Half the Price Per Token Is Not Half the Cost Per Task

On 26 October 2026, OpenAI released GPT-6.1 Terra and led with a number that sounds like a gift: standard API prices 50 percent lower than both GPT-5.6 Terra and GPT-6.1 Sol. One week later, an open-source project called NorthCUA shipped a 397-billion-parameter computer-use agent that it says beats trillion-parameter closed models on the two benchmarks the field trusts.

Both claims are real. Both are also, in a specific and instructive way, not the claim a buyer needs.

The computer-use category just had its most consequential fortnight of the year. It also produced, in the same eight days, a near-perfect demonstration of how benchmark numbers mislead once they leave the vendor's chart.

What actually shipped

GPT-6.1 Terra is a mid-tier model built for visual reasoning and precise computer use, positioned between GPT-6 Luna and the Sol and Astra flagships. Standard pricing is $1 per million input tokens, $5 per million output, and $0.05 per million cached input.^[1]

The capability claims are specific. On Agents' Last Exam, a benchmark of professional tasks, OpenAI reports Terra at medium reasoning effort scoring 15 percentage points above GPT-5.6 Terra at about one-third the cost. At max effort it performs on par with GPT-6 Astra at roughly one-tenth the price per token, and completes tasks more than three times faster than both Astra and Sol.^[1]

There is one headline worth isolating because it is unusual honesty in a launch post. On OpenAI's internal benchmark of error-inducing prompts, the error rate at low reasoning effort fell from 53.6 percent with GPT-5.6 Terra to 16.7 percent with the new model. At maximum effort it fell 18 points to 4.3 percent.^[1] That is a hallucination reduction, not a capability gain, and for an agent that clicks things on a screen it matters more than a leaderboard point.

NorthCUA is the other half of the fortnight. Released on 3 November, it is a fully open suite built on Qwen3.5 in three sizes, shipping weights, the Harness execution framework and the training pipeline. The training story is the interesting engineering: an Experience Flywheel generates executable tasks, a Harness sustains long interactions and checks outcomes, and Staged Learning moves from grounding mid-training through long-horizon supervised fine-tuning into joint desktop and web reinforcement learning. The report describes more than four million GUI interaction steps collected across dozens of applications in isolated sandboxes and filtered by outcome.^[2]

Terra against its own family

The first table anyone should build is not Terra versus competitors. It is Terra versus the models it sits between, at the same setting, with cost attached.

Read the top half and the launch story is entirely true. Terra beats its immediate predecessor by 17 points on desktop work at roughly a quarter of the cost per task, and by 9 points on professional tasks at under a third. That is a genuine generational improvement, not a marketing claim.^[1][4]

Read one line further and a different fact appears.

Now the part that should change how you read all of it

Sol at max reaches 78.0 percent for $1.15 per task. Terra at max reaches 74.2 percent for $1.19.^[5]

Sol is both more accurate and cheaper. The model with the higher list price wins on the only two numbers a buyer cares about, because it gets there in fewer steps. Terra needs its maximum reasoning effort to reach a score that Sol achieves for less.

That is not a knock on Terra. It is a demonstration that "half the price per token" and "cheaper to finish the job" are different sentences, and that a launch post will always quote the first because it is the one that sounds like a discount.

The metric nobody states

Before comparing any two computer-use scores, establish which metric they are. On the same benchmark release, from the same run, one model produced these two figures.

Thirty-three points apart. Same weights, same tasks, same run.^[3]

The benchmark paper's primary metric is binary. Vendors lead with partial, because partial is the larger number and produces a friendlier chart. Neither figure is dishonest on its own. Presenting them without the label is.

That single distinction accounts for more confusion in this category than any other, and it costs one sentence to resolve.

Three ways the numbers stop being comparable

The benchmarks are not the same benchmark. The two live versions measure different work at different scales.

A score near 85 percent is almost certainly Verified. A score near 80 percent partial on 2.0 is a far harder result, because each task requires roughly ten times the tool calls. Placing both in one column is a category error, and it happens constantly.^[3]

The releases are not comparable either. Between the August 2026 task set and v2.1, fourteen tasks were updated to prevent reward hacking and thirty-one were modified for more robust scoring. One model's binary completion moved from 31.43 to 44.33 percent across that change with no alteration to the model at all.^[3] The test moved, not the agent.

Effort level is a hidden multiplier. Same weights, same benchmark, more than double the score.

A vendor can publish its best configuration and a buyer can deploy its cheapest, and both statements will be true while describing entirely different products. A score without a stated effort level is a brochure, not a specification.^[3]

The cost illusion, worked through

This is the section I would put in front of anyone approving a computer-use budget.

Click to expand

Take six configurations of two models from OpenAI's own chart, and sort them by what they actually cost to complete a task.

Now trace the frontier rather than the rows. Every configuration below about $0.65 is a bargain you accept with a capability ceiling. Between $0.65 and $1.19, Sol dominates Terra completely: more accurate and cheaper at both ends of that range. Above $2, you are buying certainty, and Sonnet 5.5 delivers the highest score on the chart.

Three practical conclusions follow.

Terra's cheapest configuration is its most interesting one. At medium effort it costs $0.50 against Sol's $0.65 while giving up 6.8 points. If your failures are cheap to retry, that is a reasonable trade and the one OpenAI should be selling.

Terra's most expensive configuration is its weakest. At max effort it is beaten on both axes by a model with twice the list price. Anyone standardising on "max effort for quality" would have deployed the wrong model.

Above roughly $36 per failure, accuracy stops being a luxury. That is the crossover where the slower, more expensive option earns its price, and it is a business fact rather than a benchmark fact. A botched invoice, a misfiled form, a corrupted model file: price each of those once and the choice stops being close.^[5]

What the open-source result actually means

NorthCUA matters, and its practical significance is narrower than the headline.

If a 397B open model genuinely matches trillion-parameter closed systems, three things follow. Computer-use capability stops being a rented commodity and becomes something you can host. The per-screenshot token economics of cloud vendors become contestable. And QA teams get a plausible path from brittle selector scripts to agents that adapt when a UI changes.

Two caveats belong in the same paragraph. The technical report states its leadership claims as comparative verdicts rather than absolute scores in the abstract, so anyone building a comparison table needs the full paper rather than the announcement. And at least one independent write-up noted, at publication, that public weights and a working demo had not been verified.^[2][6] Benchmark parity is a claim until you reproduce it.

Worth noting too: a 397B model is open in licence and not small in practice. The 9B Nano in the same release is the piece most teams will actually run on their own hardware.

How to read the next launch

Four questions, in order. They take ten minutes and will save you a wrong decision.

Which benchmark, and which release of it. If the answer is not stated, the number is not usable.

Partial or binary. This is the single highest-value question, because it accounts for thirty-three points of apparent difference.

Which effort or reasoning setting. A score without a setting is a brochure.

Cost per completed task at that same setting, including retries. Not price per token. The two figures describe different businesses.

If a vendor cannot answer all four, that is itself information.

Why this matters beyond the comparison charts

Computer-use agents are the category where benchmark literacy has the most direct financial consequence, because these agents touch real systems. The difference between a 44 percent and a 78 percent completion rate is not a marketing distinction. It is the difference between an agent that needs supervision on every second workflow and one that needs it on every fourth.

That framing also clarifies what to build around. A measured success rate predicts how much human review you must staff. A measured cost per task predicts your unit economics. Neither is the number in the launch post.

OpenAI's most useful disclosure in the whole Terra announcement was the hallucination drop, from 53.6 percent to 16.7 percent at low effort, because an error rate in an agent that acts on a screen is not an abstract quality metric. It is a count of things you will have to undo.

The next time a launch tells you a model is half the price, ask what a finished task costs. The answer is usually in the same chart, a few lines down, in the configuration nobody screenshotted.