· AI
OpenAI Says GPT-6.1 Terra Costs Half as Much. An Outside Lab Checked, and the Real Number Is Bigger.
A vendor's price cut is a claim until someone outside the vendor checks it. This time, someone did, and the claim held up better than advertised.
OpenAI released GPT-6.1 Terra on October 26, pricing it at $1 per million input tokens and $5 per million output tokens, half of what the model it replaces, GPT-5.6 Terra, charged at $2 and $12. On the surface that is one more entry in a price war that has been running all year: Anthropic cut Claude Haiku to the same $0.10 floor two weeks earlier, and Chinese open-weight labs have been undercutting both for months. A 50 percent token-price cut from a frontier lab is not news anymore. What is worth a second look is that this particular claim did not stay a press release.
Artificial Analysis, a benchmarking firm with no stake in OpenAI's pricing, ran its own Intelligence Index test on Terra against its predecessor and found it gained 6 points in capability while the cost per task fell from $1.40 to $0.40. That is a 71 percent drop in what it actually costs to get a task done, well past the 50 percent cut in the per-token rate card. OpenAI's own numbers point the same direction: on an internal hallucination benchmark, Terra's error rate at low reasoning effort fell from 53.6 percent to 16.7 percent, and on Terminal-Bench 4.0 it solved 50.3 percent of tasks at maximum effort against 21.5 percent for the model it replaced.
Here is the steelman for not trusting any of this, and it is earned. Frontier labs have a long recent history of announcing a benchmark that only they ran, graded against tasks they picked, and letting the headline number do the marketing. I have made exactly that case on this blog before, when Notion's own router claimed a 90 percent cost cut on its own blog and 87 percent on its own help page, both self-graded by a judge Notion built. Most of what gets called a benchmark in this industry is a company checking its own work.
That is precisely why the Artificial Analysis number matters more than the other figures in this story. It did not come from OpenAI. A price cut you can only verify against the vendor's own claim is worth exactly as much skepticism as the vendor's incentive to inflate it. A price cut that an outside lab reproduces independently, and finds is actually larger than advertised, is a different kind of evidence. It does not prove every number in OpenAI's announcement holds up, and it should not. But it tells you the core claim, that this model does more work per dollar than the one before it, survived contact with someone who had no reason to be generous.
The number to budget against is still not the per-token price. It is cost per task, and that number moves with how much reasoning effort a job actually needs. Terra's low-effort error rate of 16.7 percent is real, and a business routing high-stakes work there on the strength of a price chart alone will eat that error rate as rework. The efficiency gains are also concentrated in exactly the work OpenAI built this model for, computer use and agent tasks. For harder coding work, OpenAI itself still points customers to the pricier Sol and Astra tiers. A 71 percent independently verified drop in cost per task on one benchmark is a strong signal. It is not a guarantee that your specific workload sees the same number.
So the practical move is the same one I would have told you before this model existed: test the cost per task on your own work, at the reasoning effort your task actually needs, before you move production traffic. What changed this week is that you now have one external data point that says, for this model, that number is likely to beat what the rate card implies rather than fall short of it. That is the useful kind of evidence, and it is rare enough in this industry to call out when you see it.
This is the exact check MojoAI runs before recommending a client change model or vendor: not whether the vendor's own number is impressive, but whether an outside source, or our own test on the client's actual tasks, agrees with it. If your AI budget is still set by last quarter's price chart, that conversation is free.
Sources
References used in this article. Links also appear alongside the relevant claims.
Let's make it happen.
You don't need a polished brief. A couple of lines about where you want your company to go is plenty, and we'll come back with what we'd tackle first.
We scope the work around your goals and timeline before quoting anything, so you know exactly what you're getting.