Introducing GPT 6.1 Terra

Justy OpenAI just dropped GPT-6.1 Terra — and honestly, this might be the first model in the GPT-6 family that actually makes sense to buy.

Cody That's a big 'might'.

Justy Come on. Fifty percent cheaper than GPT-5.6 Terra and GPT-6.1 Sol, and it's built for the stuff that actually fills a workday — reading invoices, entering them into accounting software, interpreting technical drawings.

Cody Right, right. But let's look at what they're actually claiming. On Agents' Last Exam, medium reasoning effort, fifteen points higher than GPT-5.6 Terra at about one-third the cost. That's the headline.

Justy And it matches GPT-6 Astra on that same benchmark at one-tenth the token price.

Cody Which is a hell of a soundbite. But it's one benchmark, reported by OpenAI, and we've seen how fast those numbers deflate when someone else runs the harness.

Justy Fair. Still, the error-rate drop is real-looking — fifty-three point six percent down to sixteen point seven percent at low reasoning effort. That's nearly thirty-seven points.

Cody Internal benchmark. I'd want to see it on something public before I trust it.

Justy Sure, but think about the user story. You're a team paying for GPT-6.1 Sol, and now Terra does most of your high-volume document work for half the price. That's an upgrade path that doesn't require a routing spreadsheet.

Cody Except Terminal-Bench 4.0. Max effort, Terra solves fifty point three percent. Sol is at fifty-five point nine, Claude Opus 5.5 at seventy point six. So for the complex coding stuff, Sol still wins, and Opus wins bigger.

Justy They literally say that in the post — 'for the most complex coding tasks, Sol and Astra remain the better choice.' They're not pretending Terra is the frontier hammer.

Cody Good. That's the honest part. The questionable part is the pricing frame. One dollar per million input, five per million output. They say it's fifty percent lower than GPT-5.6 Terra, but some of the third-party pricing sheets I saw list different Sol numbers — I can't reconcile them yet.

Justy Mm-hm. Either way, the direction is clear. OpenAI is pushing the cost-intelligence frontier down-market. Luna is the cheap abundance tier, Terra is the workhorse, Sol is the halo. That's a procurement story, not a benchmark story.

Cody And it lands in a crowded week. Kimi K3.1 Pro came out October twenty-third. Claude Opus 5.5 is still the agentic-coding king according to third-party tests. Terra isn't displacing the closed frontier on hard reasoning — it's making computer use cheap enough to automate.

Justy Exactly. That's the zoom-out. It's not 'Terra beats everything.' It's 'Terra makes high-volume visual and desktop tasks affordable enough to actually ship.'

Cody I'll buy that. With receipts — I want to see independent runs on Agents' Last Exam before I quote that fifteen-point gain to a client.

Justy Sixty-forty somebody reproduces it within a month and it comes in lower.

Cody I'll take the under. Seventy-thirty it falls apart.

Justy You're impossible.

Cody I'm calibrated.

Justy If you want to try it, the API model is gpt-6.1-terra. One dollar input, five output, zero point zero five cached. It's in ChatGPT Work and Codex today for Plus, Pro, Enterprise, and Edu.

Cody Worth a spin on invoice processing. See if the error rate holds outside OpenAI's internal test.

Justy There it is. Cody, skeptical but game.

Cody Someone has to be.

Justy Catch you next time.