Eliza Ward: Hey. Two things dropped yesterday and I keep bouncing between them because I can't tell if they're actually one story.
Brian Reed: The price cut and the math thing.
Eliza Ward: Yeah. OpenAI launched GPT-6.1 Terra on October 26th — cut the API price in half compared to GPT-5.6 Terra — and on basically the same day submitted AI-generated results to outside reviewers. Including work on something called the Yang-Mills problem, which is — wait, actually let me just say what that is first.
Eliza Ward: The Clay Mathematics Institute has seven open problems. A million dollars each for a verified solution. Yang-Mills is one of the seven. Nobody has solved it. OpenAI is saying its AI produced results related to it.
Brian Reed: Hang on — so the price cut for a workplace automation tool and a potential Millennium Prize solution are both landing in the same 48-hour window? That's a lot to hold together.
Eliza Ward: And here's the intuition click, I think. GPT-6.1 Terra is a computer-use agent — meaning it can literally sit at your computer and do the clicking for you. Apps, browsers, whatever. That's the product. OpenAI just made renting it dramatically cheaper. The science submission is meant to be the credibility layer underneath that.
Brian Reed: So the pitch is: trust this AI to run your workflows because look, it might have cracked a problem mathematicians haven't touched in decades. The question I'd want answered is — who's actually doing the verifying, and does the Yang-Mills paper ever become public?
Eliza Ward: That's the part that doesn't actually work — the public part. OpenAI published a framework, no NDAs, no academic credit claims, recognition for the reviewers. That's real. Frontier labs basically never do that. But AGMAI put out a clarifying statement the day before the release specifically laying out how these results should be handled, naming which experts should shepherd them into the public sphere. And OpenAI's approach to two of the three results violated that statement.
Brian Reed: Wait — the day before?
Eliza Ward: The day before. And AGMAI specifically called on OpenAI to release the Yang-Mills paper. The quote is — I mean, pretty direct — 'there is no scientific reason for delaying any further.' As of October 27th, the paper was not public.
Brian Reed: So the governance body that's supposed to be shepherding this is saying release it, and OpenAI just... didn't. And they're still calling the process open. The part I don't get is — what does open verification even mean if nobody outside OpenAI can read the manuscript?
Eliza Ward: Right — and that's the distinction. Appointing reviewers is not the same thing as making the document exist for anyone to check. Those are two completely different claims.
Brian Reed: And as of October 28th — yesterday — no independent verification of any of the three results had been completed. So we're in this position where OpenAI has said 'experts are checking it,' AGMAI has said 'release the paper,' and the answer from OpenAI is basically silence. I mean — this is me guessing — but that looks like OpenAI controlling the release calendar while calling the framework open.
Eliza Ward: Yeah, and I think that's — actually, no, I don't want to overstate what we know. The sourcing doesn't tell us why the paper's still sealed. It just tells us it is.
Brian Reed: Fair. But the gap is real whether or not we know the motive.
Eliza Ward: The concrete thing to watch: does the Yang-Mills paper get released before verification finishes, or does OpenAI get to claim the discovery and set the publication date? That's the actual test of whether this framework means anything.
Brian Reed: But the benchmark story — that's where I think the circulating take is actually just wrong. Not shaded wrong. Wrong.
Eliza Ward: The 'OpenAI beat Anthropic' headline.
Brian Reed: GPT-6 Sol, 60.5% on OSWorld 2.0. Claude Opus 5, 60.3%. That's the headline. OpenAI wins. Except — hang on — Sol ran at 'xhigh' effort. Opus 5 ran at medium. Those are not the same test.
Eliza Ward: And the 80% lower cost-per-task claim comes from OpenAI's own comparison. Not Gartner, not an independent lab. OpenAI's numbers, comparing their xhigh run to Anthropic's medium run.
Brian Reed: Is this a technicality though? Like — does effort level actually change the conclusion, or is this just how benchmarks work and everyone knows it?
Eliza Ward: No, I — wait. OSWorld 2.0 is evaluating agents operating graphical environments, apps, websites. Real interface interaction. Higher effort almost certainly means more compute per task. So if you're comparing cost per task, effort level is exactly what changes the math. You can't just — you can't separate those.
Brian Reed: And Arun Chandrasekaran at Gartner flagged GPT-6.1 Terra's latency advantage, but also said enterprises need caution in agentic workflows. Which is — I mean, that's not exactly a ringing endorsement of the benchmark narrative.
Eliza Ward: And then there's the naming problem. Sources are calling it GPT-6 Sol, GPT-6.1 Terra, GPT-6 broadly — we actually don't know with certainty whether the model that scored 60.5% is the same model the price cut applies to. That's not resolved.
Brian Reed: Competitive parity is probably the honest description. Not a win. But honestly — what this double announcement actually means for enterprise buyers, and whether token price is even the right lever here, that's the strange part. We'll get there.
Eliza Ward: And the stranger part is — Google's not even at the table anymore. Google killed its standalone Gemini Computer Use on July 28th. Folded everything into Gemini 3.x Flash, still labeled Preview. So the field OpenAI is pricing into has already shrunk.
Brian Reed: Wait — July 28th? That's three months before this announcement.
Eliza Ward: Three months. So this isn't a three-way race OpenAI is winning. It's a two-way race, and Amazon Nova Act is also in it. Which — I mean, that actually makes the selection problem worse for buyers, not simpler.
Brian Reed: Right — because the price cut looks decisive until you realize you're choosing between OpenAI, Anthropic, and a third option in Nova Act, with Google having just quietly exited. The sticker price is the first number anyone sees, but let me put this concretely. Say you're a procurement lead at a mid-size logistics company, you're evaluating computer-use APIs this week. The 50% cut on GPT-6.1 Terra grabs you. Except — hang on — your actual cost exposure isn't per-token list price. It's how many tokens your agent loop burns. And a badly designed planning loop can burn ten times what a well-structured one does at the same task.
Eliza Ward: And that's what analysts have actually flagged — planning efficiency, context management. Those are the real levers, not the list price.
Brian Reed: So she's looking at two numbers that don't answer her question. GPT-6.1 Terra's latency advantage over Sol and Astra is real for certain agentic tasks — Chandrasekaran at Gartner said that. But 'certain tasks' is doing a lot of work there. If her workflow doesn't fit that task design, the latency edge disappears and the cost math changes completely.
Eliza Ward: No, that's — actually that's the thing I don't think the price war story accounts for. The cut creates urgency without creating clarity.
Brian Reed: And this is me reading analyst commentary, not certainty — but the working interpretation is that a 50% price cut compresses margins on high-volume deployments before enterprises even have the tooling to know whether their planning loops are efficient enough to benefit.
Eliza Ward: The concrete thing to watch: does any independent benchmark — not OpenAI's comparison — run GPT-6.1 Terra and Claude Opus 5 at matched effort levels on OSWorld 2.0. That's the number that would actually move a procurement decision.
Brian Reed: So where does that leave the science side? As of October 28th, none of the three results — the complexity theory, the Yang-Mills work, the quantum field theory — none of them had independent verification completed. And the Millennium Prize carries a million dollars per problem. That's not a ceremonial prize. Clay Mathematics Institute set that up in 2000 precisely because these problems are — I mean, they're the hardest open questions in math. If OpenAI's AI actually cracked any part of Yang-Mills, that's the story. But AGMAI said release the paper. It hasn't been released. So what do we actually have?
Eliza Ward: Partially claimed. That's — wait, is that even a category? It might be now. If reviewers come back and say the complexity theory result checks out, but the Yang-Mills manuscript is still sealed — does 'one of three verified' become the new bar for announcing a breakthrough?
Brian Reed: And I genuinely don't know the answer to that. We don't know yet.
Eliza Ward: Neither do I. The open question — the real one — is whether OpenAI can hold the release calendar on a Millennium Prize Problem result, have external experts bless the process, and still call it open. And whether the 50% cut on GPT-6.1 Terra means anything to a buyer before that credibility question is settled. Those two things are connected and neither is resolved.
Brian Reed: That's an honest place to stop.
