The short version: GPT-6.1 Terra pricing is $1 per million input tokens and $5 per million output, half of what GPT-6.1 Sol costs and half of the old GPT-5.6 Terra input rate, according to OpenAI’s API pricing page as read on October 27, 2026. The headline number most coverage repeated, cached input at five cents per million, is only half of how caching now bills. OpenAI’s prompt caching guide says that on GPT-5.6 and later, writing a prompt into the cache costs 1.25 times the normal input rate, and the default cache lasts 30 minutes. A prompt that is written and never reused costs 25 percent more than no caching at all.
For most owners this works out to pennies. For a business whose AI assistant carries a large instruction set, a full price book or a product manual, and answers customers a few times a day rather than a few times an hour, the setting the software vendor picked can make the same month cost $20 or under $5. That is the arithmetic below, plus the second dial that decides how long a customer waits before Terra says anything.
What GPT-6.1 Terra is, and where you will meet it
OpenAI released GPT-6.1 Terra on October 26 as its middle model: cheaper than GPT-6.1 Sol, more capable than the GPT-6 Luna tier, and pitched at work done on a computer, such as filling in forms, drafting proposals and moving through software screen by screen. The model page lists a context window of 1,050,000 tokens, up to 128,000 tokens of output, a knowledge cutoff of June 1, 2026, and the computer use tool among its supported tools.
OpenAI’s launch announcement claims Terra scored 15 percentage points higher than GPT-5.6 Terra on Agents’ Last Exam, a test built from professional software tasks, at about a third of the cost per task, as reported by Techstrong.ai. Those are OpenAI’s own benchmark figures and nobody outside OpenAI has reproduced them yet.
Most small businesses will not touch the API directly. They will meet Terra in three places. It is in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu subscribers, where a flat monthly plan covers usage. It is in GitHub Copilot, where the GitHub changelog says it is “billed at provider list pricing under usage-based billing,” which means the API rates below apply directly. And it will arrive, over the coming months, inside the software you already rent: the website chat widget, the review responder, the AI receptionist, the quoting tool. Those vendors pay OpenAI by the token and pass the cost to you in their own pricing. The settings in this article are theirs to choose, which is why they are worth asking about.
GPT-6.1 Terra pricing next to the models around it
Here is the standard rate card for prompts under 272,000 input tokens, read from OpenAI’s pricing page on October 27, 2026, with Anthropic’s closest models from Anthropic’s pricing page the same day for comparison. All figures are dollars per million tokens.
| Model | Input | Cache write | Cache hit | Output |
|---|---|---|---|---|
| GPT-6.1 Terra | $1.00 | $1.25 (30 min) | $0.05 | $5.00 |
| GPT-6.1 Sol | $2.00 | $2.50 (30 min) | $0.10 | $10.00 |
| GPT-5.6 Terra (previous) | $2.00 | 1.25x input | $0.20 | $12.00 |
| Claude Sonnet 5.5 | $2.00 | $2.50 (5 min) | $0.10 | $10.00 |
| Claude Haiku 5.5 | $0.10 | $0.125 (5 min) | $0.01 | $0.50 |
Two things stand out. Terra is priced at exactly half of Sol on every line, so the choice between them is purely about whether the job needs Sol’s extra capability. And both labs now charge more to write a cache than to send the same text uncached. Anthropic’s shortest window is 5 minutes and its longer one is an hour; OpenAI’s are 30 minutes and 12 hours, with the 12-hour write billed at 2.5 times the input rate. Haiku 5.5 rates apply to prompts up to 100,000 tokens; we covered how the two flagships compare in Opus 5.5 vs GPT-6 Sol.
How does prompt caching work on GPT-6.1 Terra?
Every time a business’s AI assistant answers a customer, the software sends Terra the same long block of instructions first: who the business is, the service area, the price book, the policies, how to hand off to a person. Only the last few lines, the customer’s actual message, are new. Caching lets OpenAI keep the processed version of that repeated block so it does not have to read it from scratch every time.
OpenAI’s guide says caching is “enabled by default for supported OpenAI models,” with a minimum of 1,024 tokens. On GPT-5.6 and later the first send of the block is a cache write, billed at $1.25 per million on Terra instead of $1.00. Every reuse inside the window is a cache hit at $0.05. The guide is careful to say the write is “not an additive fee,” meaning the write rate replaces the normal input rate rather than stacking on top of it. But it is still a 25 percent premium over sending the block uncached, and it only pays off if the block gets reused before the window closes.
That is the whole mechanism, and it makes caching a bet on timing. If the next customer message arrives inside 30 minutes, the bet pays: the block costs five cents per million instead of a dollar. If it arrives an hour later, the cache has lapsed, the block is written again at the premium price, and the business has paid more than if caching had been off.
What GPT-6.1 Terra costs a busy shop and a quiet one
To make the bet concrete, we worked two examples. Both use the same assistant: a 60,000-token instruction block, roughly a full residential price book plus equipment notes and policies for a heating and cooling company. Each customer message adds 300 new tokens and each reply, including Terra’s reasoning at the default medium setting, runs 400 output tokens. These token counts are our assumptions for illustration, not measurements; the per-token prices are OpenAI’s. Both shops work 22 days a month.
- The busy shop gets 40 messages a day, close enough together that the cache never lapses during business hours. That is 880 conversations a month.
- The quiet shop gets 12 messages a day, spread hours apart, so every message arrives after the 30-minute window has closed. That is 264 conversations a month.
Here is the monthly bill for each under three settings: OpenAI’s default 30-minute cache, the 12-hour cache, and caching switched off for those requests, which the guide says happens when explicit mode is set with no cache breakpoints.
| Monthly cost on Terra | 30-min cache | 12-hour cache | No cache |
|---|---|---|---|
| Busy shop, 880 messages | $6.25 | $7.90 | $54.82 |
| Quiet shop, 264 messages | $20.41 | $4.63 | $16.45 |
The quiet shop sends a third as many messages and, on the default setting, pays more than three times as much as the busy one. Each of its messages lands after the cache has lapsed, so each one rewrites the 60,000-token block at $1.25 per million, which is 7.5 cents a message before the customer’s words are even counted. Switching caching off would save it about $4 a month. Switching to the 12-hour cache does far better: one write a day at $0.15, then eleven hits at three-tenths of a cent each, for a total of $4.63.
The busy shop shows the reverse. Its messages keep the 30-minute cache warm all day, so the default is already the cheapest option, and the 12-hour write would add $1.65 a month for retention it never needs. Neither setting is right for everyone. The right one depends on the shape of a business’s day, and the vendor setting up the assistant is the only one positioned to know it.
The worked math, for anyone who wants to check it: the quiet shop’s default bill is 264 cache writes of 60,000 tokens at $1.25 per million ($19.80), plus 264 messages of 300 fresh input tokens at $1 per million and 400 output tokens at $5 per million ($0.61). With a small 6,000-token instruction block the same pattern holds at a tenth of the size: about $2.59 on the default against $1.01 on the 12-hour cache. Pennies either way, which is why this matters most to businesses whose assistants carry a large reference library.
The second dial: how long Terra thinks before it answers
Terra has six reasoning settings, from none through low, medium, high and xhigh to max, with medium as the default, per the model page. Higher settings think longer, write more reasoning tokens, and bill those tokens as output.
Artificial Analysis, an independent benchmarking firm, publishes measurements of the max setting. When we read its Terra page on October 27, it listed a median time to first answer token of 77.96 seconds, against a median of 3.69 seconds across the models it tracks. It also counted 99 million output tokens to run its test suite, against a median of 63 million, and called the model “somewhat verbose.” Its figures are rolling medians and they have already moved since launch day, so treat them as a snapshot, not a constant.
The practical reading is simple. A minute and a quarter of silence is fine for a job that runs overnight: reconciling a month of supplier invoices, filling in a stack of warranty forms, checking every quote sent last week against the current price book. Terra’s computer-use focus suits that kind of patient back-office work, and that is where its benchmark claims are pointed. The same wait is unacceptable for a customer typing into a chat box on your website. A tool that answers customers live should run Terra at low or medium. We made the same point about a model that always reasons by default in Grok 4.7 pricing; with Terra the dial is wider, and it sets both the bill and the wait.
Three questions to ask any vendor that switches you to Terra
If a tool you pay for announces it now runs on GPT-6.1 Terra, these three questions take five minutes and tell you whether the savings reach you.
- Which reasoning setting does the customer-facing part use? Anything above medium on a live chat or text conversation is a slow answer waiting to happen.
- Which cache setting do you use, and did you match it to how often my customers write in? A vendor serving thousands of businesses may have picked one setting for all of them. If your inquiries come in a few times a day, the 12-hour cache is likely the cheaper one for you.
- Does a lower model cost change what I pay? Terra at least halves the per-token price against Sol and against GPT-5.6 Terra. If the vendor bills you per conversation or per seat, ask whether the price moves. Some will pass it on; some will keep it as margin, which is their right, but you should know which.
The larger story here is good news. A model that OpenAI says keeps pace with its premium tier on one benchmark of professional tasks now costs a dollar per million tokens going in, and the arithmetic for a real small business assistant lands under $10 a month on the right setting. The fine print simply moved from the price to the settings, and the settings are a conversation an owner can have. For more on how plan pricing is moving at the subscription level, see our look at ChatGPT Pro 200 against Claude Max.
Frequently Asked Questions
How much does GPT-6.1 Terra cost?
On OpenAI’s API pricing page, read October 27, 2026, GPT-6.1 Terra costs $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens for prompts up to 272,000 tokens. Cache writes cost $1.25 per million for the default 30-minute retention or $2.50 for 12 hours, and batch and flex processing are 50 percent cheaper.
Is GPT-6.1 Terra cheaper than GPT-6.1 Sol?
Yes. Terra is priced at exactly half of Sol on every line of the rate card: $1 against $2 for input and $5 against $10 for output. OpenAI positions Terra for computer-based work such as forms, proposals and software tasks, and Sol for jobs that need more capability.
Can prompt caching make an AI assistant more expensive?
It can. OpenAI’s caching guide says that on GPT-5.6 and later models a cache write costs 1.25 times the normal input rate, and the default cache lasts 30 minutes. If a business’s messages arrive more than 30 minutes apart, every message pays that write premium and the cache is never reused, so the bill ends up higher than with caching off.
Why does GPT-6.1 Terra sometimes take over a minute to answer?
At its highest reasoning setting, max, Terra thinks before it writes. Artificial Analysis measured a median of about 78 seconds before the first answer token at max on October 27, 2026. The default medium setting is far quicker, and customer-facing tools should use low or medium.
If your software vendor has already moved you to Terra, did they tell you which settings they chose, and did your bill change?
