GPT-6.1 Terra cuts the price of AI agents. Reliability is the harder test.

Lower inference costs change which tasks make sense

Reading invoices, transferring details into accounting software and completing multi-step actions on a desktop do not always need a frontier-priced reasoning model. Yet model fees are only one part of an agent’s cost: retries, tool calls and human corrections often decide whether automation pays off. GPT-6.1 Terra targets exactly this operational trade-off.

Announced on October 26, the model combines visual understanding, computer use and writing. OpenAI describes applications ranging from technical drawings to form entry and editing tasks. That is a model capability, not a ready-to-use enterprise integration. Actual deployments still need tool connections, scoped credentials, approval rules and an accountable owner.

Reading the benchmarks without turning them into promises

The stated standard API rates are $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens. OpenAI says standard token pricing is 50% below GPT-5.6 Terra and GPT-6.1 Sol. The complete bill for one finished task also includes repeated steps, external services, validation and potential recovery work.

On Agents’ Last Exam at medium reasoning, OpenAI reports a 15-percentage-point improvement over GPT-5.6 Terra at roughly one-third of the cost per task. On Terminal-Bench 4.0 at maximum reasoning effort, it reports 50.3% of tasks solved versus 21.5% for its predecessor. These are vendor-published evaluations under defined settings, not independently verified success rates for every website, finance workflow or enterprise desktop.

The company also reports that the factual-error rate on its internal difficult-prompt benchmark fell from 53.6% to 16.7% at low reasoning effort. That is a result for a specific internal test, not a universal hallucination rate. OpenAI itself still points developers towards stronger models such as GPT-6.1 Sol or GPT-6 Astra for the most demanding coding work.

Where it is available, and where it is not automatically included

Terra is accessible through the OpenAI API as gpt-6.1-terra and is rolling out through ChatGPT Work and Codex. The ChatGPT release notes list Plus, Pro, Business, Enterprise and Edu plans, while making access contingent on rollout, region and workspace settings. That does not mean every account sees the model immediately, or that it is now available in the ordinary Chat tab. API use is billed separately from ChatGPT subscriptions.

Computer use is not equivalent to permission to perform any computer action. The selected application or organization still has to define which data the agent may read, which commands it may run, when to seek approval and how its activity can be audited.

Editorial view: price per successful task beats price per token

A faster, cheaper model can make high-volume routine work attractive. But procurement teams should evaluate the cost of one correctly completed and reviewable job, not merely the token line on a price sheet. An agent that retries after interface changes or needs frequent manual cleanup can erase its nominal model-price advantage.

A useful pilot would choose a narrow reversible task—for instance handling a sample invoice without issuing payment. Measure completion rate, duration, token and tool cost, exception handling, trace quality and human intervention. These figures would establish whether OpenAI’s benchmark advantages translate to the organization’s actual workflow. AgentenCode has not performed an independent Terra benchmark.

Consequential actions require a separate approval layer. An agent may prepare a payment draft or update proposal, while authorization to move money or modify live master data should still follow the organization’s process. Our evidence-based Agent Check helps turn those requirements into concrete questions.

Primary sources and verification scope

• OpenAI: GPT-6.1 Terra product announcement (26 Oct) ↗
• OpenAI: GPT-6.1 Terra – deutsche Primärquelle ↗
• OpenAI: ChatGPT release notes, 26 Oct 2026 ↗

Claims are drawn from dated first-party publications; the editorial assessment is not an independent software test or certification.