OpenAI’s GPT-6.1 Terra went live on October 26 with a price that undercuts the company’s own flagship. Developers now pay $1 per million input tokens and $5 per million output tokens, half of what GPT-6.1 Sol charged when it launched on September 29. For companies running AI agents that click through forms and screens, that gap is the story this week.
What OpenAI launched on October 26
OpenAI announced Terra in a blog post on October 26 and pitched it for computer use, visual reasoning, coding and routine office work, according to SiliconANGLE. The company called it “a fast and efficient model that combines detailed visual reasoning” with computer use. Seeking Alpha reported the stated aim as “delivering fast and capable computer use at scale.”
Terra is a middle-tier model. It sits below the flagship GPT-6.1 Sol and the GPT-6 Astra tier, and OpenAI pitches it as the everyday workhorse for simpler jobs such as filling out forms and reading on-screen content. It succeeds GPT-5.6 Terra, so the Terra name carries over while the version number moves up.
The specifications are substantial for a cheaper model. MarkTechPost reported a context window of 1,050,000 tokens, output of up to 128,000 tokens and a knowledge cutoff of June 1, 2026. Terra is closed-weight and runs only on OpenAI-hosted services, so businesses cannot self-host it or run it inside their own data centers.
How much GPT-6.1 Terra costs in the API
The price is the headline. Sources agree on $1 per million input tokens and $5 per million output tokens. They disagree on the yardstick. gHacks Tech News measured the cut against GPT-5.6 Terra and put it at 50 percent lower, while MarkTechPost framed it as halving the API price of GPT-6.1 Sol.
The Sol comparison is the sharper one. According to DIYAI, Sol launched on September 29 at $2 per million input tokens and $10 per million output tokens. Terra cuts both rates in half.
OpenAI also says Terra matches GPT-6 Astra on Agents’ Last Exam while charging one-tenth of Astra’s per-token price. For a team sending large volumes of agent traffic, a per-token gap that size shows up quickly on the invoice. Teams should model their own token counts rather than rely on a headline percentage.
Which users get Terra first
Paying subscribers can pick Terra now. According to Blockchain.News, the model is rolling out to Plus, Pro, Enterprise and Edu subscribers in ChatGPT Work and Codex. Developers can call it through the API using the model string gpt-6.1-terra.
Free accounts are the open question. None of the reports confirmed whether free ChatGPT users get Terra, and none gave a date for when the ChatGPT rollout will finish. If you are on a free plan, check the model picker directly rather than assume access.
What the benchmarks show and who is measuring them
Most of the performance numbers come from OpenAI or from charts OpenAI supplied. The table below uses figures that MarkTechPost reported, which it describes as independent testing on computer-use tasks, along with its Agents’ Last Exam chart.
| Test | Model | Score | Cost per task |
|---|---|---|---|
| Computer-use tasks, medium effort | GPT-6.1 Terra | 57.6% | $0.29 |
| Computer-use tasks, medium effort | GPT-5.6 Terra | 42.6% | $0.84 |
| Agents’ Last Exam, max effort | GPT-6.1 Terra | 60.0% | $1.01 |
| Agents’ Last Exam, same chart | Claude Opus 5.5 | 63.2% | $8.82 |
Two details matter here. The 15-point gain OpenAI cites over GPT-5.6 Terra on Agents’ Last Exam was measured at medium reasoning effort, so it does not line up directly with the 60.0 percent max-effort score in the table. Also, Claude Opus 5.5 scored higher on that same chart, at 63.2 percent for $8.82 per task, nearly nine times Terra’s $1.01 cost for that test.
On Terminal-Bench 4.0, OpenAI’s figures show Terra solving 50.3 percent of tasks at maximum reasoning effort, up from 21.5 percent for GPT-5.6 Terra, as SiliconANGLE reported.
Speed is the claim with the least support so far. OpenAI says Terra finishes certain computer-use tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. BleepingComputer’s headline repeated the “3x faster” framing, but none of the reports cited independent verification of that figure.
Reliability on messy, multi-step work is untested. DIYAI makes the key point: faster and cheaper only helps if the agent avoids repeated attempts and misread screens. A cheap run that retries several times can cost more than its per-task price suggests, and a wrong form entry costs more than the tokens that produced it.
Anthropic is the rival to watch. OpenAI’s comparisons center on Claude Opus 5.5 and its own older models, and Anthropic is building its own agents, including Claude Duet, which watches work apps and asks before it acts.
Why agent security belongs in the upgrade decision
A faster agent that can click through screens and submit forms can also make mistakes faster. USA 24 has reported that major U.S. banks are blocking unvetted AI agents from customer accounts, a sign that institutions are tightening access rather than widening it.
Treat a model switch as a permissions review. Confirm which systems the agent can open, which actions it can take without approval, what data it can read from screens and whether every action is logged. A cheaper model does not change those answers.
How to test Terra before you switch
Vendor benchmarks are a starting point, not a verdict. Teams running agents through the API can take these steps:
- Point a copy of your current agent workflow at gpt-6.1-terra and run it on a sample of real tasks, not demo prompts.
- Record success rate, cost per completed task, retries per task, and any misread screens or wrong form fields.
- Compare the results with your current model on the same tasks, using the same permissions.
- Keep the existing model as a fallback until Terra matches its failure rate on your work.
- Review the logs for every action Terra takes on a screen, and remove any permission the agent does not need.
What to watch next
Three things will settle the story. First, independent evaluations of the three-times speed claim, which have not yet appeared in the reporting. Second, whether OpenAI extends Terra to free ChatGPT accounts and how quickly the rollout reaches paying users. Third, whether rivals answer with price cuts or with agent results of their own.
If Terra keeps its cost advantage on multi-step tasks outside OpenAI’s charts, the per-task math will push many agent budgets toward it. If retries add up, the lower price will matter less than it looks today.
