OpenAI's GPT-6.1 Terra lands at $1/$5 — and nothing in OpenAI's catalogue explains the 10x spread above it

OpenAI’s changelog entry for 26 October 2026 runs to one sentence. gpt-6.1-terra, released on the Responses and Chat Completions endpoints: “a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing.”

The rate card is the more interesting document. GPT-6.1 Terra lists at $1.00 per million input tokens and $5.00 per million output — exactly half of GPT-6.1 Sol, which shipped four weeks earlier at DevDay. It has the same 1,050,000-token context window as Sol. The same 922,000-token maximum input. The same 128,000-token output cap. The same 272K-token surcharge cliff. The same text-and-image input and text-only output. The same endpoints, the same feature list, the same tool list and the same published rate limits.

What it does not have, and what Sol and the $10/$50 GPT-6 Astra do not have either, is a published capability grade. OpenAI’s catalogue does not rate models qualitatively at all: the field labelled “Reasoning” on each model card is a link to the reasoning-effort guide followed by that model’s effort ladder. So across a tenfold spread in list price there is no vendor number that speaks to capability — and the structured field that does vary runs against price rather than with it.

Correction (27 October 2026): an earlier version of this article said OpenAI’s model catalogue rates GPT-6 Astra, GPT-6.1 Sol and GPT-6.1 Terra all as “Highest” reasoning and GPT-6 Luna as “High,” and built its argument on that badge spanning a 10x price range. No such rating is published. OpenAI’s catalogue carries no qualitative capability grade: the field labelled “Reasoning” on each model card is a link to the reasoning-effort guide followed by that model’s effort ladder, and the four models’ ladders are low–max on Astra and Sol and none–max on Terra and Luna. The article also said all four share “the same six reasoning-effort levels”; Astra and Sol have five. The headline, the comparison table and the argument have been rewritten against the ladders OpenAI actually publishes, which cut across the price order rather than following it. Every price, context, cutoff and tier figure in the original was re-verified and stands. Corrections log.

The four rungs, side by side

Model	Effort levels	Input / output per 1M	Cached input	Cache as % of input	Context	Knowledge cutoff	Ultrafast tier
`gpt-6-astra`	5 (`low`–`max`)	$10.00 / $50.00	$1.00	10%	1,050,000	30 Apr 2026	Yes
`gpt-6.1-sol`	5 (`low`–`max`)	$2.00 / $10.00	$0.10	5%	1,050,000	30 Apr 2026	Yes
`gpt-6.1-terra`	6 (`none`–`max`)	$1.00 / $5.00	$0.05	5%	1,050,000	1 Jun 2026	No
`gpt-6-luna`	6 (`none`–`max`)	$0.10 / $0.50	$0.01	10%	1,050,000	18 May 2026	Yes

Prices and specifications read from OpenAI’s model pages and pricing page on 27 October 2026. Batch and Flex are 50% below standard on all four; Fast mode is 2x standard; regional-processing endpoints add 10%.

Read down the columns and the shape of the problem is clear. The context window column is constant. The effort-level column splits the family in half, but not along the price line — the two models that accept none, and can therefore be run as non-reasoning models, are the two cheapest. The columns that vary in a way a buyer can use — knowledge cutoff, the effort floor and which speed tiers exist — appear on neither the pricing page nor in any summary table OpenAI publishes. And the cutoff column is inverted against price: the cheapest model in the family’s upper tier knows the most recent things.

Terra in Fast mode is Sol at list

The cleanest piece of arithmetic here is one OpenAI does not draw attention to. Fast mode bills at 2x standard. So:

GPT-6.1 Terra, Fast mode: $2.00 input, $10.00 output per million.
GPT-6.1 Sol, standard: $2.00 input, $10.00 output per million.

To the cent. The decision to buy lower latency on Terra is therefore the same spend as running Sol at list, which turns a tier upgrade into a model choice. At that price point the two options differ on: knowledge cutoff (1 June against 30 April, favouring Terra), stated tuning (vision and prose against coding, favouring whichever matches your work), throughput (Fast delivers up to roughly 2.5x standard speeds, favouring Terra), and escalation headroom.

That last one is the asymmetry worth planning around. Terra does not appear in the Ultrafast pricing table at all — only Astra and Sol do. A team that standardises on Terra because it is cheap, and then discovers eighteen months later that its latency budget has tightened, cannot buy its way up a tier. It has to re-qualify a different model. Sol at the same effective price keeps that door open. Whether that option is worth paying for depends entirely on how stable your latency requirements are, but it is a real difference hiding behind two identical numbers.

The cache multiplier does not follow the price

The second detail likely to misprice a deployment is that cached input is not a constant fraction of base input across the family.

Astra: $1.00 cached against $10.00 base — 10%
Sol: $0.10 against $2.00 — 5%
Terra: $0.05 against $1.00 — 5%
Luna: $0.01 against $0.10 — 10%

The two middle rungs discount cache reads twice as hard, proportionally, as the flagship above them and the budget model below them. For agentic workloads — a large fixed system prompt, a stable tool schema, retrieved documents re-sent every turn — cache reads are usually the largest line on the bill, which makes that multiplier more consequential than the headline rate.

Layer batch processing on top and Terra’s cached input reaches $0.025 per million tokens. Against Astra’s $10.00 standard input, that is a 400x spread inside a single vendor’s current flagship family. Anyone who sized a budget from the base input rate of one of these models and then changed models, tiers or caching strategy has a number with no predictive value.

A worked task, and the one ratio that stays put

Take a plausible agentic workload and price it on three of the four rungs. A support-triage agent, 1,000 tasks a month, each task sending 40,000 cached input tokens (a fixed system prompt plus a stable tool schema, served from cache), 2,000 fresh input tokens and emitting 1,500 output tokens.

	Cached input	Fresh input	Output	Per task	1,000 tasks
`gpt-6.1-sol`	$0.0040	$0.0040	$0.0150	$0.0230	$23.00
`gpt-6.1-terra`	$0.0020	$0.0020	$0.0075	$0.0115	$11.50
`gpt-6-luna`	$0.0004	$0.0002	$0.0008	$0.0014	$1.35

Standard-tier rates, no batch discount, prompt under the 272K surcharge threshold.

Terra comes in at exactly half of Sol — and it will come in at exactly half of Sol for any mix of cached input, fresh input and output, because Terra halves all three rates by the same factor. That is unusual and it is useful: the Sol-to-Terra decision is the rare model comparison where you do not need to know your own token mix to know the saving. A 50% cut is a 50% cut whether your workload is cache-dominated or output-dominated.

The Terra-to-Luna comparison is the opposite. Luna is 10x cheaper than Terra on input and output but only 5x cheaper on cached input, so the gap between them depends entirely on how cache-heavy the work is. In the table above it comes out at 8.5x, not 10x. Push the cached share higher — a bigger fixed prompt, more turns per task — and it compresses further toward 5x. That is still a large gap, but a budget built on “Luna is a tenth of the price” will be wrong in the direction that matters, and wrong by more the better your caching gets.

What this means for a buyer

If you are running GPT-6.1 Sol today, the default has probably moved — but moved is not the same as decided.

For vision work, document and screenshot interpretation, computer use, and drafting or revising prose, Terra is what OpenAI built for the job, at half the price, with a newer cutoff. The burden of proof has shifted onto staying with Sol.

For complex coding, Sol is still the stated target and the one with Ultrafast behind it. Test before moving; don’t assume the halved rate card carries across a task type the vendor explicitly points elsewhere.

For anything where the model is verbose, remember that output bills at five times input on every model in the table. A difference in output-token volume between Terra and Sol will move the invoice more than the 50% cut in the rate card does. That is measurable in an afternoon on your own traffic and is not inferable from any published number.

And for anyone building a selection process: there is no published capability rating to put in it. OpenAI grades nothing qualitatively, so a procurement checklist with a “vendor capability tier” column has nothing to fill it from across a 10x price range. The axes left standing are knowledge cutoff, stated task fit, available speed tiers, the cache multiplier, the effort floor, and measured cost per completed task on your own workload. The first five are scattered across separate documentation pages. The sixth is the only one that decides anything.

A dedicated GPT-6.1 Terra vs GPT-6.1 Sol comparison sets out the tier-by-tier arithmetic, the Fast-mode collision and the effort-ladder difference in full. The API cost calculator now carries Terra’s rates alongside the rest of the GPT-6 family, and the prompt cache retention guide has the cache-write arithmetic for the 30-minute and 12-hour retention windows. For the broader product picture, the ChatGPT review tracks which models reach the consumer plans and when.

Frequently asked questions

What exactly did OpenAI ship on 26 October?

A new model ID, gpt-6.1-terra, on the Responses and Chat Completions endpoints. The changelog describes it as "a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing." The model page prices it at $1.00 per million input tokens, $0.05 per million cached input, and $5.00 per million output, with cache writes at $1.25 for 30-minute retention and $2.50 for 12-hour. It takes text and image input and emits text only. The context window is 1,050,000 tokens with a 128,000-token output cap, the knowledge cutoff is 1 June 2026, and reasoning.effort accepts none, low, medium (the default), high, xhigh and max. Prompts above 272K input tokens are billed at 2x input and cache rates and 1.5x output for the whole request. There is no dated snapshot: the only snapshot listed is gpt-6.1-terra itself.

How is it different from GPT-6.1 Sol, which is four weeks old?

On almost every structural number the two are identical — 1,050,000-token window, 922,000-token maximum input, 128K output cap, the same 272K surcharge, text-and-image in, text out, the same endpoints, features and tools, and the same published rate limits. The one structural difference is the reasoning-effort ladder: Terra accepts none, low, medium (the default), high, xhigh and max, while Sol accepts only low, medium (default), high, xhigh and max. Sol has no none rung, so it cannot be switched into non-reasoning mode at any price. Terra costs half as much ($1/$5 against $2/$10, with cached input at $0.05 against $0.10) and knows more recent things (1 June 2026 cutoff against 30 April). The differences that remain are positioning and latency headroom. OpenAI points Sol at "complex coding, computer use, and professional work" and Terra at "visual reasoning, precise computer use, and clear, natural writing" — so coding is Sol's stated territory and vision is Terra's. And Sol can be bought faster: Terra does not appear in the Ultrafast pricing table at all.

What is the Fast mode arithmetic?

Fast mode is 2x standard rates, so GPT-6.1 Terra in Fast mode costs $2.00 per million input and $10.00 per million output — to the cent, what GPT-6.1 Sol costs at standard rates. That makes the latency upgrade on Terra a decision about which model you would rather have at $2/$10, and it is a genuine coin-flip: Terra-Fast gives you the newer knowledge cutoff and roughly 2.5x standard throughput, Sol-Standard gives you the coding-oriented tuning and the option to escalate to Ultrafast later. The one asymmetry is headroom. If you build on Terra and later need Ultrafast, there is no Ultrafast to buy — you have to change model, not change tier. Teams whose latency requirements might tighten should weigh that before taking the cheaper rate card.

Does the cheapest model have the best cache economics?

Not in the way the price ladder implies. Cached input is 5% of base input on GPT-6.1 Terra ($0.05 against $1.00) and on GPT-6.1 Sol ($0.10 against $2.00), but 10% on both GPT-6 Astra ($1.00 against $10.00) and GPT-6 Luna ($0.01 against $0.10). In absolute terms the ladder still runs in price order — Luna's $0.01 cache read is the cheapest of the four and Terra's $0.05 is five times that. What the multiplier changes is how much caching is worth engineering for. On Terra, moving a token from uncached to cached saves 95% of its cost; on Luna it saves 90%. So the two middle rungs reward cache design more than the flagship above them or the budget model below them, and an agent that re-sends a large fixed system prompt every turn closes more of the Sol-to-Terra gap than the headline 50% cut suggests. Stack batch on top and Terra's cached input reaches $0.025 per million, which is one four-hundredth of Astra's standard input rate — the full spread inside a single vendor's current flagship family.

Which model should a team on GPT-6.1 Sol actually run now?

Run a measured comparison rather than a swap, but the default has probably moved. For vision-heavy work, document and screenshot interpretation, computer use, and drafting or revising prose, Terra is the model OpenAI built for the job and it costs half as much with newer knowledge — the burden of proof is now on staying with Sol. For complex coding, Sol remains the stated target and the one with Ultrafast available, so test before moving. For anything where output volume dominates, note that output bills at five times input on both, so verbosity differences between the two will move the invoice more than the halved rate card does. And for anyone budgeting from the published numbers: the effort setting on these models has historically swung cost per task by more than 5x on an unchanged rate card, which is a larger effect than the entire Sol-to-Terra price cut.

What does this do to the model-selection problem generally?

It removes a shortcut buyers often assume exists. OpenAI's catalogue publishes no qualitative capability or intelligence grade for any model: the field labelled "Reasoning" on each model card is a link to the reasoning-effort guide followed by that model's effort ladder, not a rating. So across a tenfold spread in list price — Astra at $10/$50, Sol at $2/$10, Terra at $1/$5 — there is no vendor-published capability number to compare at all, and all four models in the family carry the same 1,050,000-token window and the same 128K output cap, so those columns discriminate nothing either. The one structured field that does vary is the effort ladder, and it runs against price rather than with it: Terra and Luna, the two cheapest, accept a none rung that Astra and Sol do not. What is left to choose on is knowledge cutoff, stated task positioning, which speed tiers exist, the cache multiplier, the effort floor, and measured performance on your own traffic. Only the last of those is decisive, and none of the others appear on the pricing page together.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.