🔥 Trending on HN

GPT-6.1 Terra puts computer work, not just answers, at the center

3 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
computer use(computer use)

An AI’s ability to see a screen and take actions inside computer apps.

benchmark(benchmark)

A fixed test used to compare how well AI models perform.

prompt injection(prompt injection)

An attempt to make an AI ignore its assigned task.

What happened

OpenAI announced GPT-6.1 Terra, a model built for fast computer work. OpenAI’s announcement says Terra combines visual reasoning, precise computer use, and clearer writing. It targets workday tasks such as drafting proposals, reading technical drawings, filling forms, and editing videos. Plus, Pro, Enterprise, and Edu users can access it in ChatGPT Work and Codex. Developers can use it through the API as gpt-6.1-terra. A post about the launch drew more than 500 points on Hacker News. That number shows attention, not proof that the claims are true. The Hacker News post should be read separately from the product announcement.

Background

OpenAI is not presenting Terra as the best model for every job. It presents Terra as a balance of speed, capability, and cost. Its standard API prices are 50% lower than GPT-5.6 Terra and GPT-6.1 Sol. The listed prices are $1 per million input tokens, $0.05 per million cached input tokens, and $5 per million output tokens. A token is a small unit used to count processed text. OpenAI says this price makes high-volume work more practical, such as reading invoices and entering their details into accounting software.

Why it matters

Computer use changes the question from Can an AI write an answer? to Can it complete the next step in a work app? Terra is designed to read a screen or drawing, then act inside software. That connects language with routine office work. Lower cost could make repeated tasks easier to automate. Faster completion could also let one workflow handle more items. These are the reasons the release matters beyond another model name. They are also the reasons reliability matters. A cheap action that fills the wrong form can still create expensive work for people.

What is confirmed

OpenAI reports several improvements in its own evaluations. On Agents’ Last Exam, Terra scored 15 percentage points above GPT-5.6 Terra at medium reasoning effort, at about one-third the cost. OpenAI says it performed on par with GPT-6 Astra at one-tenth the price per token. At maximum effort, OpenAI reports more than three times the token efficiency of Claude Opus 5.5, one-fourth the price per token, and about nine times lower cost per task. It also says Terra finished tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol. Terra beat GPT-5.6 Terra on MMMU-Pro and OSWorld 2.1. On Terminal-Bench 4.0, it solved 50.3% of tasks at maximum effort, compared with 21.5% for the older model. OpenAI still says GPT-6.1 Sol and GPT-6 Astra are better for its hardest coding tasks. In an internal error-inducing test, the error rate fell from 53.6% to 16.7% at low reasoning effort, and to 4.3% at maximum effort. OpenAI also reports that Terra meets its Critical cybersecurity threshold and resisted prompt injection tests designed to pull it away from a user’s task.

What remains unclear

All of those performance numbers come from OpenAI’s evaluation setup. The release does not establish whether independent testers will see the same gaps. Benchmark tasks also cannot show every problem in a long business workflow. We still do not know how often Terra loses track of a screen, needs human help, or makes a costly mistake. A lower token price is not the same as a lower total cost after review and correction.

What to watch next

The next useful evidence will come from the system card, independent reruns, and real customer workflows. It will be important to see the kinds of errors Terra makes while filling forms, editing videos, and using desktop apps for a long time. The central test is whether Terra can become a dependable work tool, not merely a strong score on a company’s launch page.

💬 Is GPT-6.1 Terra a cheap middle-ground model?

HN commenters are split between seeing Terra as a cost-effective workhorse and questioning the benchmark selection and coding performance. The cautious view is that the article alone does not establish broad parity with Opus 5.5 or Sol.

  • Some commenters read the article’s favorable charts as showing Terra near Astra while costing less and competing with Sol. Others warn that the release may emphasize benchmarks where Terra performs well, so the full evaluation suite matters.
  • In user self-reports about Terminal Bench 4.0, Terra needed xhigh to get close to GPT-6.1 Sol, while costing more and performing slightly worse. Another user reported 50% for Terra at max effort versus 70% for Opus 5.5.
  • A separate user self-report put Terra Max at 48 and Opus 5.5 Max at 58 on the AA Index. That weakens any broad claim that Terra matches Opus 5.5.
  • Several commenters see Terra’s likely niche as computer work, automation, and technical writing rather than the hardest software-engineering tasks.
  • Supporters argue that API buyers should measure cost, time, and quality per completed task, not token price alone. Others warn that benchmark and pricing comparisons can change depending on how the task is defined.
  • One commenter suggested that Terra’s weaker Terminal Bench result might reflect less over-optimization for the benchmark. This remains a hypothesis, not a verified explanation.
  • Overall, Terra may be attractive for particular workflows, but users should compare quality, latency, and total cost on their own real tasks instead of relying only on the article’s charts.

growing digest at 405 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

GPT-6.1 Terra can look at screens and do computer work

📰 Full story: GPT-6.1 Terra puts computer work, not just answers, at the center

OpenAI has introduced a new AI model. It focuses on seeing screens and using computer apps.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
API(A-P-I)

A way for another program to use an AI model.

benchmark(benchmark)

A test used to compare AI models.

token(token)

A small piece used to count text processed by an AI.

💡 The gist

  • GPT-6.1 Terra can see screens and use computer apps.
  • OpenAI says it is faster and cheaper than earlier models.
  • Tests look better, but real-world results remain unknown.

GPT-6.1 Terra is a new AI model from OpenAI, the company behind ChatGPT. It can read drawings and screens. It can fill forms and edit videos. It can also write clearer text. OpenAI calls this computer use. It means the AI can look at an app and take steps inside it.

OpenAI wants Terra to handle many repeated tasks. One example is reading invoices. Another is entering invoice details into accounting software. OpenAI says these jobs become more practical when the model costs less. Its standard API price is $1 per million input tokens. A token is a small piece used to count text. Cached input costs $0.05. Output costs $5 per million tokens.

OpenAI compared Terra with older and larger models. In one professional-task benchmark, Terra scored 15 points higher than GPT-5.6 Terra. It cost about one-third as much in that test. OpenAI also says Terra nearly matches GPT-6 Astra at a much lower token price. On a coding benchmark, Terra solved 50.3% of tasks. The older Terra solved 21.5%.

OpenAI tested factual errors too. At low reasoning effort, the error rate fell from 53.6% to 16.7%. At maximum effort, it fell to 4.3%. These numbers come from an internal test. They do not prove that every answer will be correct. OpenAI says Terra also resisted prompt injection tests. Those tests check whether bad instructions can pull the model away from its task.

The launch became a major topic on Hacker News. The post received more than 500 points. Points measure community attention. They do not prove the model works as advertised. Independent tests must show how Terra behaves during long, real jobs. People also need to check its work. You can read the OpenAI announcement and the Hacker News post.

💬 Is Terra cheap and useful? The comparison is still hard

Some people think Terra is a cheaper model that can handle many jobs. Others question whether the tests were fair and how well it codes.

  • The article’s charts made some users think Terra is close to larger models for less money. Other users worry that the most favorable tests were selected.
  • A user reported that Terra needed xhigh on Terminal Bench 4.0, yet cost more and did slightly worse than Sol. Another user reported 50% for Terra and 70% for Opus 5.5.
  • On the AA Index, a user reported 48 for Terra Max and 58 for Opus 5.5 Max. These are user reports, not final proof.
  • Terra may fit computer work, automation, and technical writing better than the hardest coding. People also disagree about whether to measure token price or the total cost of finishing a job.
  • A low coding score might mean less benchmark gaming, one commenter guessed, but the safest test is still your own work.

growing digest at 405 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

🔥 Trending on HN

GPT-6.1 Terra helps with computer screens

📰 Full story: GPT-6.1 Terra puts computer work, not just answers, at the center

OpenAI made a new AI that can look at screens and help with work.

1 min read Tiny Why Newsroom · By Curio, Martian correspondent

Words
GPT-6.1 Terra(G-P-T six point one Terra)

The name of OpenAI’s new AI.

Hacker News(Hacker News)

A website where many technology readers discuss news.

OpenAI, the company behind ChatGPT, made a new AI. Its name is GPT-6.1 Terra.

It can look at a screen. It can read a drawing. It can fill boxes. It can help edit videos. It can use computer apps.

OpenAI says it is faster and cheaper than before. OpenAI also says it made fewer mistakes in tests. Tests are not every real job. The AI can still make mistakes.

Many tech readers discussed it on Hacker News. That shows attention, not truth. A person should check the AI’s work.

Read the OpenAI announcement.

💬 What are people saying about Terra?

Some people see Terra as a cheaper helper. Others want more proof that it is strong.

  • One commenter thinks Terra may be good for computer chores, automation, and technical writing.
  • Users reported weaker results on some coding tests: 50% for Terra versus 70% for Opus 5.5, and 48 versus 58 on another test. These are personal reports, not a final verdict.
  • People worry that the article chose friendly tests. Another person guessed that a low score could mean less test-gaming, but that is only a guess.
  • People also debate whether to count tokens or the whole finished job. Trying Terra on your own work is the clearest check.

growing digest at 405 comments (revision 2). We fetched 300 comments and sampled 120 across the thread. These are HN users’ reports, not independently verified facts.

Sources