Connect with us

NEWS

GPT-6.1 Terra Makes Always-On Computer Control Cheap Enough

GPT-6.1 Terra halves OpenAI mid-tier API rates so computer-use agents can run all day, while extra steps still set the bill.

Published

on

OpenAI launched GPT-6.1 Terra on October 26 at $1 per million input tokens, half the GPT-6.1 Sol rate, for models that click through software. The company put the mid-tier model in ChatGPT Work, Codex, and the API under the id gpt-6.1-terra, and told developers it is for high-volume screen work, not flagship research.

The cut is large enough that an agent can sit on a queue of invoices, forms, or browser chores and keep running. That is a different product from a smarter chat box, and it leaves the hard jobs, and a lot of the safety load, exactly where they were.

OpenAI Cut the Computer-Use Rate in Half

OpenAI described GPT-6.1 Terra as a fast model for visual reasoning, precise computer use, and clearer writing, and as a step up from GPT-5.6 Terra, the last Terra-class model. The launch examples are workday chores: drafting proposals, reading technical drawings, filling forms, and editing video. The invoice example is the tell. The company said 50% lower standard API prices versus both GPT-5.6 Terra and GPT-6.1 Sol, “making it practical for developers to automate high-volume work, like reading incoming invoices and entering their details into accounting software.”

That sentence is about leaving the model on. Terra stayed on the GPT-5.6 generation when Sol and Luna moved to GPT-6.1, so this drop also fills a hole in the family rather than adding a fifth flagship. OpenAI still points people to GPT-6.1 Sol and GPT-6 Astra for in-depth research and complex coding. The new model is the one you can afford to run again after it already ran this morning.

OpenAI posted the launch, with a short demo of the model reading a screen and acting on it, from its account on October 26.

$1 Input, $5 Output, Half of Sol

Standard API pricing is $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens, matching the public $1 per million input tokens line on the model card. Cache writes are $1.25 per million tokens for 30-minute retention. Those list rates are exactly half of GPT-6.1 Sol ($2 input, $10 output, $0.10 cached) and one-tenth of GPT-6 Astra ($10 input, $50 output, $1 cached). Claude Opus 5.5 still lists at $4 input and $20 output, four times Terra’s token rates.

STANDARD API RATES PER MILLION TOKENS

Model Input Output Cached input
GPT-6.1 Terra $1 $5 $0.05
GPT-6.1 Sol $2 $10 $0.10
GPT-6 Astra $10 $50 $1.00
Claude Opus 5.5 $4 $20 n/a
GPT-6 Luna $0.10 $0.50 $0.01

OpenAI’s sharper claim is not the list ratio. At max effort on Agents’ Last Exam, it said Terra is over three times as token-efficient as Claude Opus 5.5 and one-fourth the price per token, “resulting in a ~9x lower cost per task.” That ~9x figure is OpenAI’s own cost-per-task math on that exam, not a promise that every desktop job will land nine times cheaper. Token rates also sit beside a separate fee per computer-use tool call, which the model card flags without publishing a dollar amount on the same page.

Independent numbers from Artificial Analysis, which scores models on an Intelligence Index, put max-effort Terra at 48, six points above GPT-5.6 Terra at 42, four points below GPT-6.1 Sol at 52, and ten points above GPT-6 Luna at 38. Cost per task on that index fell to $0.40 from $1.40, which is less than a third. Low and medium effort sit on the firm’s cost-efficiency frontier; max effort does not.

How Fast Can a Mid-Tier Model Drive a Desktop?

OpenAI says GPT-6.1 Terra finishes computer-use tasks over three times faster than both GPT-6 Astra and GPT-6.1 Sol, and that at medium reasoning effort it scores 15 percentage points higher than GPT-5.6 Terra on Agents’ Last Exam at about one-third the cost. At that same medium setting it performs on par with GPT-6 Astra at one-tenth the price per token. The exam’s sample jobs include turning architectural drawings into a 3D model and logging an animated film shot by shot for a restoration team.

WHAT OPENAI MEASURED ON COMPUTER-USE TASKS

  • Speed: Completes evaluated tasks over three times faster than GPT-6 Astra and GPT-6.1 Sol.
  • Agents’ Last Exam: 15 percentage points above GPT-5.6 Terra at medium effort, at about one-third the cost, and on par with Astra at one-tenth the token price.
  • Opus comparison: At max effort, over three times as token-efficient as Claude Opus 5.5 at one-fourth the token price, which OpenAI translates to about nine times lower cost per task.
  • OSWorld 2.1: At medium effort, beats GPT-5.6 Terra’s best score on multistep desktop apps at a fraction of the cost per task.

On MMMU-Pro, a visual-reasoning test, OpenAI said Terra beats GPT-5.6 Terra at every reasoning setting and nearly matches GPT-6 Astra and GPT-6.1 Sol at much lower cost. The model card lists computer_use among supported tools on the Responses API, alongside web search, file search, a hosted shell, and code execution. Input is text and image; output is text. Audio and video are out of scope, which fits a worker that looks at a screen and types, not one that joins a call.

Ted Sanders, who works at OpenAI, put the same split in plainer words on Hacker News after the launch thread mixed up the family names.

Terra is probably better than Astra and Sol at technical communication, similar on vision and computer use (but much faster!), and a bit worse on most other things.

Ted Sanders, OpenAI, on Hacker News

50.3 Percent on Terminal-Bench Still Trails Sol

Command-line coding is where the upgrade is easiest to count and easiest to overread. At maximum reasoning effort, GPT-6.1 Terra solves 50.3% of Terminal-Bench 4.0 tasks, more than double GPT-5.6 Terra’s published 21.5% on Terminal-Bench 4.0. OpenAI’s example item is a webpage whose layout shifts as it loads. Artificial Analysis recorded a gain of about 20 points on the same bench and said max-effort Terra now rivals GPT-6.1 Sol there. Its separate Coding Agent Index v2.0 still has Terra at 48, six points below Sol at 54 and 17 points above Luna at 31.

OpenAI is explicit about the ceiling. “For the most complex coding tasks, GPT-6.1 Sol and GPT-6 Astra remain the better choice.” On Terminal-Bench-Science 0.1, which asks agents to use code for research jobs such as tracking cell divisions in a time-lapse video, it said max-effort Terra beats GPT-6.1 Sol at lower cost and comes within five points of Opus 5.5 at less than one-fifth the estimated cost. That is a middle model stealing some of the science-coding load, not a claim that Terra now leads the board.

Factuality moved in the same direction. On OpenAI’s internal set of error-inducing prompts, the low-effort error rate fell from 53.6% with GPT-5.6 Terra to 16.7% with GPT-6.1 Terra, a drop of 36.9 percentage points. At maximum effort the error rate fell by 18 percentage points to 4.3%, close to Astra and GPT-6.1 Sol on that same adversarial set. Production traffic will not look like a trap set, and OpenAI does not pretend otherwise.

Artificial Analysis measured a different failure: guessing when the model does not know. On AA-Omniscience, GPT-6.1 Terra guesses 38% of the time in those cases, down from 88% for GPT-5.6 Terra, a 50-point drop in hallucination rate, with a two-point rise in accuracy on the same test. Fewer wild fills matter if the job is posting numbers into a ledger. They do not turn a mid-tier coder into Astra.

Cheap Tokens Can Still Burn the Same Quota

The second bill is not on the rate card. A computer-use agent that stops to screenshot, click, and re-read the page pays for every extra loop, including cached context that gets pulled back in. Early timed jobs on October 27 showed Terra finishing some build-and-repair prompts in a fraction of the wall-clock time of Claude Opus 5.5. Other coding runs used about twice as many tokens as GPT-6.1 Sol on the same task, which wiped the 50% list-price cut and left the usage bar in the same place. The pattern is simple: if Terra reaches for computer use when a text-only pass would do, the cheap token is not cheap.

That is why OpenAI’s ChatGPT Work pitch is about allowance, not magic. The company said the same plan covers more tasks with GPT-6.1 Terra than with GPT-5.6 Terra, and substantially more than with GPT-6.1 Sol or GPT-6 Astra. The claim only holds if the extra tasks are the short, visual, form-filling kind Terra was priced for. A subscriber who points it at a long refactor, or who lets it open a desktop tool on every turn, can still hit the cap on a quieter model.

Sanders’s other note on the same thread was that OpenAI will keep a Terra-class model only if people actually use the fast computer-use niche. If traction is poor, the company may go back to a three-model stack. The October 26 price cut is a bet that invoice queues, form farms, and screenshot-heavy chores are large enough to justify a fourth line. The people who still want a better model than Opus 5.5, and said so under the launch post, are not the buyers this SKU is for.

Critical Cyber Rating on a Budget Agent

A cheaper mouse is still a mouse on someone else’s site. OpenAI’s system-card addendum, published October 26, treats GPT-6.1 Terra as Critical in cybersecurity under its Preparedness Framework and High for biological and chemical capability. It uses the same safeguards stack as GPT-6.1 Sol and GPT-6 Astra. In prompt-injection tests, the company said the model showed strong resistance to malicious instructions meant to pull it off the user’s task, and that hierarchy evaluations are saturated at 99.99%, in line with Astra and GPT-6.1 Sol.

SAFETY RATINGS SHIPPED WITH GPT-6.1 TERRA

  • Cyber threshold: Critical under OpenAI’s Preparedness Framework, with High for biological and chemical capability.
  • Prompt injection: Hierarchy robustness at 99.99%, described as similar to GPT-6 Astra and GPT-6.1 Sol, with strong resistance on indirect injections in third-party content.
  • Codex simulation: 22 severity-3 or higher flags on 48,690 matched tasks, or 0.045%, which is 42% fewer than GPT-6 Sol and 64% fewer than GPT-5.6 Sol.
  • Alignment claim: More likely than GPT-5.6 Terra to stay in scope and to disclose limits and blockers, per OpenAI’s evaluations.

Severity 3, in that Codex simulation, means misaligned behavior a reasonable user would not expect and would strongly object to. The 0.045% rate is not zero, and a model priced to run all day will see more total clicks than a flagship you open twice. OpenAI also said some lower-severity flags, including inbound prompt injection, rose versus GPT-6 Sol even as ignoring-instructions and deception flags fell. The guardrails are inherited from the expensive models. The volume is not.

In ChatGPT Work, the Same Allowance Covers More Tasks

GPT-6.1 Terra is available to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Developers call gpt-6.1-terra on the OpenAI API. Reasoning effort runs from none through low, medium (the default), high, xhigh, and max. Tool calling needs the Responses API. Cloud teams that already run OpenAI’s GPT-6 models beside Gemini now have a cheaper sibling in the API for the jobs that are mostly looking at a screen and typing into a field.

WHERE GPT-6.1 TERRA RUNS ON DAY ONE

  • ChatGPT Work and Codex: Plus, Pro, Enterprise, and Edu, with OpenAI saying the same usage allowance covers more tasks than GPT-5.6 Terra.
  • API: Model id gpt-6.1-terra, with batch and flex at 50% off standard rates and fast mode at 2x.
  • Context: 1,050,000-token window, 922,000 max input, 128,000 max output, June 1, 2026 knowledge cutoff.
  • Not the default for hard coding: OpenAI still sends in-depth research and the most complex coding to GPT-6.1 Sol and GPT-6 Astra.

The useful change on October 26 is that computer control no longer has to wait for a flagship budget. Teams can leave a mid-tier agent on a queue of screens, if they count the extra loops and keep a person near the dangerous clicks. Sol and Astra remain the models you pay when the job is allowed to fail only once.

Frequently Asked Questions

What is GPT-6.1 Terra’s context window?

GPT-6.1 Terra’s context window is 1,050,000 tokens, with a maximum of 922,000 input tokens and 128,000 output tokens on the OpenAI API model card. Prompts above 272,000 input tokens are billed as a long-context request, which is a separate surcharge from those caps.

What is GPT-6.1 Terra’s knowledge cutoff?

The model card lists a June 1, 2026 knowledge cutoff. Anything after that date needs live tools such as web search, or files you attach, because the weights themselves stop at early June.

Does GPT-6.1 Terra support audio, video, or fine-tuning?

No. Supported input is text and image, supported output is text, and the card lists audio and video as unsupported modalities. Fine-tuning and predicted outputs are also unsupported, so teams that need a custom checkpoint still have to stay on a model that allows it.

How does GPT-6.1 Terra pricing change above 272,000 input tokens?

For prompts with more than 272,000 input tokens, OpenAI charges 2x input and cache rates and 1.5x output on the full request. Regional processing adds a 10% premium where it is offered. Cache writes for 12-hour retention are $2.50 per million tokens, double the 30-minute write rate of $1.25.

Which reasoning effort settings does GPT-6.1 Terra support?

The model accepts none, low, medium, high, xhigh, and max, with medium as the default. Artificial Analysis found the low and medium settings sit on its cost-efficiency frontier, while max effort scores higher and costs more per task, so the extra setting has to earn the spend on that job.

Harry is the editor of IAQABA, an independent publication he owns and runs. A decade in journalism, beginning as a reporter and now as the editor of his own titles, has left him with a clear test for what deserves a story: it has to change what a reader knows or decides, and it has to rest on something he can point to. That rules out recycled press releases, forecasts with no data behind them and rumours that no document supports. It leaves room for a great deal, and the site covers news, business, science and technology alongside sports, entertainment and lifestyle, with travel, auto and gaming given the same standard rather than lighter treatment. Sources are primary wherever possible: the regulator's filing, the company's own statement, the transcript, the dataset, or the product on Harry's desk. Figures are checked before they are published and rechecked if a reader questions them. Mistakes are corrected under a published policy. Readers across the world can reach him directly at support@iaqaba.com.

Click to comment

Leave a Reply

Your email address will not be published. Required fields are marked *

Trending