# GPT-6.1 Terra Makes Computer Use Cheap Enough to Build On > GPT-6.1 Terra cuts computer use to $1 input and $5 output per million tokens. Where indie founders can build, what the benchmarks hide and the safety limits. - Canonical: https://zeroraised.com/blog/gpt-6-1-terra-computer-use - Published: 2026-10-27 - Author: ZeroRaised Team (https://zeroraised.com/blog/authors/zeroraised) - Publisher: ZeroRaised (https://zeroraised.com/) - Section: Growth - Tags: artificial-intelligence, indie-apps, developer-tools, saas GPT-6.1 Terra cuts the standard API price for computer use to $1 per million input tokens and $5 per million output tokens, half the standard price of GPT-5.6 Terra and GPT-6.1 Sol ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). That changes the economics of automating back-office work inside apps that never shipped an API. The opportunity is narrow, vertical automation of a job someone currently does by hand, not another general assistant. **TL;DR** - Terra lists at $1 input and $5 output per million tokens, half the price of the previous Terra and Sol standard rates. - The best first products target specific jobs in legacy apps with no API, such as invoice entry. - Vendor benchmarks are a reason to test, not proof that the model handles your workflow. - Terra meets OpenAI's Critical cybersecurity threshold, so build approval gates and limits before launch. **Disclosure:** ZeroRaised publishes this analysis and is a launch site for bootstrapped products. Facts come from the linked sources as checked on 27 Oct 2026; ZeroRaised has no commercial relationship with the companies named. **Computer use** is a model operating a browser or desktop interface by taking actions such as clicks, keystrokes and screenshots, rather than calling a structured API ([OpenAI computer use guide](https://developers.openai.com/api/docs/guides/tools-computer-use)). **Cost per completed task** is the total bill for one finished job, including screen images, retries and failed attempts, not the advertised rate for a single token. ## What did OpenAI actually cut? Terra's standard API price is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). The concrete numbers are $1 input and $5 output per million tokens ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). OpenAI also says Terra is more than three times faster than GPT-6 Astra and GPT-6.1 Sol ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). Speed matters for computer use because every action waits on a screenshot and a model response. A faster loop means fewer timeouts and more completed work. Terra is rolling out in ChatGPT and Codex, but the API is where indie products can use it: it supports computer use and MCP tools through the Responses API ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/), [model docs](https://developers.openai.com/api/docs/models/gpt-6.1-terra)). ## Why does cheaper computer use matter for a small product? Much business work still runs through software with no usable API. Invoices arrive as PDFs, orders sit in supplier portals and reports live in old admin dashboards. A person opens the app, reads the screen and types the data somewhere else. Automation used to require brittle screen scraping or a custom integration. OpenAI describes the pattern as sending a task, executing actions, returning a screenshot and repeating ([OpenAI computer use guide](https://developers.openai.com/api/docs/guides/tools-computer-use)). The example OpenAI gives for Terra is "reading incoming invoices and entering their details into accounting software" ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). At $1 and $5, a workflow can survive a low price per document if it succeeds often enough. At the old rates it needed a higher price or a cleaner task. ## Which jobs should an indie founder automate first? Pick one job with three properties: no API, a clear success test and a human who already does it every week. Good candidates include invoice entry, purchase order matching, claims intake, order entry into supplier portals, expense reconciliation and assembling reports from legacy dashboards. Then charge for the outcome, not the seat. A bookkeeping firm will pay per processed invoice or reconciled batch. Our guide to [pricing an AI product](https://zeroraised.com/blog/how-to-price-an-ai-product) covers the packaging, and [AI agent vs AI workflow](https://zeroraised.com/blog/ai-agent-vs-ai-workflow) helps decide how much autonomy the job really needs. Keep the scope to one app and one job. "Automate accounts payable" is a platform, not a first product. ## What do the benchmarks leave out? OpenAI reports that on Agents' Last Exam, Terra at medium effort scores 15 points above GPT-5.6 Terra at about a third of the cost, and is on par with GPT-6 Astra at a tenth of the price per token ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). At max effort it is more than three times more token-efficient than Claude Opus 5.5, at a quarter of the price per token, or about nine times lower cost per task ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). All of those numbers are vendor-reported, and The Verge frames the computer-use claim as OpenAI's own ([The Verge](https://www.theverge.com/ai-artificial-intelligence/1018854/gpt-6-1-terra-is-very-good-at-computer-use-openai-says)). None of them prove Terra can read the particular invoice layout your customer uses. Build a test set of 30 real tasks, count passes, then measure cost per pass. ## Where is Terra still the wrong tool? First, the hardest coding work: OpenAI says plainly that "for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice" ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). If your product is a coding agent, the price cut does not make Terra the best default. Second, high-stakes autonomy. Terra meets the Critical cybersecurity threshold under OpenAI's Preparedness Framework ([OpenAI](https://openai.com/index/introducing-gpt-6-1-terra/)). It means the model should not be given unattended access to production systems, payment rails or sensitive accounts. There is also no fine-tuning ([model docs](https://developers.openai.com/api/docs/models/gpt-6.1-terra)), so behaviour comes from prompts, tools and the interface you wrap around it. Screenshots count as image input ([OpenAI computer use guide](https://developers.openai.com/api/docs/guides/tools-computer-use)), so a screen-heavy task can cost far more than the text tokens suggest. ## What safety limits do you need before shipping? OpenAI's own guidance is a reasonable baseline: use an isolated browser or VM with an allow list, treat everything on screen as untrusted, confirm consequential actions such as purchases and data transmission, and set step, time or cost limits while verifying the actual outcome ([OpenAI computer use guide](https://developers.openai.com/api/docs/guides/tools-computer-use)). Treat screen content as a prompt injection surface. A web page can contain instructions aimed at the model, and a computer-use agent may follow them. The fix is least privilege: one account, one app and only the permissions the job needs. Add a human approval step for anything that moves money, changes a record or sends data outside the customer's systems. Keep an audit log of every action and screenshot, and review the [vibe-coded app security checklist](https://zeroraised.com/blog/vibe-coded-app-security-checklist) before you take on real accounts. ## What this means for bootstrapped founders 1. Choose one back-office job in an app with no API. 2. Build a 30-task test set with a pass or fail rule before you compare models. 3. Measure cost per completed task, including screenshots, retries and human review time. 4. Start at low or medium reasoning effort and escalate only when the task fails. 5. Put approval gates and spend limits around anything that touches money or customer data. 6. Keep a second model tested, because the price leader changes often. The [Anthropic S-1 analysis](https://zeroraised.com/blog/anthropic-s1-lessons-for-founders) explains why your supplier's incentives are shifting. For the build sequence, see [how to launch an AI app](https://zeroraised.com/blog/how-to-launch-an-ai-app); [how to reduce LLM costs](https://zeroraised.com/blog/how-to-reduce-llm-costs) covers the ongoing bill. ## Where ZeroRaised fits ZeroRaised is a launch and discovery site for products built without institutional venture capital or accelerator equity. If you have a working automation people can try, [submit it here](https://zeroraised.com/submit). The [newsletter](https://zeroraised.com/newsletter) carries new analysis like this, and the [pricing page](https://zeroraised.com/pricing) explains the listing options. ## Sources - [OpenAI: introducing GPT-6.1 Terra](https://openai.com/index/introducing-gpt-6-1-terra/) - [OpenAI: GPT-6.1 Terra model docs](https://developers.openai.com/api/docs/models/gpt-6.1-terra) - [The Verge: GPT-6.1 Terra and computer use](https://www.theverge.com/ai-artificial-intelligence/1018854/gpt-6-1-terra-is-very-good-at-computer-use-openai-says) - [OpenAI: computer use guide](https://developers.openai.com/api/docs/guides/tools-computer-use) ## Frequently asked questions ### How much does GPT-6.1 Terra cost? OpenAI lists standard API pricing at $1 per million input tokens and $5 per million output tokens. That is 50% lower than GPT-5.6 Terra and GPT-6.1 Sol, so the same automation budget buys more work. ### What is computer use AI? Computer use is a model operating a browser or desktop interface by taking actions such as clicks, keystrokes and screenshots, instead of calling a structured API. OpenAI says it is useful for filling out forms, testing user flows and completing tasks in an application. ### Is GPT-6.1 Terra the best model for coding? No. OpenAI says that for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice. Terra is positioned around computer use and agentic work at a lower price, not as the top coding model. ### What are the main risks of building on computer use? The model can act on live interfaces with real consequences, so a wrong click can send money, delete data or leak information. OpenAI recommends an isolated browser or VM, an allow list, treating screen content as untrusted, confirming consequential actions and setting step, time or cost limits. ### Does Terra support fine-tuning? No. The model documentation says fine-tuning is not supported. It offers a 1,050,000 token context window, up to 128,000 output tokens, a 1 June 2026 knowledge cutoff and reasoning effort from none to max, with medium as the default.