GPT-6.1 Terra Is Fantastic at Some Work and Terrible at Others

About This Video

A 22-minute video from Matt Maher on GPT-6.1 Terra, OpenAI's mid-tier model, which he says landed on OpenAI's API on October 26th. Maher spent several days running Terra on his real work (he says he drained his account to zero several times) to learn which jobs a cheaper model can take from the larger Sol. Everything below reflects his personal experience and opinions, not a controlled benchmark. Model names are spelled as they appear in the auto-generated captions, which sometimes render "Sol" as "sole" or "Saul" and "Terra" as "Tara"; check OpenAI's own materials for official names and pricing.

Summary

Maher's conclusion is that Terra is a strong worker model. When the work is bounded, describable, or already broken into pieces by a larger model, it does it well. When the work is open-ended or needs lateral thinking, it tends to stay inside a "box" of rules, real or imagined, and produces reasonable but not best results. He prefers it to Sol for writing, but says he cannot simply switch off Sol and move to Terra.

The Lineup

• Luna (00:33). The most efficient model at the bottom of the lineup.
• Terra (00:33). Aimed at visual reasoning, working a computer, and natural writing. He says it costs about 10 times more than Luna per token at standard API rates.
• Sol (01:05). Costs twice as much as Terra. This is the model Maher works with all day on extra-high reasoning.
• Astra (01:05). The most capable model above all of them.
• Analogy (01:30). He says choosing between Terra and Sol is much like choosing between Opus and Sonnet on the Anthropic side.

Where Terra Worked Well

• Full release pipeline (02:35). Maher describes changes that go through CI/CD, an app store whose version and pricing must match, a third-party payment system with sandbox testing, Apple's publishing requirements, and a website update. He says the workflows were built earlier with other models, "very likely Sol extra high" (04:33).
• Terra shipped changes to customers (05:01). Following a process that was already established, it took changes all the way out, including a new monitor in his Agent Runway software, built with Terra and OpenAI's image model for the graphics.
• Bounded work (05:35). He defines this as a known pattern that is describable, where you know what needs to happen and perhaps the whole workflow. He says Terra is not as good as Opus at coding but Opus is overkill when the plan is already clean.
• Testing (06:32). Terra did the coding, the CI/CD process, and a great many evaluations and tests, which he says it takes "a little bit too seriously."

Writing: Terra Beats Sol

Maher records himself talking for an hour on a walk, then hands the recording to a model to find the narrative structure (06:32 to 07:34). He says the model is mostly reconstructing a story that is already in his own talk, and that he keeps and lightly touches up large chunks. In his view Terra is "markedly better" than Sol at this, consistent with OpenAI saying natural writing was a design goal (07:34 to 08:04). He adds the caveat that Opus is still the best writing system in his view.

Where Terra Struggled

• Open-ended problems (08:04). In his "director pattern" one agent spreads work across others. When something goes wrong he asks the model to evaluate the pattern and propose approaches rather than telling it what to fix.
• Rule confusion (09:32). Terra is highly affected by rules, or rules it thinks it has. He says it declined to use sub-agents because it was not sure it was allowed to, even though he had told it that it could.
• Its own sandbox (10:03). It often invents a box and gets stuck inside it. If the box happens to cover the problem that is fine; if not, it does not step outside.
• Same problem, bigger model (10:30). Giving the problem to a bigger model often produced something very different, with more ideas and more pushback.
• Other models compared (11:04). He says even Astra sits in a box and mostly tries to get work done, and that Anthropic's Fable is the one he can give unbounded work without it being overly polluted by the structure in front of it. He rates Opus better than Sol at opening up an idea space.
• Multi-agent attempt (11:30). He had a group of Terra sub-agents with different considerations trade ideas, but each stayed stuck in its own orientation and rule set.
• Quality ceiling (12:05). He never saw bad work, but also never the best work. With tight token limits he would probably use it, but would not be confident it was the best answer.
• Trust (12:34). The bigger the work, the harder the decisions underneath are to inspect, and trust habits built on stronger models do not transfer easily to weaker ones.

Computer Use

• Unbounded in practice (13:32). Unfamiliar editors, graphics software, or websites it has not seen make a seemingly bounded job open-ended.
• Succeeds or stops (14:01). On a task like fetching an API key from a website, it either succeeded or said it could not, and he says it never handed back the wrong key. Other times it went in circles in menus.
• Affinity Designer example (14:34). Asked to lay objects out in a grid, it searched the menus repeatedly, concluded there was no way, and even after being told there might be no menu option, advised him to drag each object himself. It never reached the brute-force approach of moving each object.

Mixing Models

• Worker model (15:33). Good for bounded work or pieces a larger model has already broken down. The wrong choice when the work might need creativity or lateral decisions.
• Reasoning level (16:04). Sol on medium probably works very well, but lower reasoning reduces how many times a model can reconsider. He says calling a low-reasoning model "just not as smart" is too simplistic.
• Build the mix into the system (17:00). He gives his director guidance on which work is simple or a known pattern and which model suits it, and prefers these to be suggestions rather than rules.
• Escalation points (17:30). In a workflow he adds a step where a higher-model sub-agent makes a judgment call (his example is what to highlight on the website), then the workflow continues.
• Debugging by asking (18:01). When output is wrong, he asks the system which model did the work. If it was too low, he pushes that kind of work up a model.
• Bigger is not a cure (18:32). He says Astra, used to refactor the director pattern, discarded half the original requirements and cut a test suite of roughly a hundred cases to far fewer. It became myopic about speed and fewer instructions and forgot a refactor should keep behavior the same.
• What he will not accept (20:04). Terra getting stuck in its own thinking box often enough that he had to steer mid-process, which he is not willing to do.

The Compounding Problem

From 21:01, Maher argues that "reasonable" results compound. One pretty-good result is easy to absorb. Fifty in one request are hard to review, and if you dislike the end result you cannot tell which decisions caused it. He concludes he cannot turn off Sol and move to Terra "in no possible way" (21:33), but encourages trying Terra for things like writing and staying aware of what you give up by saving tokens (22:03).

Key Takeaways

1. Use a mid-tier model for bounded, known-pattern work, especially when a larger model has already planned the steps.
2. Terra was better than Sol for natural writing in Maher's daily use, though Opus remains his favorite writer.
3. Open-ended problems exposed Terra's tendency to stay inside rules it has or imagines.
4. In computer use, Terra tended to succeed or stop, but could loop and hand work back rather than improvise.
5. Build model choice into the workflow as suggestions, and add escalation steps to a stronger model at decision points.
6. When output disappoints, check which model produced it before blaming the approach.
7. Many "pretty good" results stacked in one request can add up to a result you cannot diagnose.

Reflect & Discuss

• Which of your tasks are truly bounded, and which only look bounded until the model meets something unfamiliar?
• Where in your workflow would an escalation step to a stronger model be worth its cost?
• How would you notice a model has built an imaginary rule box around a problem?
• If a single request chains fifty decisions, how would you review them?

Notes on the Source

No sponsor segment is present in the transcript. Maher mentions offering a link to Agent Runway, his own software, in the description (05:35); that is self-promotion. Pricing relationships are relative multiples as he states them, not dollar figures.