GPT-6.1 Terra Ties GPT-6 Astra on Agent Work at 1/17 the Cost. On the Desktop, Sol Still Wins.

What OpenAI shipped

GPT-6.1 Terra launched on October 26. OpenAI’s model page describes it as “a fast, efficient model for visual reasoning, precise computer use, and clear, natural writing”, with examples like acting on application screens and reading technical drawings. It takes text and images, writes text, and supports the computer use tool through the Responses API.

It is rolling out in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu users, and it is in the API as gpt-6.1-terra at $1 per million input tokens and $5 per million output. That is half of GPT-6.1 Sol ($2 / $10) and a tenth of GPT-6 Astra ($10 / $50).

We already covered the coding side in Terra versus Sol on Terminal-Bench 4.0, where Terra read 5.5 times the tokens and ended up costing more per task than Sol. This post is about the job OpenAI actually built Terra for: operating software.

Agents’ Last Exam: Astra’s score for 6% of the cost

Agents’ Last Exam is a benchmark of professional computer work done by an agent. OpenAI’s launch chart plots each model’s score against its estimated cost per task, at several reasoning efforts.

Source: OpenAI’s Agents’ Last Exam v1 chart from the GPT-6.1 Terra launch, as transcribed by CellCog. OpenAI ran every model, including Claude Opus 5.5, and estimated the costs.

Three things stand out. Terra at medium ties Astra at medium for $0.29 against $4.82, about 1/17 of the cost. Terra beats Sol at every effort level: even Terra at low effort (55.9%) is ahead of Sol at medium (53.8%). And Claude Opus 5.5 still has the top score, 3.2 points above Terra at max, for about 8.7 times the cost per task.

Divide the score by the cost and the gap gets bigger. With $10 you can run about 34 Terra tasks at medium effort and expect roughly 20 to succeed. The same $10 buys about one successful Astra or Opus task.

OSWorld 2.1: on the desktop, Sol still wins

OSWorld is the standard computer-use benchmark: the model sees screenshots of a real desktop and has to finish tasks by clicking and typing in ordinary apps. OpenAI reported OSWorld 2.1 on its offline task set.

Source: OpenAI’s OSWorld 2.1 chart from the GPT-6.1 Terra launch, as transcribed by Kingy AI; BenchLM lists the same 74.2% for Terra. Costs are shown only where the transcription gives them.

Here the half-price token rate doesn’t turn into a cheaper task. At max effort Terra costs $1.19 per task and Sol $1.15, so Terra is 3.5% more expensive and 3.8 points lower. Per $10, that is about 6.2 successful desktop tasks for Terra and 6.8 for Sol. At half the token price and a slightly higher cost per task, Terra must be reading roughly twice the tokens Sol does, by our math, the same pattern we saw on Terminal-Bench.

The jump over the old Terra is real, though. Terra at medium effort (66.7%) clears GPT-5.6 Terra at max (57.1%) by 9.6 points, which matches the claim in OpenAI’s launch text. If you run GPT-5.6 Terra in a computer-use agent today, switching is an easy win.

One more data point: the top of this chart isn’t OpenAI at all. Claude Sonnet 5.5 tops the OSWorld 2.1 chart OpenAI published, at 83.9%, 9.7 points above Terra.

Effort is the dial that sets your bill

Terra has six reasoning efforts, from none to max, with medium as the default. On Agents’ Last Exam, going from low to max adds 4.1 points and multiplies the cost per task by 6.3. Going from medium to max adds 2.4 points for 3.5 times the cost.

For a computer-use agent that clicks through forms all day, medium is the sweet spot on these numbers. Save max for the steps that fail at medium, not for every screenshot.

The fine print that hits computer-use loops

A computer-use agent re-sends its history, screenshots included, on every step. That makes two lines on Terra’s price card matter more than the headline rate.

Source: OpenAI’s GPT-6.1 Terra model page. Batch and Flex are 50% off, Fast mode is 2x, regional processing adds 10%. Cache writes cost $1.25 (30-minute retention) or $2.50 (12-hour retention) per million tokens.

The 272K line. OpenAI’s page says prompts with more than 272K input tokens are “priced at 2x input and cache rates and 1.5x output for the full request.” The whole request, not just the overflow. A long session that keeps every old screenshot will cross it. Trim or summarize old steps and you stay on the base rate.

The cache. Cached input is $0.05 per million, 5% of the normal rate. An agent whose system prompt and early steps stay identical gets most of its input at that price. Change the start of the prompt every step, by putting a timestamp first for example, and you lose it.

Prompt injection. A model reading web pages can be told what to do by those pages. OpenAI’s system card, as reported by The Decoder, puts Terra’s defense against indirect prompt injection at 98.8%, up from 96.2% for GPT-5.6 Terra. That still leaves about 1 attack in 83 getting through in OpenAI’s own test. Keep a human confirmation on anything that sends, buys or deletes.

Fun facts

• OpenAI’s lineup is a small solar system: Astra (Latin for “stars”) on top, then Sol (the Sun), Terra (the Earth) and Luna (the Moon).
• Terra in Fast mode costs $2 / $10 per million tokens, exactly the standard price of GPT-6.1 Sol.
• Terra has a 1,050,000-token context window, but only the first 272K input tokens are at the base rate. Of its 922,000-token input limit, about 70% sits in the 2x zone.
• A million cached input tokens on Terra cost a nickel.
• On Agents’ Last Exam, GPT-6 Luna at max effort ($0.14 a task) is cheaper than Terra at low effort ($0.16), and scores 5 points lower.
• Terra’s knowledge cutoff is June 1, 2026, almost five months before its launch. Its training data ends before its own name ever leaked.
• On Artificial Analysis’s knowledge test, Terra guesses instead of saying “I don’t know” 38% of the time when it lacks the answer. GPT-5.6 Terra did that 88% of the time.

What this means if you build agents

For browser and office agents, start with Terra at medium. On OpenAI’s Agents’ Last Exam numbers it matches Astra and beats Sol for a fraction of the cost per task. That is the case Terra was built for.

For pure desktop control, test Sol too. On OSWorld 2.1 at max effort, Sol scores 3.8 points higher at about the same cost per task. If your agent lives in a desktop OS, the half-price rate card is not the deciding number.

Watch the context, not just the rate. Screenshot-heavy loops grow fast. Past 272K input tokens the whole request costs 2x input and 1.5x output, which can erase Terra’s price advantage on long sessions.

Wait for independent runs. Every computer-use number in this post comes from OpenAI. The Terminal-Bench numbers held up within a point when Vals.ai ran Terra itself, so these are a reasonable starting point. They are still the vendor’s own.

Measure cost per finished task on your own workflows. Run a week of real tasks on two models and compare the bill and the pass rate. That beats any chart, including this one.

Where MegaBrain fits

Full disclosure: we sell an AI coding agent. MegaBrain’s terminal coding agent and editor extensions run GLM, Kimi, MiniMax, DeepSeek, Qwen, Grok and Auto on the paid plans, Claude, GPT and Gemini paid from your credits from Pro, and any model through your own API key or local models, free on every plan. So you can put two models on the same task and compare the bills yourself. See the plans and subscribe. MegaBrain is not affiliated with OpenAI, Anthropic or any of the sites cited here.

Sources

All pages opened October 27, 2026.

• OpenAI — GPT-6.1 Terra model page (description, computer use tool, context window, effort levels, prices, long-context rule, Batch / Flex / Fast pricing)
• OpenAI — API pricing (GPT-6.1 Sol, GPT-6 Astra, GPT-6 Luna, GPT-5.6 Terra)
• OpenAI — “Introducing GPT-6.1 Terra” launch post with the interactive benchmark charts (it returned HTTP 403 to us; the numbers below come from the transcriptions that follow)
• CellCog — “GPT-6.1 Terra: $1/$5 Price, Benchmarks vs Sol and Opus” (Agents’ Last Exam scores and cost per task read from OpenAI’s chart)
• Kingy AI — “GPT-6.1 Terra: Benchmarks, Pricing, Specs, and How It Compares” (OSWorld 2.1 offline-set scores read from OpenAI’s chart)
• BenchLM — GPT-6.1 Terra benchmark page (OSWorld 2.1 Offline 74.2%, sourced to OpenAI)
• The Decoder — “GPT-6.1 Terra halves OpenAI’s API prices and still gains ground on its larger models” (availability, Artificial Analysis scores, system card prompt-injection numbers)
• DataCamp — “GPT-6.1 Terra: Benchmarks, Pricing and Access”