GPT-6.1 Terra: $1 per million input tokens, half of Sol's price

OpenAI filled the missing rung in the middle of its model line today, and priced it below everything else it sells.

OpenAI released GPT-6.1 Terra on Monday at $1 per million input tokens, $5 per million output tokens and $0.05 per million cached input — exactly half the price of GPT-6.1 Sol on every line, and a tenth of the flagship GPT-6 Astra. It replaces GPT-5.6 Terra, the one tier OpenAI left behind when it moved Sol and Luna onto the GPT-6 family in September; as 9to5Mac noted at the time, Terra was excluded from that update and has been running a generation stale ever since. The model is live now for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and for developers through the API as gpt-6.1-terra, OpenAI said in its announcement.

The pitch is narrow and specific: not a smarter flagship, a cheaper one that can actually drive software. OpenAI describes Terra as combining detailed visual reasoning with "precise computer use" and clearer, more natural writing, aimed at the tasks that fill a workday — drafting proposals, reading technical drawings, filling in forms, editing video.

The price ladder, in one place

The family now runs, per million tokens: Astra at $10 in and $50 out with cached input at $1; Sol at $2 and $10, cached at $0.10; Terra at $1 and $5, cached at $0.05; Luna at $0.10 and $0.50, cached at $0.01. Terra therefore sits at a tenth of Astra, half of Sol, and ten times Luna on every line item.

OpenAI says the new prices are 50% lower than both GPT-5.6 Terra and GPT-6.1 Sol. The model it replaces was listed at $2 in and $12 out per million tokens, according to Artificial Analysis, so the cut is half on input and better than half on output. Cached input collapsing from $0.10 to $0.05 matters more than the headline for anyone running agents that re-read the same context across dozens of turns.

This is the fourth pricing move on OpenAI's mid-tier in about six weeks. We covered the last one when OpenAI halved GPT-6 Sol and Luna prices hours after Anthropic's cut — and the pattern has not changed: every cut lands on the models that carry agent traffic, not on the flagship.

What the benchmarks claim

The headline evaluation is Agents' Last Exam, a set of professional computer-use tasks that includes turning architectural drawings into a 3D model and reviewing an animated film shot by shot to build a time-coded log for a restoration team. At medium reasoning effort, OpenAI says Terra scores 15 percentage points above GPT-5.6 Terra at about one-third the cost, and performs on par with GPT-6 Astra at one-tenth its price per token. It also finishes tasks more than three times faster than both Astra and Sol.

On MMMU-Pro, a visual-reasoning benchmark, OpenAI reports Terra beating GPT-5.6 Terra at every reasoning setting and nearly matching Astra and Sol. On OSWorld 2.1, which tests whether agents can complete multistep desktop work, Terra at medium effort surpasses GPT-5.6 Terra's best score at a fraction of the cost per task.

Coding is where the claims get interesting rather than triumphant. On Terminal-Bench 4.0, which tests command-line problem solving, Terra solves 50.3% of tasks at maximum reasoning effort — more than double the 21.5% OpenAI published for GPT-5.6 Terra. OpenAI is explicit that Sol and Astra remain the better choice for the most complex coding work, which is an unusual thing for a launch post to concede and a reason to take the rest of the numbers more seriously.

The factuality numbers are the most consequential for buyers. On OpenAI's internal set of error-inducing prompts, Terra's error rate at low reasoning effort falls from 53.6% with GPT-5.6 Terra to 16.7%, a drop of nearly 37 percentage points; at maximum effort it falls 18 points to 4.3%, close to Astra and Sol. Low effort is where cheap high-volume automation actually runs, so that is the number that decides whether a document-processing pipeline is viable.

On Terminal-Bench-Science 0.1, which covers research workflows like tracking cell divisions in a time-lapse video, OpenAI says Terra at max reasoning surpasses GPT-6.1 Sol at lower cost and comes within five points of Claude Opus 5.5 at less than one-fifth the estimated cost. For scale, the Sol launch post priced an average max-effort task on that benchmark at $5.47 for Sol, against $23.21 for Opus 5.5 and $23.80 for Astra. Against Opus 5.5 specifically, OpenAI claims Terra at max effort is over three times as token-efficient and a quarter of the per-token price, for roughly nine times lower cost per task.

The safety footnote nobody will read

Buried in the middle of the post: GPT-6.1 Terra meets the Critical threshold in cybersecurity under OpenAI's Preparedness Framework, and carries over the protections built for GPT-6 Astra. The full detail sits in a system card addendum published alongside the model.

That sentence deserves more attention than it will get. Critical is the band OpenAI's own framework treats as requiring the heaviest deployment scrutiny — the level at which a model is presumed capable enough to matter to an attacker. Until this year that label was reserved for flagships; now it belongs to the cheap tier. The capability did not get safer, it got affordable. Cheap access to competent computer use is mostly a good thing for people automating invoices and bad-intent-adjacent tasks; the governance assumption that dangerous capability lives behind the expensive model is what just broke.

The problem with a four-tier line

The tiers now overlap in ways that make the price sheet a poor guide to the product. Terra matches Astra on one benchmark at a tenth the per-token price, and beats Sol outright on Terminal-Bench-Science at max reasoning. A cheaper model inside the same family beating the more expensive one is good for buyers and confusing for anyone trying to pick a default.

Dan Elitzer, an investor at Delphi, put the buyer's version of that on X: "I'm so confused by OpenAI's model release strategy. We should all just start hammering Terra at Max and ignoring Sol and Astra for now?" It reads as a joke with a purchasing decision inside it.

The trap is treating per-token price as per-task price. OpenAI's own numbers show why the two diverge: a model that reasons longer can cost more per task at a lower rate per token, and a model that is three times faster can cost less per task even at the same rate. The honest comparison is dollars per completed task at the reasoning effort you would actually run, measured on your own work — not the price sheet, and not a vendor's aggregate.

The counter-case

Every performance figure above is OpenAI's, run in OpenAI's research environment or through its API, with competitor scores taken from the competitors' own published reports. "On par with Astra" holds on one benchmark at medium reasoning effort. Astra still holds the highest Terminal-Bench-Science score OpenAI has reported at 68.1%, and OpenAI itself routes the hardest coding to Sol and Astra. Independent cost-per-task evaluations of Terra have not been published yet, and the gap between a vendor benchmark and a production pipeline is usually where the savings go to die.

What to watch: whether Terra cannibalizes Sol, since a half-price model that lands within a point or two will pull volume no matter what the positioning deck says; whether Anthropic answers on Sonnet 5.5 pricing, which now looks expensive next to Terra at the same nominal tier; and the first independent cost-per-task run on real agent workloads.

Would you move a production agent onto a mid-tier model this week, or wait for independent cost-per-task numbers? Write to us via our contact page.