GPT-6.1 Terra: A Dollar Model That Uses Your Desktop

Posted on by
Contents show
GPT-6.1 Terra

GPT-6.1 Terra costs one dollar per million input tokens, five cents per million cached input tokens, and five dollars per million output tokens. Those are standard API prices, not a promotional window, and they are half what OpenAI charges for GPT-5.6 Terra and for its own GPT-6.1 Sol. On the benchmark that measures whether a model can operate a desktop, Terra scores fifteen percentage points above its predecessor at roughly one third the cost per task, and matches GPT-6 Astra while costing one tenth as much per token.

The pricing is the story because of what it now buys. Two years ago a model at this price could summarize documents. Today it can read an architectural drawing, build a 3D model from it, review an animated film shot by shot, produce a time-coded restoration log, fill out forms, and enter invoice details into accounting software, at more than three times the speed of the frontier models that were state of the art last quarter.

This article covers what GPT-6.1 Terra actually does well, the benchmark results that support the claims, the factuality improvement that matters more than any capability score, where the model should not be trusted, the unusual safety classification attached to a budget model, and how teams should route work between Terra and its more expensive siblings.

What a dollar per million tokens now includes

The GPT-6.1 Terra specification reads like a mid-tier model and the benchmark table reads like a frontier one, and reconciling the two is the point of the release.

GPT-6.1 Terra is positioned as a fast, efficient model combining detailed visual reasoning, precise computer use, and clearer, more natural writing. The tasks OpenAI names are the ones that actually fill a workday: drafting proposals, interpreting technical drawings, filling out forms, editing videos. None of these are research problems; all of them are the administrative sediment of professional work, and all of them have resisted automation because they require looking at a screen and understanding what is on it.

The speed claim is as consequential as the price. OpenAI reports GPT-6.1 Terra completing tasks more than three times faster than both GPT-6 Astra and GPT-6.1 Sol. For an agent operating a desktop, latency is not a comfort feature; it is the difference between a workflow that replaces five minutes of human clicking and one that takes longer than doing it manually. A model that is three times faster at one tenth the token cost changes the economics of automation from marginal to obvious.

Availability is broad from day one: Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex, plus the API under the identifier gpt-6.1-terra. That is not a research preview or a waitlist; it is a shipping product, which tells you OpenAI expects volume rather than evaluation.

Desktop work is the hard part

Reading a screenshot is easy, and GPT-6.1 Terra is the first budget model that reliably gets past it. Operating an application is not, and the benchmarks OpenAI chose reflect that distinction.

Agents' Last Exam measures professional tasks end to end, including turning architectural drawings into a 3D model and reviewing an animated film shot by shot to produce a time-coded log for a restoration team. These are tasks with deliverables, not questions with answers, and they require the model to hold a goal across dozens of interface interactions while noticing when the screen stops matching its expectation.

At medium reasoning effort, GPT-6.1 Terra scores fifteen percentage points higher than GPT-5.6 Terra at about one third the cost. It performs on par with GPT-6 Astra at one tenth the price per token. At maximum effort it is more than three times as token efficient as Claude Opus 5.5 and one fourth the price per token, producing roughly nine times lower cost per completed task.

OSWorld 2.1 tests the same capability more mechanically, measuring how well an agent uses desktop applications to complete multistep jobs such as building a 3D model from a technical drawing. At medium effort, GPT-6.1 Terra surpasses the previous model's best score at a fraction of the cost per task, which is the cleanest statement of the generational jump: the cheap configuration now beats the expensive configuration of one generation ago.

MMMU-Pro, the visual reasoning benchmark, shows GPT-6.1 Terra outperforming GPT-5.6 Terra at every reasoning setting and nearly matching Astra and Sol at much lower cost. The pattern across all three benchmarks is consistent, and it is the pattern that matters for adoption: capability that was frontier six months ago is now commodity, and the price fell faster than the capability rose.

The factuality number that matters most

The most important figure in the GPT-6.1 Terra release is not a capability score. It is an error rate that fell from 53.6 percent to 16.7 percent.

On OpenAI's internal benchmark of error-inducing prompts, low reasoning effort with GPT-5.6 Terra produced errors in more than half of cases. GPT-6.1 Terra produces them in about one sixth, a drop of nearly thirty-seven percentage points. At maximum effort the error rate falls by eighteen points to 4.3 percent, nearly matching the frontier models.

That measurement deserves emphasis because error rates, not capability ceilings, are what block enterprise deployment of cheap models. A budget model that hallucinates in half its responses is unusable for invoice entry, regardless of how fast it is, because every output needs verification and verification costs more than the labor saved. A model that errs in one sixth of cases crosses the threshold where spot-checking beats manual entry, and at maximum effort the 4.3 percent rate puts it in territory where the model's output can be trusted for routine structured work.

The mechanism is worth noting for anyone evaluating similar claims. Error-inducing prompts are designed to tempt a model into fabricating rather than admitting uncertainty, so a large improvement implies better calibration about what it knows. That connects directly to the alignment claims in the same release: GPT-6.1 Terra is described as more likely to stay within scope and to disclose limitations and blockers honestly, which is the behavioral counterpart to the factuality metric.

For high-volume automation, this is the number that determines whether a workflow gets approved. A finance team can build a business case around a model that reads invoices correctly five times out of six with human review, and cannot build one around a coin flip.

Where GPT-6.1 Terra should not be used

A cheap fast GPT-6.1 Terra model invites overextension, and OpenAI's own framing draws the boundary explicitly enough that buyers should respect it.

The coding results are strong but not frontier. Terminal-Bench 4.0, which tests command-line problem solving such as fixing a webpage whose layout shifts as it loads, shows GPT-6.1 Terra solving 50.3 percent of tasks at maximum reasoning effort, more than double GPT-5.6 Terra's published 21.5 percent. That is a genuine leap for a budget model. It is also a score that leaves half the benchmark unsolved, and OpenAI states plainly that for the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.

The distinction is architectural rather than promotional. GPT-6.1 Terra is optimized for breadth at cost, which means it handles the long tail of ordinary work well and the narrow head of extremely hard work less well. A team that routes everything to Terra to save money will discover the boundary in production, on the one task category where the model is weakest and the consequences are largest.

Terminal-Bench-Science 0.1 offers the more interesting read. GPT-6.1 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 scientific coding workflows that are routine but technical, that combination is close to ideal, and it suggests the model's visual and procedural strengths transfer into lab automation more than the raw coding score implies.

The practical routing rule is to send structured, repetitive, screen-mediated work to GPT-6.1 Terra and to reserve the frontier models for novel architecture, ambiguous debugging, and anything where a wrong answer costs more than the token savings. That rule is easy to state and surprisingly hard to enforce without instrumentation, which is why routing layers have become their own product category.

The invoice test: reading the cost claim correctly

A worked example makes the pricing argument concrete, and it also shows where the savings actually come from.

Consider an accounts payable workflow reading 10,000 invoices a month. Each invoice runs roughly 3,000 input tokens of document image and instructions and produces 400 output tokens of extracted fields and confidence notes. At GPT-6.1 Terra rates that is 30 million input tokens at one dollar per million, or thirty dollars, plus 4 million output tokens at five dollars per million, or twenty dollars. Fifty dollars of inference for ten thousand invoices, before caching, which at five cents per million for cached instructions cuts the input line to a fraction.

The same workload on a frontier model at ten times the per-token price runs five hundred dollars, and on a model priced at nine times the cost per task, four hundred fifty. The comparison that matters is not the invoice total but the threshold: at fifty dollars, the workflow costs less than a single hour of the clerk time it displaces, which means the automation decision no longer requires a business case meeting. At five hundred dollars it still wins, but it needs approval, a pilot, and a procurement cycle.

The second half of the arithmetic is verification. If the model errs in one sixth of cases at low effort, a review sample of a few hundred invoices catches systematic problems at a cost far below full manual processing. At a 4.3 percent error rate at maximum effort, the review can be sampled more lightly, and the model's own confidence notes tell the reviewer where to look. That is the design pattern GPT-6.1 Terra makes affordable: cheap generation plus targeted verification, rather than expensive generation trusted absolutely.

Teams running this calculation should use their own token counts rather than vendor examples, because document-heavy workflows vary by inputs, and the ratio of input to output tokens drives the blended cost more than the headline rate does.

A Critical cyber rating on a budget model

The GPT-6.1 Terra safety classification is the detail most likely to be overlooked, and it deserves attention precisely because of the price.

GPT-6.1 Terra meets the Critical threshold in cybersecurity under OpenAI's Preparedness Framework. That is the same classification assigned to the company's dedicated security model released a week later, and it means a model available to any Plus subscriber at a dollar per million input tokens carries capability that OpenAI itself treats as requiring strong safeguards during development and before release.

The company states that GPT-6.1 Terra maintains the protections developed for GPT-6 Astra, and that it shows strong resistance to prompt injection in internal tests. Both claims matter for desktop agents in particular, because an agent that reads email, web pages, and documents is an agent that ingests untrusted instructions constantly, and a cheap model deployed at volume multiplies whatever weakness it has by the number of sessions running.

The policy question this raises is not hypothetical. When frontier cyber capability becomes commodity priced, the gating strategy that works for a specialized model, vetting customers and limiting access, cannot apply to a product sold to millions. The substitute is layered defense inside the model and the platform, which is what OpenAI describes, and the honest assessment is that this shifts the burden from access control to detection and monitoring at a scale nobody has operated before.

For enterprise buyers, the practical implication is that a cheap model still requires the same governance as an expensive one. Logging, scoped credentials, sandboxing, and human approval for irreversible actions are not optional because the price is low, and the cost of those controls now exceeds the cost of inference in most deployments, which is the inversion that defines this phase of the market.

What this does to the agent economy

A GPT-6.1 Terra with these economics changes which automations are worth building, and the change is larger than a price cut.

The threshold for automating a task is roughly that the model plus oversight costs less than the labor it displaces. At five dollars per million output tokens with three times the speed of last quarter's frontier, a large category of tasks that failed that test now passes: invoice data entry, form filling, video logging, drawing interpretation, first-pass document drafting, and the long tail of screen-mediated administrative work that consumes professional hours without producing professional value.

The second-order effect is on agent architecture. When inference is cheap relative to coordination, it becomes rational to run verification passes, redundant checks, and multiple attempts, because the marginal cost of a second opinion is pennies. Architectures that were uneconomical at frontier prices, where every retry was a budget line, become standard practice at GPT-6.1 Terra prices, and quality improves through redundancy rather than through a better single model.

The third effect is competitive. OpenAI's framing of GPT-6.1 Terra as advancing the cost-intelligence frontier across its model families is a statement about a price war, and every competitor now has to answer at this price point or explain why not. The open-weight tier already undercuts it, and the closed frontier tier cannot match it on cost, which leaves the middle of the market, models that are neither cheapest nor best, in the most exposed position.

For readers tracking how these economics are reshaping infrastructure and tooling, the pricing cliff on Google's competing frontier model is detailed in our Gemini 4 Argon cost analysis, the security model that shares GPT-6.1 Terra's Critical classification is examined in our GPT-6 Cyber coverage, and the routing layer that decides which model sees which request is unpacked in our AI gateway comparison.

What to watch over the next quarter

Four developments will determine whether GPT-6.1 Terra's economics hold, and each is observable within months.

The first is whether competitors match the rate. A dollar per million input tokens with three times the speed of last quarter's frontier sets a reference point that every other vendor must now answer, and the shape of those answers, matching prices, bundled allowances, or capability arguments, will reveal how much margin the frontier tier actually carries.

The second is whether the error rate holds in production. Benchmark improvements on error-inducing prompts are encouraging, but real invoices, real forms, and real drawings contain messier ambiguities than any evaluation set, and the first credible third-party audit of GPT-6.1 Terra's factual accuracy on production workloads will be more informative than the launch numbers.

The third is how quickly the Critical cyber classification becomes a procurement question. Enterprises that would never deploy an ungoverned frontier model may deploy a cheap one casually, and the gap between the classification and the controls actually in place is where incidents originate.

The fourth is routing maturity. The value of a model tier only materializes if requests reach the right tier, and the tooling that decides which request goes where is still immature. Teams that build routing discipline now will capture the savings the price list promises; teams that standardize on one model will either overpay or underperform, depending on which one they chose.

For anyone planning 2027 budgets, the honest forecast is that these prices keep falling and the capability tier keeps rising, so the safe assumption is not that GPT-6.1 Terra is cheap but that the price of last quarter's frontier capability approaches zero on a predictable curve.

Questions readers are asking

How much does GPT-6.1 Terra cost?

One dollar per million input tokens, five cents per million cached input tokens, and five dollars per million output tokens at standard API rates, with no expiry date.

Who can use it today?

Plus, Pro, Enterprise, and Edu subscribers through ChatGPT Work and Codex, and developers through the API under the model identifier gpt-6.1-terra.

Is GPT-6.1 Terra better than GPT-6 Astra?

It matches Astra on some visual and computer-use benchmarks at a fraction of the cost and is more than three times faster, but Astra and GPT-6.1 Sol remain stronger on the hardest coding and reasoning tasks.

What is the catch?

Half the Terminal-Bench coding tasks remain unsolved at maximum effort, error rates rise at lower reasoning settings, and the model carries a Critical cybersecurity classification that demands the same governance as a frontier deployment.

Sources
Subscription Form