OpenAI released GPT-6.1 Terra on October 26, 2026, bringing a lower-cost model for visual reasoning, computer use and coding to ChatGPT Work, Codex and the API. Standard API pricing starts at $1 per million input tokens and $5 per million output tokens, half the prices OpenAI reports for GPT-5.6 Terra and GPT-6.1 Sol.
The announcement positions Terra for high-volume professional work: reading invoices, entering information into accounting software, interpreting technical drawings and operating desktop applications. OpenAI claims substantial gains over the previous Terra generation, including better computer-use performance at roughly one-third the cost per task.
The deployment question is whether those savings survive a real workflow’s retries, verification and permissions. Terra also reaches OpenAI’s Critical cybersecurity threshold, making its security controls a central part of the release.
Where GPT-6.1 Terra Is Available
OpenAI lists three access paths:
- ChatGPT Work: Available to Plus, Pro, Enterprise and Edu users.
- Codex: Available to users on those same listed plans.
- OpenAI API: Available to developers under the model identifier
gpt-6.1-terra.
The announcement names ChatGPT Work and Codex specifically; it does not establish availability in the standard ChatGPT model picker. OpenAI also says the same usage allowance covers more work with Terra than with GPT-5.6 Terra, and substantially more than with Sol or Astra, without publishing a fixed task multiplier. Source: OpenAI
For developers, the API model identifier provides a way to select Terra in an integration. It is not a complete automation setup: a computer-use workflow also needs an environment to operate, application access, tools and a way to check whether the requested work was completed.
I’d separate those integration decisions from model selection. A cheaper model can improve an existing workflow’s economics, but it cannot resolve ambiguous instructions or supply missing application permissions.
API Pricing: Token Savings Versus Task Savings
OpenAI publishes these standard API rates:
| Token category | GPT-6.1 Terra price per million tokens |
|---|---|
| Input | $1.00 |
| Cached input | $0.05 |
| Output | $5.00 |
Cached input costs 95% less than uncached input at those rates. That rate applies to input billed as cached; developers should not budget as though every repeated instruction automatically receives it. The announcement does not provide cache eligibility or retention details. Source: OpenAI
A simple calculation shows the token economics. A workload with one million uncached input tokens and 100,000 output tokens would cost:
- Input: $1.00
- Output: $0.50
- Model-token subtotal: $1.50
If 800,000 of those input tokens were instead billed at the cached rate, the subtotal would fall to $0.74: $0.20 for uncached input, $0.04 for cached input and $0.50 for output. These are illustrations calculated from the published rates, not measured workflow costs. They exclude surrounding infrastructure and any separately charged services.
OpenAI’s claim of 50% lower standard pricing than GPT-5.6 Terra and GPT-6.1 Sol implies a $3 subtotal for the same uncached token quantities on those models. That comparison holds token consumption constant.
The more consequential claim changes the denominator: on Agents’ Last Exam, OpenAI says Terra costs about one-third as much per task as GPT-5.6 Terra at medium reasoning effort. If a comparable benchmark workload previously cost $300, that ratio would imply roughly $100. It is a translation of OpenAI’s result, not a forecast for an accounting or browser-automation deployment.
Half-price tokens and one-third-price tasks are different claims. Task cost also depends on how many tokens an agent consumes, its reasoning setting and how much work it needs to finish. The announcement does not provide enough detail to reconstruct every contributor to the reported savings.
My view: cost per successfully completed, verified task is the useful purchasing metric. A low token bill loses its appeal if the workflow repeatedly needs correction or human cleanup.
What the Computer-Use Benchmarks Show
OpenAI’s strongest case for Terra comes from benchmarks involving visual interpretation and actions across applications.
On Agents’ Last Exam, which includes professional tasks such as turning architectural drawings into a 3D model and producing a time-coded film-restoration log, OpenAI reports:
- A 15-percentage-point improvement over GPT-5.6 Terra at medium reasoning effort.
- Approximately one-third the cost per task at that setting.
- Performance on par with GPT-6 Astra at one-tenth its price per token.
- Task completion more than three times faster than GPT-6 Astra and GPT-6.1 Sol.
These are OpenAI-reported results, not independent measurements. The speed claim concerns the evaluated tasks, and the Astra comparison is a token-price comparison, not a demonstrated tenfold reduction in every workflow’s total cost. Source: OpenAI
Two other evaluations support the visual-work positioning. On MMMU-Pro, OpenAI says Terra beats GPT-5.6 Terra at every reasoning setting and nearly matches Sol and Astra. On OSWorld 2.1, a desktop-agent benchmark, Terra at medium effort reportedly exceeds GPT-5.6 Terra’s best score at a fraction of the task cost.
The reasoning setting matters here. A score obtained at maximum effort should not be attached to a deployment configured for lower effort to save time or money.
For a prospective rollout, I’d build the evaluation around complete jobs: read an invoice, extract its fields, enter them into a test accounting system and verify the resulting record. That tests the advertised combination of visual understanding and application interaction. Checking extraction alone would miss errors introduced during entry.
Coding Gains Still Leave Room for Sol and Astra
On Terminal-Bench 4.0, OpenAI reports that GPT-6.1 Terra solves 50.3% of tasks at maximum reasoning effort, compared with GPT-5.6 Terra’s published 21.5%. That is a 28.8-percentage-point gain and more than double the earlier score.
Terminal-Bench tests command-line software work, including tasks such as fixing a webpage whose layout shifts while loading. Its result is relevant to coding-agent selection, but it is not a measurement of success across every repository or development process.
OpenAI keeps a clear distinction between Terra and its higher-end models:
Source: OpenAI“For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.”
That supports a tiered deployment: evaluate Terra for bounded fixes and repeatable coding tasks, with escalation to Sol or Astra when complexity or failed attempts justify the additional cost. This is a deployment recommendation based on the published positioning, not evidence that an automatic routing system has been tested.
OpenAI also reports improved factuality on an internal set of error-inducing prompts. At low reasoning effort, the error rate fell from 53.6% for GPT-5.6 Terra to 16.7%; at maximum effort, Terra reached 4.3%. Those results support testing Terra for document-heavy work, but they do not give a universal error rate for invoices, code or business records.
Frequently Asked Questions
4 questions
1Where Can I Access GPT-6.1 Terra?
OpenAI lists ChatGPT Work and Codex access for Plus, Pro, Enterprise and Edu users. Developers can select gpt-6.1-terra through the API.
