OpenAI released GPT-6.1 Terra on 26 October. It is the middle model in the GPT-6 family, priced at $1 per million input tokens and $5 per million output tokens, which is half what OpenAI charged for the model it replaces. The first independent test is out, and it is striking: per finished task, Terra costs about a tenth of what Claude Sonnet 5.5 or Gemini 4 Argon cost, for a score only a few points lower. There is one catch that decides whether the saving is real for you. This article covers what OpenAI shipped, what it costs, what Artificial Analysis measured, how it compares with its sibling Sol and with Anthropic and Google, and who should switch.
What OpenAI released on 26 October
OpenAI describes Terra as a fast, efficient model for "the communication and computer-based tasks that fill the workday". The facts from its launch post:
- What it is. A step up from GPT-5.6 Terra, built for reading images and screens, using software the way a person does (clicking, typing, filling forms), and writing clearer text. It reads text and images and writes text.
- Who can use it today. OpenAI says 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. The Business plan is not in that list, and the Business pricing page still showed GPT-5.6 Terra, not 6.1, when we checked on 29 October. If your team is on Business seats, check before you promise anyone the new model. - The use case OpenAI picks. "Reading incoming invoices and entering their details into accounting software". That is exactly the kind of high-volume office work a small business has, which is why the price matters more than the benchmarks.
- Safety. OpenAI says Terra meets its "Critical" threshold for cybersecurity under its Preparedness Framework and keeps the protections built for GPT-6 Astra. In plain terms, OpenAI treats it as strong enough at security work to need guardrails.
The price: $1 in, $5 out, and a cache that costs to fill
These figures are from OpenAI's API pricing page, read on 29 October. Prices are per million tokens, standard speed.
| Model | Input | Output | Cached input |
|---|---|---|---|
| GPT-6 Astra | $10.00 | $50.00 | $1.00 |
| GPT-6.1 Sol | $2.00 | $10.00 | $0.10 |
| GPT-6.1 Terra | $1.00 | $5.00 | $0.05 |
| GPT-6 Luna | $0.10 | $0.50 | $0.01 |
OpenAI says Terra's standard prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. Two details are easy to miss. First, cached input is $0.05, but writing to the cache is not free: the same page lists $1.25 per million tokens for a 30-minute cache write and $2.50 for a 12-hour write. Caching only saves money if the same long prompt is reused many times. Second, data-residency endpoints carry a 10% uplift, which matters if you need your data processed in a specific region.
To make that concrete, here is our own arithmetic on an invoice-reading job. We assume 1,500 input tokens and 300 output tokens per invoice. That is an assumption, not a measurement, and it ignores the reasoning tokens a model may add. On those numbers one invoice costs about $0.003 on Terra, about $0.006 on Claude Sonnet 5.5 and about $0.0003 on GPT-6 Luna. For 10,000 invoices a month that is roughly $30, $60 and $3. All three are small. The real differences come from how many tokens each model actually uses, which is what the next section measures.
What Artificial Analysis measured
Artificial Analysis is an independent tester. Its Intelligence Index (version 5.0) averages nine evaluations of agentic and professional work, including coding in a terminal, spreadsheet and document tasks, and workflow automation. It also records what it cost to run each task. These figures are from its model pages, read on 29 October 2026.
Terra scores 51, which Artificial Analysis ranks 12th of 52 models and "well above average" for its price class (the median is 42). It generates 91 output tokens a second, a little faster than the average of 87. At $0.96 per task it is the cheapest model on the chart that scores above 50.
One caution on the Gemini number. Artificial Analysis lists Argon at $4 in and $20 out per million tokens. Google's pricing page shows that as the price from 1 December; until 30 November the introductory price is $2 and $10. So Argon's cost per task is stated at the price it will have in a month, so what you would pay while the introductory price lasts is lower than the chart shows.
Terra against Sol: the saving is smaller than the price list says
On the price list Terra is half of Sol. Per task it is not. Terra costs $0.96 per task against Sol's $1.27, a saving of about a quarter, and it scores 51 against 55. The reason is in the token counts. To complete the index Terra produced 99 million tokens; Sol produced 64 million. Terra is more talkative, so half the price per token becomes only about 24% less per job.
- Pay about $0.31 more per task, get four more points. If one wrong answer costs you more than a few cents, Sol is the better buy for the hard jobs: contract review, coding, anything a person will not double-check.
- Terra is faster to the end of a task, by OpenAI's measure. OpenAI says Terra completes tasks over three times faster than both Astra and Sol. That is OpenAI's own timing, but Artificial Analysis does show Terra at 91 tokens a second against Sol's 81, which points the same way.
- OpenAI itself says Sol stays better for hard coding. The launch post says that "for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice".
Terra against Claude and Gemini: the real news
The headline comparison is with the rival mid-tier and top-tier models. Claude Sonnet 5.5 costs $2 in and $10 out per million tokens, the same as Sol, yet Artificial Analysis measured $10.02 per task for a score of 54. Terra gets 51 for $0.96. That is about a tenth of the cost for three points less. Claude Opus 5.5, at $4 and $20, scores 58 at $10.04 per task.
The gap is almost entirely verbosity. Sonnet 5.5 produced 340 million tokens to finish the index, against Terra's 99 million. Opus 5.5 produced 220 million. Every extra token is billed at the output rate. We saw the same effect in Mistral Large 4: the price list tells you the cost of a token, and only a task-level test tells you the cost of a job.
OpenAI makes a bigger claim. It says that at maximum effort Terra is over three times as token efficient as Opus 5.5 and a quarter of its price per token, giving "a ~9x lower cost per task". That is OpenAI's own comparison on its own chosen benchmark, Agents' Last Exam, and we have not been able to check it independently. Artificial Analysis's separate figure of $0.96 against $10.04 is about 10 times, which is consistent in direction, though from a different test.
The price per million tokens is the sticker price. The cost per task is the bill, and on the same independent test it now differs by about ten times between OpenAI and its two main rivals.
What OpenAI says it improved, and how to read it
These are all OpenAI's own numbers from the launch post. None has been independently reproduced yet, so treat them as claims.
- Fewer made-up answers. On OpenAI's internal set of error-inducing prompts, Terra's error rate at low effort fell from 53.6% (GPT-5.6 Terra) to 16.7%. At maximum effort it is 4.3%, which OpenAI says nearly matches Astra and Sol. A test built by the vendor is a direction, not a guarantee.
- Terminal coding. On Terminal-Bench 4.0, Terra solves 50.3% of tasks at maximum effort, against 21.5% for GPT-5.6 Terra.
- Computer use. On Agents' Last Exam at medium effort, OpenAI says Terra scores 15 percentage points above GPT-5.6 Terra at about a third of the cost, and performs on par with Astra at a tenth of its price per token.
Run your own check before you trust any of them. Take 50 real documents, run them through the old model and the new one, and have a person mark both. It costs an afternoon and tells you more than every chart above.
Who should switch, and who should wait
- High-volume office work: switch and test. Reading invoices, extracting fields from forms, drafting routine replies, filling in web forms. Terra is the cheapest model above a score of 50 on the independent test, and these jobs are exactly what OpenAI built it for. Document-heavy work is where model price shows up on a bill, which is why we watch it for products like IO Snack Accounts.
- Hard, high-stakes work: stay on Sol, Astra or Opus. For legal review, difficult code or analysis nobody re-checks, the four extra points of Sol, and the further gap above it, are worth paying for.
- Very simple jobs: look at Luna or Haiku. GPT-6 Luna scores 28 at $0.15 a task and Claude Haiku 5.5 scores 36 at $0.41. If the work is sorting or tagging, you do not need Terra. See our Haiku 5.5 test.
- Gemini 4 Carbon: wait. Google announced it on 28 October, ahead of Terra on the index at 59, with the same introductory price as Argon. But it is only open to a small group of security defenders for now, and Google gives no date for wider access. It is not a model you can buy today.
- Whatever you pick, cap the output. A verbose model is a billing risk. Set a maximum output length per request and a monthly spending cap, and log the cost of each job.
Your next step: measure your job, not the launch
Launch week always produces a new "cheapest model". The one that is cheapest for you is the one that finishes your own jobs in the fewest tokens at acceptable quality. We explain how to measure that in the cheapest AI model for a business, by cost per task. For the other half of the mid-tier fight, read Claude Sonnet 5.5 against GPT-6.1 Sol, and for the monthly bill of the apps themselves, see what AI costs a small business per month.
If you would rather use AI than manage models, that is what our products are for. IO Snack Accounts uses AI to categorise entries and flag anomalies as you post them, and the AI Chat Assistant handles customer questions on your website. If you are weighing a bigger project and want a second opinion on which model fits it, talk to us.
