GPT-6.1 Terra Costs $1 Per Million Input Tokens: The Cheap Computer Use Model, and the 272,000 Token Trap

OpenAI released GPT-6.1 Terra on 26 October 2026, and the number that matters is on the price card, not in the launch video. Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. OpenAI says that is 50 percent lower than GPT-5.6 Terra and GPT-6.1 Sol, and about one tenth the token price of GPT-6 Astra. For builders, that changes which jobs are worth automating at all. A job that was too expensive to run on a flagship model every hour can become affordable when the input price falls to $1 per million tokens, provided the model can actually finish the job.

Terra is built for computer use. That means work that combines seeing what is on screen with taking action across apps: drafting a proposal from a set of documents, reading a technical drawing, filling out a form, entering invoice data into accounting software, or editing a video by following on screen steps. It is not positioned as the model for the hardest coding or the deepest research. OpenAI itself says that for in depth research or complex coding, GPT-6.1 Sol or GPT-6 Astra remain the better choices. That honesty is useful. Terra is the workhorse tier, not the flagship tier, and judging it against the flagship on the hardest tasks misses the point.

Where can you actually use Terra today?

Terra launched on several surfaces on the same day, which is unusual and helpful. In ChatGPT Work and Codex, it is available to Plus, Pro, Enterprise, and Edu users. Developers can call it through the OpenAI API under the model name gpt-6.1-terra. Amazon Web Services made it generally available on Amazon Bedrock on 26 October, with the option to point ChatGPT Work and Codex at the Bedrock hosted model. GitHub made it generally available in GitHub Copilot the same day for Pro, Pro+, Max, Business, and Enterprise plans, with a gradual rollout. Microsoft's Foundry documentation also lists gpt-6.1-terra, version 2026-10-26.

That spread matters for a practical reason. If your team already runs on AWS, you can test Terra under your existing AWS access and audit setup instead of opening a new vendor relationship. If your developers already live in Copilot, they can try Terra in the client they already use, subject to their admin's model policy and the gradual rollout. If you are an individual Plus user, the fastest test is in ChatGPT Work or Codex, where OpenAI says your existing usage allowance goes further with Terra than with GPT-5.6 Terra, and far further than with Sol or Astra. Availability can still vary by plan, region, workspace settings, and rollout progress, so if Terra has not appeared in your picker yet, that is a rollout delay, not proof that your plan is excluded.

What does $1 per million tokens actually buy?

Token math is where most pricing posts go wrong, because they quote the rate and stop. A token is a chunk of text, roughly three quarters of a word in English on average, but code, numbers, and long documents do not follow that average neatly. The table below uses OpenAI's standard rates and simple, stated assumptions, so you can replace them with your own job sizes. These are arithmetic examples on the published rates, not a promise about your bill, because cache use, tools, and the long prompt tier change the result.

Read the cache row twice, because it is the biggest saving in the table. Cached input at $0.05 per million tokens is 95 percent cheaper than fresh input at $1 per million. If your agent rereads the same documents, the same system instructions, or the same reference material across many tasks, cache is what turns Terra from cheap into very cheap. If every task is brand new material with no reuse, you pay the full input rate every time, and the output side dominates the bill faster than people expect, because output at $5 per million is five times the input rate.

For comparison, OpenAI's framing puts Sol at double these standard rates, so the same 1,000 short tasks would cost about $20 on Sol instead of $10 on Terra, before cache and before any Batch or Flex discount. Batch and Flex processing cost 50 percent less than Standard on Terra, according to the pricing notes, which would cut the Standard example further for work that can wait. Fast mode costs twice the Standard rate, so do not turn it on by default. Those modifiers are why two teams can run the same model and get very different bills.

What is the 272,000 token trap?

The context window is 1,050,000 input tokens, with a maximum output of 128,000 tokens. That sounds like you can throw a whole archive at the model and pay the headline rate. You cannot, not at the headline rate. Prompts with more than 272,000 input tokens attract double input and cache rates and 1.5 times the output rate for the full request. In plain terms, once you cross that line, the whole request reprices, not just the part above the line. A 300,000 token prompt does not pay $1 per million on the first 272,000 and a higher rate on the rest. It pays the higher rate on everything.

That changes how you should design agent work. Do not dump a full document store into one prompt because the context window allows it. Split the work: retrieve the relevant pages first, then send Terra only what the task needs, and keep repeated instructions in cache. A task that fits under 272,000 tokens stays at the cheap rate. A task that sprawls over it pays double input and 1.5 times output, which can wipe out the saving versus Sol in one step. There are also separate charges for cache writes and some tools, with 30 minute cache writes listed at $1.25 per million tokens in the pricing notes, so cache is not free to create, even though cached reads are cheap.

The practical rule is simple. Measure your prompt size before you scale a workflow. If your typical prompt is under 272,000 tokens, Terra's headline price is your price, subject to cache and tools. If it is over, either trim the prompt or budget for the higher tier. Never estimate a bulk agent budget from the $1 headline alone without checking prompt sizes first.

Are OpenAI's performance claims worth trusting?

OpenAI reports that Terra scored 15 percentage points higher than GPT-5.6 Terra on Agents Last Exam at medium reasoning effort, at about one third the task cost, and performed on par with GPT-6 Astra on that test at one tenth the token price. On Terminal-Bench 4.0, a coding test, OpenAI says Terra solved 50.3 percent of tasks at maximum reasoning effort, up from 21.5 percent for GPT-5.6 Terra. On its internal set of prompts designed to trigger errors, OpenAI says the error rate at low reasoning effort fell from 53.6 percent with GPT-5.6 Terra to 16.7 percent with Terra, and to 4.3 percent at maximum effort.

Treat all of those as vendor reported results until independent tests confirm them. They are useful as a reason to test Terra, not as proof that it will pass your tasks. The claim that matters most for a workhorse model is not a benchmark score. It is cost per finished task on your own work, including retries, corrections, and human review time. A model that is three times faster on OpenAI's timing runs, as OpenAI claims versus Astra and Sol, still costs you money if its answers need heavy correction. Run the same five real tasks on Terra and on your current model, count the finished tasks that needed no fix, and compare the total bill including output tokens and retries. That is the test that decides a switch.

There is also a safety note that should not be skipped. OpenAI classifies Terra at the Critical cybersecurity capability threshold under its Preparedness Framework, the same threshold it applies to Astra, and says Terra retains protections developed for Astra. That classification is about what the model could do in the wrong hands, not about whether it is safe to use for office work. The practical consequence for builders is the same as for any computer use model: keep a human approval step before any action that moves money, sends a message to a customer, changes a record, or touches a production system. Cheap automation that acts without review is not a saving. It is a risk at scale.

Who should switch, and who should stay put?

Switch first for high volume, repetitive computer use work where the task is well defined and the output is easy to check: form filling, invoice entry, document drafting from templates, proposal sections from source packs, and first pass summaries. Those jobs benefit most from the lower input price, and their quality is easy to verify. Test Terra in ChatGPT Work, Codex, Copilot, or Bedrock, whichever surface your team already uses, and measure cost per finished task before moving a whole workflow.

Stay put, for now, for the hardest coding, deep research, and any task where a confident wrong answer is expensive. OpenAI itself points those jobs to Sol or Astra. Also stay put if your prompts routinely exceed 272,000 tokens and you cannot trim them, because the higher tier changes the math. Finally, do not switch on price alone if your current workflow depends on a feature Terra does not support in your chosen surface, such as a specific tool, a regional deployment, or a model policy setting. Price is only one part of a working agent.

Key takeaways:

• GPT-6.1 Terra launched 26 October 2026 at $1 input, $0.05 cached input, and $5 output per million tokens, half the standard price of GPT-6.1 Sol.
• It is available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu, through the API as gpt-6.1-terra, on Amazon Bedrock, and in GitHub Copilot.
• Cached input is 95 percent cheaper than fresh input, so reuse of documents and instructions is the biggest saving.
• Prompts over 272,000 input tokens reprice the whole request at double input and cache rates and 1.5 times output, so keep prompts under that line where possible.
• Benchmark and speed claims are OpenAI's own; measure cost per finished task on your own work before switching a whole workflow.

Sources: OpenAI, Introducing GPT-6.1 Terra, 26 October 2026, and OpenAI API model documentation and pricing; OpenAI Developer Community announcement, 26 October 2026; Amazon Web Services, OpenAI GPT-6.1 Terra generally available on Amazon Bedrock, 26 October 2026; GitHub Changelog, GPT-6.1 Terra in GitHub Copilot, 26 October 2026.

Related links: Top Use AI Blog, Open-source LLM API pricing, Best local coding model by VRAM