GPT-6.1 Terra Explained: $1/$5 Pricing vs Claude Sonnet 5.5 and Haiku 5.5

GPT-6.1 Terra is OpenAI's new mid-tier model, launched October 26, 2026, priced at $1 per million input tokens and $5 per million output tokens: half of what GPT-5.6 Terra cost and roughly a tenth of what GPT-6 Astra charges per token. It scores 50.3% on Terminal-Bench 4.0 at maximum reasoning effort, more than double GPT-5.6 Terra's 21.5%, and OpenAI says it runs more than three times faster than both GPT-6 Astra and GPT-6.1 Sol.

That's the pitch in one paragraph. The rest of this post is about whether the pitch holds up once you put Terra next to Claude Sonnet 5.5 and Claude Haiku 5.5, because that's the comparison most developers actually care about right now.

What Is GPT-6.1 Terra Built For?

Terra isn't trying to be OpenAI's smartest model. It's the one built for the boring-but-constant stuff: reading what's on a screen, clicking through forms, drafting emails, matching product photos to catalog entries, turning a technical drawing into something usable. OpenAI's own announcement calls it "a fast, efficient model that combines detailed visual reasoning and precise computer use alongside clearer, more natural writing."

One OpenAI employee, Ted Sanders, put it more bluntly on the company's community forum: 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." That's an honest trade-off description, and it tells you exactly where this model belongs: high-volume, low-to-medium complexity work, not your hardest coding problem.

How Much Does GPT-6.1 Terra Cost?

The context window is 1.05 million tokens total, split into up to 922,000 input tokens and 128,000 output tokens. That's a big window for a model priced this low. Terra also supports reasoning.effort:"none", one of six effort levels (none, low, medium, high, xhigh, max) and a toggle Astra and Sol don't get.

The headline number (50% cheaper than GPT-5.6 Terra) checks out against OpenAI's own pricing page. The "one-tenth the price of Astra" claim also holds, though keep in mind Astra is OpenAI's flagship, so that's not really a fair fight on capability.

How Does Terra Compare to Claude Sonnet 5.5 and Haiku 5.5?

Here's the table that actually matters if you're choosing a model for coding work:

Sonnet 5.5 still wins on raw coding capability, by a wide margin. Terra lands between Sonnet and Haiku on the benchmark that measures actual agentic terminal work, but it's priced closer to Haiku territory than Sonnet territory on output tokens. That's an odd spot to occupy, and it's part of why the comparison is interesting instead of a blowout.

OpenAI's own numbers back up the "Terra for routine, Sonnet for hard problems" split. Terra's error rate drops from 53.6% at low effort (on the prior generation) to 4.3% at max effort, nearly matching Astra and Sol. But OpenAI's release notes still recommend Sol or Astra for the hardest coding tasks, and nothing in the public benchmarks suggests Terra closes the gap against Sonnet either. Push either model into a genuinely hard refactor or multi-file agentic coding task, and Sonnet pulls ahead.

Against Haiku 5.5, the picture flips. Haiku is cheaper per token below 100K context ($0.10/$0.50 versus Terra's $1/$5) and still beats Terra on Terminal-Bench 4.0 by 11 points at matched context. If your workload is small, cheap, high-volume calls, Haiku is the better default. Terra's case is the 1M-token context window and the computer-use layer. Haiku doesn't do screen understanding and app navigation the way Terra does.

What About Agents' Last Exam?

Terra's numbers look best here. On Agents' Last Exam (OpenAI's benchmark built around realistic workday tasks, like turning architectural drawings into 3D models or reviewing an animated film shot by shot), Terra at medium effort scores 57.6% for about $0.29 per task, matching GPT-6 Astra's medium-effort score at a fraction of Astra's $4.82-per-task cost. At maximum effort, Terra hits 60.0% for $1.01 per task.

Claude Opus 5.5 still scores higher on the same benchmark, 63.2%, but at $8.82 per task — nearly nine times Terra's cost for roughly three extra points. If you're running this kind of task at scale, that math is hard to ignore. If you're running it occasionally and need the best possible answer, Opus is still the safer pick.

Why Is Everyone Arguing About the Name Instead of the Model?

Because OpenAI now has Luna, Terra, Sol, and Astra, each with its own version numbers, and six reasoning effort levels stacked on top. One commenter on the launch thread did the math: "You have Luna, Terra, Sol and Astra, each with 4+ reasoning levels. That's ~20 options to choose from." The Hacker News discussion around the launch pulled 566 points and 353 comments, and from what I can tell, most of it wasn't about benchmarks at all — it was people arguing over whether Terra 6.1 beats Sol 6.0, or how a mid-tier model keeps leapfrogging the tier above it.

Honestly, I get the frustration. "Does the mid-tier model from this version beat the big model from last version" shouldn't require a spreadsheet. Anthropic's Opus/Sonnet/Haiku split, even with the 5.5 version jump, is still easier to reason about at a glance. OpenAI's constellation naming is cute until you need to pick a model for a Tuesday afternoon task and have to go look up a chart first.

Should You Actually Use GPT-6.1 Terra?

My take: yes, for the specific lane it's built for. If you're running high-volume computer-use automation, form filling, document drafting, or screen-reading agents, Terra's price-to-performance on Agents' Last Exam is genuinely good, and the 1M context window gives you more room than Haiku 5.5 offers below its own pricing cliff.

For coding work specifically, I wouldn't switch off Sonnet 5.5 if you're doing anything past routine edits. The 20-point Terminal-Bench gap between Terra and Sonnet is too wide to paper over with a lower price tag. And Greg Brockman's own comment about Terra, that "computer use with terra has crossed a qualitative threshold where it feels faster at using a computer than what i can do unassisted," is specifically about computer use, not code generation. That's a tell. OpenAI is selling Terra on speed and screen understanding, not on beating Claude at coding, and the benchmarks agree with that framing.

If you're already comparing Claude's lineup, our breakdowns on Claude Haiku 5.5's pricing and benchmarks and Claude Opus 5 vs Fable 5 cover where Anthropic's own tiers split the same way. And if you want the other side of OpenAI's lineup, GPT-6 Pro and parallel agents in ChatGPT covers the flagship end of this same naming mess.

Watch: The Announcement

Terra isn't a Claude killer and OpenAI isn't pretending it is. It's a cheaper, faster model for the bulk of work that doesn't need a frontier model's full reasoning budget, and on that narrow job, the price makes sense.