OpenAI’s GPT-6.1 Terra lists at $1.00 per million input tokens and $5.00 per million output. What that rate does not tell you is that the whole request is re-priced once a prompt passes 272,000 input tokens.

Send the API 270,000 input tokens and ask for 4,000 output tokens on the Standard tier with no caching, and the request costs $0.29: 270,000 at $1.00 per million for input, plus 4,000 at $5.00 per million for output. Send 280,000 input tokens with the same 4,000-token answer, and the bill is $0.59.

The 10,000 extra tokens are 3.7% more input, and they add $0.30, roughly doubling the cost of the request. The rule sits on the Terra model page: “Prompts with more than 272K input tokens are priced at 2x input and cache rates and 1.5x output for the full request.” The last two words carry the cost. Past 272,000 input tokens the full request is billed at $2.00 per million for input and $7.50 per million for output. Those final 10,000 tokens work out to an effective $30.00 per million, because they carried the rest of the prompt across the line with them.

Cache writes cost more than cache reads, and both move above 272K

Cached input on Terra costs $0.05 per million tokens, one twentieth of the standard input rate. That is a read price. Writing to the cache is billed separately: $1.25 per million tokens for 30-minute retention and $2.50 per million for 12-hour retention, and the model page quotes both figures “before long-context adjustments”.

That qualifier matters because the long-context rule doubles input and cache rates. Applying the documented multipliers gives $2.50 per million for a 30-minute write above 272K and $5.00 per million for a 12-hour write, with cached reads rising from $0.05 to $0.10. OpenAI does not print those post-cliff write figures on the model page, so treat them as derived from the rule rather than quoted from the price table.

The write charge is not an add-on to the input charge. Each input token is billed at exactly one of three rates: uncached input, cached input, or cache write. A request can carry up to four writes, and the minimum cacheable prefix is 1,024 visible input tokens, as the prompt caching guide sets out.

Caching pays off when a prefix is reused, and it is a losing trade on a single long request. Take the 280K-token prompt again with 250,000 tokens served as cache reads, 30,000 fresh input tokens and 4,000 output tokens: 250,000 at $0.10 plus 30,000 at $2.00 plus 4,000 at $7.50 comes to $0.115. The write that made it possible costs $0.625 at 30-minute retention or $1.25 at 12-hour retention, so running that request once is cheaper without caching. Run it repeatedly against a shared prefix and the write pays for itself.

Two operational limits matter for agents. Cache entries live on individual machines and are not shared across organizations, so a request has to route to the machine holding the entry. Caches also cannot cross regional-processing boundaries.

Batch and Flex halve the bill, Fast doubles it

“Batch and Flex prices are 50% lower than Standard,” the pricing page says. That discount applies to whatever rate the long-context rule has already set. The same 280K-in, 4K-out request at Batch pricing is 280,000 at $1.00 per million plus 4,000 at $3.75 per million, which is the long-context output rate halved by the Batch discount, or $0.295. For work that can tolerate queued processing, Batch and Flex roughly cancel the input side of the cliff.

Fast mode is the opposite lever. Fast mode prices are 2x Standard, which puts Fast Terra at $2.00 per million input tokens and $10.00 per million output tokens in the short-context regime. Combine that with the long-context multipliers and the 280K-token request costs $1.18, with input at $4.00 per million and output at $15.00 per million. OpenAI does not publish that combined Terra figure. The $1.18 total is arithmetic on the documented multipliers, following the precedent on the previous flagship, GPT-5.6 Sol, where the Fast mode guide prices long-context requests at double the short-context input rate and $60 per million output.

OpenAI’s pricing page lists Ultrafast, the fastest and most expensive tier, for GPT-6 Astra at $60.00/$300.00 per million tokens and GPT-6.1 Sol at $12.00/$60.00. Terra is not on that list, so Fast is the only speed upgrade available on this model.

Regional processing adds 10%, and you have to opt in

Regional processing endpoints carry a 10% uplift for models released on or after 5 March 2026, and FedRAMP endpoints carry the same uplift. It is not an automatic surcharge on non-US accounts. Data residency is configured per project, and requests without it use global processing. Using residency in any region outside the United States requires approval for abuse-monitoring controls and a Modified Retention amendment, so the surcharge comes with paperwork before it comes with a bill. The data controls guide sets out both requirements.

Terra supports EU data residency and regional processing on Standard, Flex, Batch and Fast. Fast mode offers EU residency for Sol, Terra and Luna, but not Astra, which matters when a speed tier and a residency requirement both apply to the same workload.

Rate limits scale with your spend tier, not your country

Tier Requests per minute Tokens per minute Batch queue limit
Free Not supported Not supported Not supported
Build 5,000 1,000,000 3,000,000
Launch 10,000 4,000,000 200,000,000
Grow 15,000 40,000,000 15,000,000,000

There is no free tier. These limits are identical to the previous GPT-5.6 Terra and follow the usage tier attached to the account. They constrain long-document work separately from price: a Build account’s ceiling of 1 million tokens per minute allows about three and a half 280K-token prompts a minute, and its 3 million-token batch queue holds around ten of them.

What Terra does not support, and the fine-tuning shutdown behind it

The Terra model page marks fine-tuning as not supported. That is a Terra-specific line. Separately, OpenAI says it is winding down the fine-tuning platform: it is no longer accessible to new users, existing users can create training jobs for the coming months, and fine-tuned models remain available for inference until their base models are deprecated. The deprecations page gives the dates. New organizations were blocked on 7 May 2026. From 2 July 2026, an organization that had not run inference on a fine-tuned model in the previous 60 days could no longer create training jobs. From 6 January 2027, the remaining existing customers lose the ability to create new training jobs at all. The only model in the fine-tuning price table is o4-mini-2025-04-16.

So the answer to “can I fine-tune Terra” is no, and the answer to “can I start fine-tuning anything else” is increasingly also no. Terra takes text and images in and returns text, and reasoning effort runs from none to max. Audio and video are not supported either.

For procurement reviews, the data controls guide states that API data has not been used to train or improve OpenAI’s models since 1 March 2023 unless the customer opts in, and that abuse-monitoring logs are retained for up to 30 days by default. Zero Data Retention and Modified Abuse Monitoring are available on approval.

An independent test of the max-effort variant

OpenAI’s own model page describes Terra as a “fast, efficient model for visual reasoning, precise computer use, and clear, natural writing”. That is vendor copy, not a measurement.

Artificial Analysis has run the model at maximum reasoning effort. Its snapshot lists an Intelligence Index score of 48, 25th of 221 models, at $0.40 per Index task. The blended rate on the same page is $0.73 per million tokens at a 7:2:1 cache-hit, input, output ratio, an assumption about workload mix rather than a price any single request receives. Output speed is 90.9 tokens per second, and time to first token is 133.64 seconds against a tier median of 3.78 seconds, which is consistent with a reasoning model thinking before it answers. The run produced 130M output tokens on the Index against a median of 81M, so this is a verbose configuration.

The limits: one evaluator’s ten-test suite, one test on that suite marked under review, figures taken from a first-party API, and no coverage of the >272K regime or region-specific deployments. Read 48 and $0.73 as that suite’s results under those conditions.

How Terra compares with the model it replaces

The direct predecessor, GPT-5.6 Terra, was priced at $2.00 per million input tokens, $0.20 cached and $12.00 output. Terra cuts input by half and output by 58%. The old long-context rule said “2x input and 1.5x output for the full request”, with no mention of cache rates; the current wording adds “and cache rates”.

Terra accepts 1,050,000 tokens of context and returns up to 128,000 output tokens, with a knowledge cutoff of 1 June 2026. Among the current models, Terra sits between Sol at $2.00/$10.00 and Luna at $0.10/$0.50, with Astra at the top at $10.00/$50.00.

Terra’s $1/$5 rate is also the concrete counter-datapoint to the open-weight cost argument we covered earlier, the claim by IBM’s chief executive that open-weight models reach roughly 90% of frontier effectiveness at 10 to 25% of frontier token cost. Neither side of that comparison has a like-for-like benchmark, and the answer shifts with which frontier and open-weight models you place side by side, so the two figures are best read as pricing positions.

What could not be verified

The launch post at openai.com returned HTTP 403 on 26 and 27 October, with and without a trailing slash, so we could not read it, and its claims are not quoted here. Techmeme and Hacker News carried the launch, and those are aggregator references rather than primary verification. Apart from the IBM comparison above, every figure in this article comes from OpenAI’s developer documentation and Artificial Analysis, both read live on 27 October 2026. The derived figures, meaning the post-272K cache-write prices and the Fast plus long-context combination, are labelled where they appear, because OpenAI prints the multipliers and not the combined result.