The Black Cloud

AI - OpenAI launches GPT-6.1 Terra, a fast model for computer use

OpenAI has released GPT-6.1 Terra, and this time the headline is not “here is our most powerful model”. Terra is pitched as a fast, affordable model for computer use and everyday work. Given how quickly agentic tasks can burn through tokens, a useful model that costs less and finishes sooner could be more important than another leaderboard-topping model 😊

What GPT-6.1 Terra is aiming for

Terra is designed to understand what is on screen and take actions across applications, alongside visual reasoning and more natural writing. OpenAI positions it as a model for the computer-based tasks that make up a working day, from filling in forms and editing videos to interpreting technical drawings.

The launch focuses on a few practical advantages:

OpenAI’s most striking comparison is against Claude Opus 5.5 at maximum reasoning effort: it says Terra is over three times as token-efficient, costs a quarter as much per token, and reaches roughly nine times lower cost per task. That’s a vendor-reported benchmark result, not a guarantee that every real task will be nine times cheaper. Task length, reasoning settings, retries and how much quality matters can all change the final bill.

Official announcement

OpenAI introduced GPT-6.1 Terra as a faster, lower-priced model for computer use:

GPT-6.1 Terra: delivering fast and capable computer use at a lower price.

GPT-6.1 Terra understands what’s on screen, takes action across apps, writes naturally, and is a more capable model for your whole workday with the efficiency to take on more of it.

The full Introducing GPT-6.1 Terra post includes benchmark details, API prices and the system card. Sam Altman also highlighted computer use as a particular focus:

TestingCatalog summarises OpenAI’s price and speed comparisons:

Min Choi also focused on the claimed speed and price advantage:

NOTE
GPT-6.1 Terra is available in ChatGPT Work and Codex for Plus, Pro, Enterprise and Edu users. Developers can access it through the API as gpt-6.1-terra. The published standard API prices are per million tokens; cached input has a much lower rate, but overall task cost still depends on how many tokens and model turns the task takes.

Efficiency versus capability

The efficiency numbers are appealing, especially for agents that need to inspect a screen, make a series of changes and check the result. OpenAI reports that, at medium reasoning effort, Terra scores 15 percentage points higher than GPT-5.6 Terra on Agents’ Last Exam at about a third of the cost, and performs on par with GPT-6 Astra at one-tenth its price per token. On OSWorld 2.1, it surpasses GPT-5.6 Terra’s best score at medium effort.

Those are OpenAI’s own evaluations, so independent use matters. Early impressions are split: some people like the speed and price, while others feel the model gives up too much capability or simply isn’t the frontier-model release they were waiting for.

Chris highlights OpenAI’s reported Agents’ Last Exam result and notes that the approximate per-task cost is around $1 in the comparison he shared:

The Browser Use account reports a score close to GPT-6.1 Sol on its browser benchmark, at lower cost:

Tony shared a separate three-model comparison using the same “Alpine Gold” prompt:

DesignArena compared Terra with GPT-5.6 Terra by asking both to recreate Belém Tower. Its post says Terra took longer and used more tokens, yet still cost less because of the lower token price:

That is useful context, but it is still one benchmark and one set of settings. A separate hands-on review from Salio is much less impressed:

Salio also posted a more general negative reaction to the benchmark results:

Other same-prompt comparisons also point out the speed-versus-quality trade-off. J A Z I I found Terra much quicker and cheaper, but preferred Sonnet’s result for the task they tested:

Atomic.chat shared another comparison in which Terra finished much faster and cheaper, but produced a weaker game:

Mr Tactical’s Blender comparison reported a similar difference in speed and cost:

And asha’s test came down firmly on the side that the speed did not make up for the capability gap:

And Chris makes a different point about the reaction: even a genuinely useful, efficient model can disappoint people if they were expecting a much larger frontier-model launch.

More first impressions

Not everyone is focused on benchmark tables. Some of the other posts you shared highlight hands-on impressions of computer use, writing and speed.

Tak argues that another important frontier is understanding what users mean, not just following the literal wording of a prompt:

Luckey Faraday says Terra was particularly quick at navigating unfamiliar computer interfaces:

Rohit shared a video of Terra drawing Sam Altman, saying it was not sped up:

Eric Provencher, who works on Codex DX at OpenAI, praised Terra’s speed, computer use and writing:

Afzal recommends Terra alongside GPT-6.1 Sol for Plus users looking for speed and lower cost:

Ben Badejo describes using Terra for phone control through Sightless, an example of a third-party application built on OpenAI models rather than a native Terra feature:

Bebu is already looking ahead to how the next model, Luna 6.1, might perform:

My take

I like seeing OpenAI compete on speed and cost as well as raw capability. If Terra can handle common computer-use and work tasks reliably, then lower per-token prices and quicker completion could make agents practical for more than occasional experiments. The benchmark claims are impressive, but I would still test it on the sort of workflows I actually need and compare the finished output, not just the stopwatch or token bill.

It also seems clear that Terra is aimed at a different question from “what is the smartest model OpenAI can release?” It’s a fast workhorse model; whether that is the model people wanted this week is another matter entirely 🤣

<< Previous Post

|

Next Post >>

#Blog #AI #OpenAI #GPT-6.1 Terra #Computer Use #Models