GPT-6.1 Terra makes computer-use automation cheaper—and governance more urgent

OpenAI has released GPT‑6.1 Terra, a faster and less expensive model aimed at the computer-based tasks that fill an ordinary workday: interpreting what is on screen, navigating software, handling documents and writing the result.

According to OpenAI’s October 26 announcement, Terra 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. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens.

The important development is not another benchmark leader. It is the combination of useful computer control, visual understanding and a price designed for repeated work. Tasks that were too expensive or slow to automate with a frontier model may now be candidates for production—provided the organisation can control what the model sees and does.

What Terra is designed to do

OpenAI positions Terra for high-volume professional work. Examples include reading invoices and entering details into accounting software, interpreting technical drawings, filling forms, editing video and completing coding tasks from a terminal.

The company reports substantial gains over GPT‑5.6 Terra. On its Agents’ Last Exam benchmark at medium reasoning effort, GPT‑6.1 Terra scores 15 percentage points higher at roughly one-third of the cost. OpenAI says it performs on par with GPT‑6 Astra on that benchmark at one-tenth of Astra’s token price and completes tasks more than three times faster than Astra and GPT‑6.1 Sol.

Those are vendor-reported benchmark results, not a promise about a company’s own workflow. They still indicate the product direction: computer use is moving from a premium demonstration toward a model tier intended for routine throughput.

OpenAI also reports lower factual-error rates on an internal set of difficult prompts. At low reasoning effort, the reported error rate falls from 53.6% for GPT‑5.6 Terra to 16.7% for GPT‑6.1 Terra; at maximum effort, Terra’s reported rate is 4.3%. For software engineering, OpenAI reports a 50.3% score at maximum reasoning on Terminal-Bench 4.0, while advising customers to choose GPT‑6.1 Sol or GPT‑6 Astra for the most complex coding work.

The practical conclusion is model routing, not wholesale replacement. Terra may be the economical choice for frequent visual and computer tasks, while harder reasoning and coding cases escalate to a more capable model or a person.

Lower token prices do not equal lower workflow cost

Computer-use agents consume more than model tokens. They need browser or desktop environments, tool integrations, identity controls, retries, monitoring, human review and incident handling. A cheap model that fails repeatedly or creates long approval queues can cost more per completed task than a more expensive model that succeeds reliably.

Measure the whole unit of work:

• model and tool usage per attempt;
• successful completion without rework;
• time spent waiting for the agent;
• human review and correction time;
• cost of failed or duplicated actions;
• infrastructure for isolated execution and logging.

The relevant metric is cost per correctly completed workflow, not cost per million tokens.

The safety profile belongs in the buying decision

The GPT‑6.1 Terra system-card addendum says OpenAI treats the model as Critical in cybersecurity and High for biological and chemical capability under its Preparedness Framework. Terra uses the same safeguards stack as GPT‑6.1 Sol and GPT‑6 Astra, with probes adapted to the model.

That classification is not a claim that ordinary business use is unsafe. It is evidence that the model is capable enough to require serious containment. The system card reports strong performance on security evaluations, including vulnerability discovery and exploitation tasks, while noting that reliable exploitation remains difficult on some recently disclosed vulnerabilities.

For a business deploying computer use, the concern is broader than deliberate cyber misuse. A capable model can also make an ordinary operational mistake faster and across more systems. If the agent can open files, use authenticated applications and send information, a prompt-injection attack or misunderstood instruction can become an access-control incident.

Where a sensible pilot starts

Choose a bounded, reversible workflow with clear evidence of success. Invoice extraction is a useful example if the first phase stops before payment or ledger posting.

A controlled pilot could let Terra:

1. open invoices from a dedicated queue;
2. extract supplier, amount, tax and purchase-order data;
3. enter a draft into a test or staging system;
4. attach the source document and confidence notes;
5. ask a person to approve the record before it affects finance.

The agent should use a restricted identity, see only the queue needed for the task and have no ability to create suppliers, change bank details or release funds. The reviewer should see the source document and the proposed fields together. Every action should be logged.

That design uses the model’s strengths while keeping the final financial decision with a person.

Controls for production computer use

Separate observation from action. Reading a screen and proposing the next step can be one permission tier. Clicking “send”, changing a record or executing a command should be another.

Constrain the environment. Run browser and desktop tasks in isolated sessions. Restrict network destinations, file access, clipboard use and credential scope. Recreate the environment rather than allowing state to accumulate indefinitely.

Treat displayed content as untrusted. A webpage, document or message can contain instructions intended for the model rather than the employee. Keep untrusted content out of system instructions and prevent it from expanding the agent’s permissions.

Require approval at consequential boundaries. Payment, deletion, external communication, account changes and production deployment deserve explicit review. The approval request should show the exact action, destination, affected data and source evidence.

Set stop conditions. Limit steps, retries, runtime and spend. An agent that cannot complete a task should stop cleanly instead of improvising across adjacent systems.

Evaluate the actual workflow. Build a test set from real examples, edge cases and adversarial inputs. Compare models at the reasoning settings you will use in production. Track accuracy, completion, intervention and cost over time.

What the release means for AI strategy

Terra strengthens the case for a portfolio of models. Organisations no longer need to send every task to the most capable and expensive option. They can route work according to risk, complexity and required latency—but only if they have reliable evaluations and clear escalation rules.

It also raises the value of process knowledge. A lower-cost computer-use model can automate a well-defined workflow more readily than an ambiguous one. Teams that know the inputs, decisions, exceptions and evidence requirements will move faster than teams that begin with “let the agent handle it.”

GPT‑6.1 Terra makes capable computer automation more economically accessible. The opportunity is real: more workflows can now justify a pilot. The corresponding obligation is to design the permissions, reviews and recovery path before scale turns a small mistake into a repeated one.

If you are deciding which workflows and models deserve a production pilot, our AI consulting and integration work starts with the process, the evidence and the human control points.