GPT-6.1 Terra for Browser Agents: The Per-Step Math Behind 74.2%

OpenAI’s pitch for GPT-6.1 Terra is not that it is smarter. It is that it can operate a screen — fill a form, read a technical drawing, click through an application — and do it cheaply and quickly. The benchmark that supports the claim is OSWorld 2.1, where Terra scores 74.2% at maximum reasoning effort, within 1.7 points of GPT-6.1 Sol at xhigh and 1.6 points of GPT-6 Astra. The number that matters for an agent, though, is not the task score. It is the per-step reliability implied by that score, because agents fail step by step, not task by task.

This piece takes the 74.2% apart: what OSWorld measures, what it implies per step, what a twenty-step task costs, where the loop breaks, and the guardrails that keep a 74%-accurate agent from becoming a 0%-accurate incident.

What the scoreboard says

• GPT-6.1 Terra — 74.2% on OSWorld 2.1 at maximum effort.

• GPT-6.1 Sol — 75.9% at xhigh.

• GPT-6 Astra — 75.8%.

• Claude Opus 5.5 — 81.8%.

• Claude Sonnet 5.5 — 80.1%.

• Kimi K3.1 Pro — 79.6%.

Three readings are worth separating. Within the GPT-6 family, Terra is a rounding error away from its more expensive siblings — the price gap buys almost nothing on desktop automation. Across vendors, the whole family trails Anthropic by six to seven points. And Terra’s own vendor-adjacent claim — roughly level with GPT-6 Astra at medium on Agents’ Last Exam, at one-tenth the per-token price — is a per-task claim that the OSWorld number largely supports.

From task score to per-step reliability

OSWorld tasks are scored pass or fail after a sequence of actions. If a task takes roughly twenty steps and Terra completes 74.2% of tasks end to end, the implied per-step success rate is the twentieth root of 0.742 — about 98.5%. That is an arithmetic illustration, not a measurement: some tasks are shorter, some are longer, and failures are not independent. But it sets the scale correctly. A model that looks like a 74% model is really a 98.5%-per-step model, and that gap is where agent design lives.

Run the math the other way for a fifty-step workflow. At the same per-step rate, end-to-end success falls to about 47%. At eighty steps it is 30%. The task score does not change; the workflow does, and that is why “good enough for an agent” depends entirely on how long the agent runs unsupervised.

What a task costs

Assume 3,000 input tokens per step — a screenshot plus task state plus action history — and 150 output tokens for the action, over twenty steps.

• Per task: 60,000 input at $1 per million is $0.06; 3,000 output at $5 per million is $0.015. Total $0.075.

• With a 30% retry rate: about $0.0975 per successful task, or $98 per thousand tasks.

• The same workload on GPT-6.1 Sol at $2/$10: $0.15 per task, $195 per thousand.

• On GPT-6 Astra at $10/$50: $0.75 per task, $975 per thousand.

• On Claude Opus 5.5 at $4/$20: $0.30 per task, $390 per thousand.

The token assumptions matter more than the price ratio. Screenshots dominate input cost, and screenshot size is a choice: a full 1440×900 PNG costs far more tokens than a cropped region of interest. A pipeline that sends one cropped screenshot per step instead of the whole desktop can cut input tokens by half, which halves the per-task cost and changes the ranking against the more accurate models.

Latency is the product

Artificial Analysis measures Terra at 90.9 output tokens per second and 133.64 seconds to first token on its reasoning benchmark, against 73.5 tokens per second and 284.31 seconds for GPT-6.1 Sol. Those latency figures come from a max-effort reasoning task, so they overstate per-step delay in an agent loop — but the ratio, a little over 2×, is the right indicator of how long each model thinks before acting.

In an agent loop, thinking time compounds twenty times per task. If Sol takes four seconds per step and Terra two, a twenty-step task finishes in 40 seconds instead of 80. OpenAI staffer Victor E. Nunez reported that Terra completed a web form more than two minutes faster than Sol; that is an anecdote, not a benchmark, but it is the kind of difference users notice without being told.

Where the loop breaks

• Irreversible actions. Submit, send, delete, pay — any action with a side effect needs an approval gate. No benchmark score changes that.

• Ambiguous interfaces. Two buttons labelled “Continue” defeat a model at any price; the fix is prompt and screenshot hygiene.

• Long horizons. The per-step math above is the whole argument against unsupervised long-horizon work at 98.5% per step.

• Stale screenshots. A frame captured before a layout shift is wrong, and a wrong frame produces a confident wrong click.

• State that lives across tabs. Screenshot-based agents see one window at a time; multi-tab workflows need explicit state tracking outside the model.

Guardrails that make 74% shippable

• Cap steps, not just dollars. Twenty is a budget and a safety control at once; a loop that cannot terminate is the most expensive failure mode.

• Gate side effects. Route any action that changes external state through a human or a policy check.

• Log the frame with the action. When the agent does something wrong, the only useful artefact is the screenshot it acted on.

• Use structured outputs for the action schema. A free-text action is unparseable at exactly the moment you need certainty.

• Measure per step, not per task. A 2% per-step regression is invisible in task scores and doubles the failure rate over fifty steps.

• Keep a second model on the hard steps. Terra for routine frames, a stronger model for the ambiguous ones, is cheaper than Terra with retries.

What to run today

GPT-6.1 Terra is not on the OrcaRouter catalogue at the time of writing, so the Terra arm has to be trialled against OpenAI directly. The comparison set is live on our side at 0% markup over provider list price — GPT-6.1 Sol at $2/$10, GPT-6 Astra at $10/$50, Claude Opus 5.5 at $4/$20, Claude Sonnet 5.5 at $2/$10 and Kimi K3.1 Pro at $2/$10 — all on one key with automatic failover.

The experiment worth running is not “which model scores highest on OSWorld”. It is your own task set, scored per step, with the screenshot policy held constant. That measurement decides whether Terra’s price and speed are worth the 7.6-point gap to Claude Opus 5.5 on the tasks you actually run — and it is the only number an agent product can be built on.