GPT-6.1 Terra Price and the 9x Cheaper Claim

OpenAI launched GPT-6.1 Terra on October 26, and most of its launch post comes down to one sentence: at maximum effort, Terra is “over three times as token efficient as Claude Opus 5.5, and one-fourth its price per token, resulting in a ~9x lower cost per task.”

That’s OpenAI’s number, from OpenAI’s benchmark, about a competitor’s model. So it’s worth asking what has been checked and what hasn’t.

What OpenAI released

Terra went live on October 26 for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, and in the API as gpt-6.1-terra. OpenAI describes it as a fast model for computer use, visual reasoning and “clearer, more natural writing”: filling in forms, reading technical drawings, entering invoices into accounting software.

Standard API prices are $1 per million input tokens, $0.05 for cached input and $5 per million output tokens, according to OpenAI’s pricing page. Its predecessor, GPT-5.6 Terra, was $2 and $12. OpenAI also says Terra finishes tasks more than three times faster than GPT-6 Astra and GPT-6.1 Sol.

The 9x figure

OpenAI measured the cost claim on Agents’ Last Exam, a benchmark of professional tasks that OpenAI’s post describes but doesn’t link to. I couldn’t find a public leaderboard for it, so nobody outside OpenAI can rerun the comparison today.

The arithmetic is also loose. Three times fewer tokens at a quarter of the price per token would come to roughly 12x, not 9x. The post says “over three times”, which if anything pushes the figure higher, so rounding doesn’t close the gap. Maybe the token mix (cached versus fresh) does, but that’s my guess. OpenAI doesn’t show the working.

What the independent group measured

Artificial Analysis, an independent benchmark group, ran Terra on its own tests. Its numbers are less dramatic, and in my view easier to check.

Measure	Artificial Analysis result
Intelligence Index, Terra at max effort	48 (model page)
Intelligence Index, Claude Opus 5.5 (top of its list)	58
Intelligence Index, GPT-6.1 Sol	4 points above Terra
Cost per Index task, Terra max vs GPT-5.6 Terra	$0.40 vs $1.40
Coding Agent Index v2.0, Terra vs GPT-6.1 Sol	48 vs 54

So Terra is ten points behind Opus 5.5 on that index, and OpenAI’s own post tells developers to pick GPT-6.1 Sol or GPT-6 Astra for the hardest coding work. Artificial Analysis’s verdict is about cost efficiency: Terra at low and medium effort sits on the frontier of intelligence per dollar. Its page doesn’t test the Opus comparison that OpenAI headlined, and I haven’t seen another group do it either.

Hallucinations: two tests, same direction

OpenAI says the error rate on its internal set of error-inducing prompts fell from 53.6% to 16.7% at low effort, and to 4.3% at max effort. Artificial Analysis used a different private test, AA-Omniscience. There, Terra guessed an answer 38% of the time when it didn’t know, down from 88% for GPT-5.6 Terra, and accuracy rose by 2 points.

The two tests measure different things, so the numbers can’t be compared with each other. Both point the same way, and that’s the part I’d trust most.

One more gap: Terminal-Bench 4.0. OpenAI reports 50.3% against 21.5% for the old Terra, a gain of about 29 points. Artificial Analysis describes a gain of roughly 20 points on its run. Different harnesses, probably, but the gap is there.

Safety line

OpenAI says Terra “meets the Critical threshold in cybersecurity” under its Preparedness Framework and keeps the protections from GPT-6 Astra. The system card is linked from the launch post. I haven’t read it end to end, so I’m only passing along the headline.

My read

Terra looks like a real price cut: $1 per million input tokens (half the previous Terra’s $2) and $5 for output (down from $12), and Artificial Analysis measured its own cost per Intelligence Index task at $0.40 against $1.40 for the old Terra, a drop of more than two thirds. That part I’d take at face value. The “9x cheaper than Opus 5.5” line I’d treat as a vendor claim until someone outside OpenAI reproduces it. If you’re picking a model for invoices and forms, run your own fifty tasks on Terra and whatever you use now. Last month’s price move is covered in OpenAI halves GPT-6 prices, and the pattern there was the same: the discount was real, the headline comparison was OpenAI’s own.

Note on sources

I read OpenAI’s launch post and pricing page, Artificial Analysis’s write-up, model page and index page, and BleepingComputer’s coverage. The Verge piece I could only read in part, its opening paragraph. I did not read OpenAI’s system card in full, and Agents’ Last Exam has no public results page I could find.

OpenAI says GPT-6.1 Terra costs about 9x less per task than Claude Opus 5.5. The independent scoreboard has Terra ten points behind.