Last Exam, a test built around professional software work, the company says Terra at max effort completes tasks for about $1.01 each, versus $3.58 for GPT-5.6 Terra, while also scoring higher. At medium effort it reports a 15 point gain at roughly a third of the cost. It also says Terra matches GPT-6 Astra on that test at one tenth the per-token price.
Those are big jumps. The Terminal-Bench 4.0 score more than doubles, from 21.5% to 50.3%, and OpenAI says the internal error rate at low effort fell from 53.6% to 16.7%. If those numbers hold up in real workloads, Terra is a serious tool for coding agents and browser automation. They are still OpenAI’s numbers, run on OpenAI’s chosen settings.
Where independent data pushes back
Third-party evaluators tell a more modest story. Artificial Analysis gives GPT-6.1 Terra at max effort an Intelligence Index score of 48, while Claude Opus 5.5 scores in the mid to high 50s depending on the index version. The version matters: one of the 48 readings is tied to an older release of the index, so exact comparisons should be treated with care. Even OpenAI’s own chart for Agents’ Last Exam shows Opus 5.5 ahead, at 63.2% against Terra’s 60.0%, though OpenAI argues Terra is about nine times cheaper per task.
| Measure | GPT-6.1 Terra (max) | Claude Opus 5.5 |
|---|---|---|
| Intelligence Index (Artificial Analysis) | 48 | 56 to 58 |
| Agents’ Last Exam (OpenAI chart) | 60.0% | 63.2% |
| Output speed | 90.9 tokens/sec | 83.6 tokens/sec |
| Blended cost per 1M tokens | $0.73 | $2.94 |
This is the real shape of the trade. Terra is faster and roughly four times cheaper on a blended token basis, while trailing the top Claude model on measured capability. For high-volume agent work where a task can be retried, that is a good deal. For hard problems where one wrong answer is expensive, the cheaper model may cost more in the end.
Where Terra fits in OpenAI’s lineup
The timing makes sense in context. OpenAI recently scrapped GPT-6.1 Astra and shipped a cheaper model instead, and Terra continues that direction: stop chasing the top slot and win on economics. The top of the capability chart is crowded anyway, with Google’s Gemini 4 Argon leading many benchmarks but hard to access. Cost pressure is also coming from the east, where DeepSeek is now within 3% of the best US model. A cheap, fast, good-enough model is exactly what that environment rewards.
Before you switch your agents
- Test on your own tasks. Vendor benchmarks reward the setups the vendor picked, and your workload will differ
- Measure cost per completed task, not per token. A cheaper model that needs more retries can erase the savings
- Watch the long-prompt tier. Anything over 272K input tokens is billed higher, which matters for big-codebase agents
- Wait for independent computer-use results. The OSWorld and Terminal-Bench jumps have not yet been confirmed outside OpenAI
Terra is not the smartest model on the market, and OpenAI does not really claim it is. It is an argument that the next phase of AI is about who can run agents at a price people will actually pay. On that argument, a $1 input rate is a strong opening move. The proof will come from teams running it for a month, not from a launch-day chart.
Sources and further reading
- gHacks: OpenAI launches GPT-6.1 Terra with API pricing 50% lower
- Technology.org: GPT-6.1 Terra, $1/$5 per million tokens
- Artificial Analysis: GPT-6.1 Terra vs Claude Opus 5.5
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
