Two claims came out of OpenAI in the same week, and both are dressed in the language of AI improving AI. One is that a frontier model rewrote the code that serves it, cutting the price of March's flagship intelligence by roughly thirteen times. The other is that an unreleased model cracked ten mathematical problems that had gone nowhere for a decade, for about two thousand dollars each.
Most readers will file these together as a single story about acceleration. That is the mistake.
Start with the number doing the most seductive work. GPT-5.4 full at xhigh scored 51, exactly where Luna max sits today. GPT-5.4 cost $2.50 per million input tokens and $15 per million output; Luna now costs $0.20 and $1.20.
So March's full flagship intelligence is on sale, four months later, at about one-thirteenth the token price. Annualize that and you get something like a 2,000x-per-year rate of decline for a fixed level of capability.
And OpenAI credits the model itself. In its own technical report, the company says GPT-5.6 Sol was pointed at production traffic to analyze load, tune request routing, and autonomously rewrite the Triton and Gluon kernels that serve OpenAI's own models—cutting end-to-end serving cost by 20%. A redesigned speculative-decoding draft model, built through hundreds of autonomous experiments, added more than 15% in generation efficiency.
This is the half of the week's announcement that holds up. The other half—the math—does not, at least not yet. And the reason to trust one and not the other is the same reason, applied twice.
Here is the test: can you verify the claim without taking OpenAI's word for it? Not "is it plausible," not "is it exciting," but is there evidence a skeptic could check independently.
The pricing story passes. The chain runs from a capability claim to a documented efficiency gain to a number on a public price list three weeks after the model shipped. CNBC reportedly reported the cuts based on OpenAI's own announcement. The Quantum Dispatch put the mechanism plainly: kernel optimization is hours of profiling for single-digit gains on specific silicon, and the engineers who do it well are scarce and expensive.
A model doing that work, on the stack that serves itself, with the result flowing to a customer invoice, is a short and legible loop. Benchmarks are contested; a bill is not. If the efficiency were fake, the margin math would not close and the price cut would be a subsidy—and OpenAI, facing pressure from Kimi K3 and Gemini 3.6 Flash on cost, would have every incentive to hide that it was bleeding.
Benchmarks are contested; a bill is not. If the efficiency were fake, the margin math would not close and the price cut would be a subsidy.
There is a second guardrail worth naming: the human is still in the loop. Per OpenAI's report, the model proposes and experiments, but optimization targets, tool permissions, eval metrics, and whether code reaches production remain human decisions. This is not GPT training GPT. It is a bounded, testable engineering task inside a supervised workflow—which is precisely why it is credible and why it will spread. Every lab runs an inference stack; every stack has unoptimized corners.
Now run the same test on Astra.
The claim is spectacular. An internal version of Astra, described as OpenAI's next major model, solved ten problems in mathematics and theoretical computer science, each open for at least a decade—the first improvement to the sphere-packing exponent since 1978, an explicit non-sofic group, a disproof of the Connes rigidity conjecture, Erdős's 183rd problem. OpenAI released Lean 4 formalizations, a 249-page paper, and a document reconstructing the reasoning. Greg Brockman reportedly put the total at about $2,000 at Sol API prices, though sources differ on whether that figure is per problem or for all ten combined.
That is more transparency than most capability claims arrive with. And it still fails the test, on four counts.
First, the model is unreleased. No external party can run Astra, so no external party can reproduce the result or probe how narrow the wins are. You are asked to grade the exam and trust the school that no one else can enter.
Second, the price figure is a fiction dressed as a fact. Two thousand dollars "at Sol API prices" prices an internal model's compute at the public rate of a different, released model. No external buyer pays that number because no external buyer can buy Astra. It is an internal cost translated into a retail-sounding figure for effect.
Third, there is no denominator. As Simon Willison flagged, OpenAI reports the ten problems it solved and says nothing about how many it attempted and failed. Ten successes out of ten is a very different claim from ten out of a thousand, and only one of those numbers is disclosed.
Fourth, there is no control group. Gary Marcus's charge, per his post, is blunt: half of the Astra problems can reportedly be solved by Fable, and OpenAI didn't even have a control group. If an existing model clears half the board, the marginal contribution of the unreleased frontier system is a fraction of what the headline implies—and without a baseline, no one can say what that fraction is.
Notice how neatly the two stories invert. The pricing claim ends in an invoice anyone can read. The Astra claim ends in a demonstration only OpenAI can stage—unreleased model, internal pricing, undisclosed failures, no baseline.
The context around the launch reinforces the reading. Per The Information, Altman previewed Astra in Washington this week to White House officials, cabinet members, and members of Congress. A capability meant to move policymakers, shown before it can be independently examined, is a PR release with a formal-verification garnish. Marcus's own coda—"poorly evidenced sensationalism makes for bad policymaking"—is the point exactly.
None of this means Astra did nothing. Lean certificates are real; a proof either checks or it does not. The problem is not fabrication but framing: a controlled reveal engineered to read as a capability leap while withholding everything a skeptic would need to size it.
The distinction is the whole story. Legible, invoice-grade evidence is the new bar for a capability claim, and this week OpenAI met that bar on cost and merely performed it on math.
So update your model of AI economics. The falling cost of fixed intelligence is now attached to a documented mechanism and a public price, and it is durable even if any single price point is not—the direction survives the marketing.
Do not yet update your model of AI capability. Not because Astra is unimpressive, but because you have not been shown enough to know how impressive it is.
The recursion worth taking seriously is the one that showed up on a customer bill three weeks after the model shipped—not the one previewed to cabinet members behind closed doors.