OpenAI's GPT-6.1 Terra Halves the Token Price. The Computer-Use Bill Is What to Watch.

Point-in-time, 28 October 2026 — two days after OpenAI launched GPT-6.1 Terra.

The numbers, plainly

• OpenAI launched GPT-6.1 Terra on 26 October 2026, describing it as a fast, efficient model for detailed visual reasoning, precise computer use, and clearer, more natural writing.
• Its standard API prices are $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. OpenAI states those prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol.
• The model is available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and to developers through the API as gpt-6.1-terra.
• The model documentation lists a 1,050,000-token context window, a 128,000-token maximum output, and a knowledge cutoff of 1 June 2026.
• OpenAI's examples of the work Terra is built for are deliberately ordinary: drafting proposals, interpreting technical drawings, filling out forms, editing videos, and, at high volume, reading incoming invoices and entering their details into accounting software.

Why price per task, not price per token, is the right unit

A token price is easy to publish and easy to compare. It is not what a company pays. A company pays for a finished task: a form filled correctly, a drawing read and turned into a model, a proposal drafted without a factual error that has to be caught later.

That is why Terra's launch matters in the same lane as the Anthropic filing two days earlier. Anthropic's half-year showed compute as about 55% of revenue. Terra shows the other side of that ledger: the supplier now cutting the unit price of the work that creates that compute demand, and aiming it at the highest-volume, most routine computer work in an office day.

If Terra does what OpenAI claims — the same computer-use quality at a third of the cost per task of its predecessor, and on par with the flagship Astra at a tenth of the price per token on one benchmark — then the bottleneck for agent work moves. It stops being "can a model do this at all?" and becomes "how many of these tasks is it worth running, and how carefully do you check them?"

That shift is also the site's physical-footprint question, stated plainly: cheaper inference does not automatically mean less inference. If a halved token price leads to twice as many agent tasks, the data-center load behind those tasks grows, not shrinks. The price cut is real. The total bill, in tokens and in megawatts, is a separate number, and nobody has published it yet.

The fine print that changes the headline price

Three details from OpenAI's own model documentation qualify the $1 / $0.05 / $5 headline, and any draft that omits them overstates the cut:

• Long prompts cost more. Prompts with more than 272,000 input tokens are priced at 2x input and cache rates and 1.5x output for the full request — not just for the part above the threshold.
• Batch and Flex are cheaper again. Batch and Flex prices are 50% lower than Standard, which matters for high-volume, non-urgent work like invoice processing.
• Fast mode is dearer. Fast mode prices are 2x Standard. A team that needs speed first will not pay the headline price.

Cache is the other lever. Cached input at $0.05 is 5% of the uncached rate, so an agent that re-reads the same documents, drawings or system prompts repeatedly will see a much larger saving than an agent that mostly reads new input. The launch's own example — invoices into accounting software — is exactly that kind of repetitive, cache-friendly work.

What OpenAI claims, and what it does not prove

All of the following are OpenAI's own reported results, not independent verification. At the time of writing, independent tests of Terra had not yet been published, as launch coverage itself noted. Treat them as the vendor's case to be tested, not as settled measurements:

• On Agents' Last Exam, a benchmark of professional tasks including turning architectural drawings into a 3D model, Terra at medium reasoning effort scored 15 percentage points higher than GPT-5.6 Terra at about one-third the cost, and performed on par with GPT-6 Astra at only one-tenth its price per token.
• Terra completed tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol, on OpenAI's measures.
• On MMMU-Pro, a visual-reasoning benchmark, Terra outperformed GPT-5.6 Terra at every reasoning setting and nearly matched Astra and Sol at much lower cost.
• On OSWorld 2.1, which tests agents using desktop apps for multi-step tasks, Terra surpassed GPT-5.6 Terra's best score at a fraction of the cost per task.
• On OpenAI's internal benchmark of error-inducing prompts, the error rate at low reasoning effort fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, and at maximum effort fell to 4.3%.
• On Terminal-Bench 4.0, a coding benchmark, Terra solved 50.3% of tasks at maximum reasoning effort, more than double GPT-5.6 Terra's published 21.5%.

OpenAI also draws its own boundary: for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice. Terra is positioned as the middle model — not the frontier, and not the cheapest — built for volume work that does not need the frontier.

What this does not settle

• Cost per finished task in production. Token prices and benchmark scores do not tell us how many tokens a real invoice, form or drawing actually takes, how often a task needs to be rerun, or how much human checking it still needs. No independent cost-per-task figure was available at launch.
• Whether speed claims survive real desktops. OSWorld and Agents' Last Exam are controlled tests. A real office desktop — with its pop-ups, permissions, slow applications and changing interfaces — is a harder test, and OpenAI has not published that result.
• Whether the factual-error improvement holds outside OpenAI's own prompts. The 53.6% to 16.7% fall is on an internal benchmark designed to induce errors. It is a promising direction, not a general error rate for customer work.
• What cheaper computer use does to total demand. If Terra makes agent work cheap enough to run everywhere, inference volume could rise faster than the price falls. That is the question this site will keep tracing: price per token down, total tokens up, and power demand following the second number, not the first.
• Safety at scale. OpenAI states Terra meets the Critical threshold in cybersecurity under its Preparedness Framework. What that threshold means in practice for agents that can click, type and move between applications at scale is a deployment question, not a launch-day answer.

What I am watching next

1. Independent benchmark and cost-per-task results for Terra, when they appear — especially on real desktop tasks, not only on OSWorld and Agents' Last Exam.
2. Whether Batch and Flex pricing, rather than Standard pricing, becomes the real price for high-volume agent work, and what that does to the headline 50% claim.
3. Any disclosure of Terra's inference volume or energy footprint, in the same way Anthropic's filing disclosed its compute bill. Without that, the physical cost of cheaper computer use stays invisible.
4. Whether Sol and Astra usage actually falls for routine work, or whether Terra simply adds a new layer of agent work on top of the existing tiers.

Sources

1. OpenAI, 26 October 2026 — Introducing GPT-6.1 Terra.
2. OpenAI API documentation — GPT-6.1 Terra Model: context, output, cutoff and pricing fine print.
3. Skimwire, 26 October 2026 — OpenAI's GPT-6.1 Terra targets everyday computer work at half the API price of its predecessor.

Risk Assessment

This post claims only what OpenAI's launch announcement and model documentation establish: the price, availability, context and output limits, the stated use cases, and OpenAI's own benchmark results, clearly labelled as vendor-reported. It does not claim an independent cost per task, a production error rate, or a total power saving from the price cut — those are precisely the figures that independent testing and deployment data will have to supply.