OpenAI’s Sub-n log n Multiplication Claim: Why It Matters—but Won’t Speed Up Your Laptop

OpenAI’s Sub-n log n Multiplication Claim: Why It Matters—but Won’t Speed Up Your Laptop

OpenAI’s 23 September 2026 preprint, later released in its mathematics collection on 6 October 2026, makes a striking claim: a deterministic algorithm for multiplying n-bit integers in O(n(log n)^(1-kappa)) bit operations, with an initial kappa of 2^-182, in specified theoretical computation models. That is a genuine asymptotic improvement over the long-standing n log n frontier. It is not, however, a 2^182 speedup, and it does not mean your laptop, Python, CPUs, or cryptography suddenly got faster.

What the claim actually says

To understand the excitement, it helps to translate the notation into plain English. Big-O notation describes how an algorithm grows as inputs become huge. If one method takes roughly n log n steps and another takes roughly n(log n)^(1-kappa) steps, the second grows a little more slowly in the limit. Little-o is even stricter: it means one growth rate becomes negligible compared with another as the input size increases. So “below n log n” does not mean “dramatically faster in practice right away.” It means “asymptotically smaller” in a mathematical model.

The exponent improvement is tiny by design. kappa = 2^-182 is extraordinarily small, so the gain is far more important as a proof of principle than as a practical recipe. The headline is about crossing a theoretical boundary, not about shaving milliseconds from ordinary calculations.

Why n log n looked like a barrier

For decades, integer multiplication was a central benchmark in complexity theory. Classic methods improved steadily, and by 2021 Harvey and van der Hoeven had proved an O(n log n) algorithm. That result was celebrated because many researchers regarded n log n as a natural frontier: not a law of nature, but a boundary that seemed hard to move below. The new preprint claims to push under that line, in a model-specific asymptotic sense, which is why it attracted immediate attention.

That does not automatically make the new claim correct, complete, or the final word. In theoretical computer science, major advances often require careful scrutiny before the community accepts them. The more surprising the statement, the more important the verification process becomes.

What David Harvey said on 20 October 2026

The freshest expert signal is David Harvey’s 20 October 2026 note. His assessment is cautious but important. He says the claimed overall shape looks plausible, but he also says he has not checked most details, has not read all three papers in full, and has identified major unchecked work, including the costs of data rearrangement. He further states that a formalized proof of the integer-multiplication result was not yet available to him.

That distinction matters. Harvey’s note is not an endorsement that every step is settled. It is an expert’s early read that the broad direction may be right, while the proof still needs line-by-line validation. In other words: plausible, yes; independently confirmed, no.

Why the Lean formalization does not settle the main claim

One subtle point is easy to miss. A companion Fourier-circuit result has a Lean formalization associated with it, but that is not the same as an independently formalized proof of the main integer-multiplication theorem. Formal verification of one component does not automatically certify the entire argument, especially when the full claim depends on additional reductions, model assumptions, and cost accounting.

So if you see “formalized” mentioned in connection with the broader story, check exactly what was formalized. Was it the exact main theorem and model, or only a related subresult? Those are very different standards of evidence.

Why this will not speed up your laptop

Here is the practical bottom line: this result does not currently make laptop multiplication faster. It does not change Python’s arbitrary-precision integers, GMP or FLINT deployments, CPU multiplication pipelines, cryptographic libraries, or AI inference kernels. The reason is simple. This is an asymptotic theorem in a theoretical model, not a production implementation with known constants, crossover points, and benchmarked performance.

In computer science, an algorithm can be asymptotically better and still be useless at realistic sizes if the hidden constants are enormous or the threshold where it wins is astronomically large. That is a common pattern. Without an implementation, explicit constants, and empirical benchmarks against established libraries, there is no basis for claiming a practical speedup.

There is also no known production system replacing standard integer-multiplication software because of this preprint. The everyday tools used in real systems are still the established ones, because they are engineered, tested, and fast at the sizes that matter now.

Why theorists still care a lot

Even if it changes nothing immediate in software, the result would matter deeply if it holds up. It would challenge a long-standing perceived boundary, redirect research questions, and likely influence work on FFTs, circuit complexity, data rearrangement, and lower bounds. In theory, moving below a famous frontier can reshape what researchers think is possible and which techniques deserve attention.

It also matters because integer multiplication is not a toy problem. It sits near the intersection of algorithms, number theory, and computational models. Progress here often reveals methods that later prove useful elsewhere, even if not in the obvious product-facing way.

Headline translation

1. “Below n log n” -> an asymptotic statement in a theoretical model. 2. “Plausible” -> not independently verified. 3. “Tiny exponent improvement” -> theoretically boundary-breaking, practically negligible today. 4. “AI-generated mathematics” -> requires the same or greater scrutiny, not automatic rejection or acceptance.

What would count as stronger evidence next?

For readers trying to separate a promising theorem from a fully accepted result, the next evidence should look like this:

• Independent line-by-line review by experts who are not authors of the preprint.
• Formalization of the exact main theorem and the exact computation model, not just a related subresult.
• Reproduced proof artifacts that others can inspect and check.
• Explicit constants and crossover estimates showing where, if ever, the asymptotic gain might matter.
• An implementation of the algorithm in real software.
• Benchmarks against established arbitrary-precision libraries such as GMP or FLINT, using meaningful input sizes.

Until that exists, the right posture is neither hype nor dismissal. It is measured skepticism: the claim may be profound, but it is still a claim under review.

Sources

• OpenAI official overview: https://openai.com/index/sharing-ai-progress-in-mathematics/
• OpenAI preprint: https://github.com/openai/math/blob/main/preprints/Integer-multiplication-below-n-log-n-September-23-2026/paper.pdf
• David Harvey’s 20 October 2026 expert note: https://web.maths.unsw.edu.au/~davidharvey/research/subnlogn.pdf
• Harvey–van der Hoeven 2021 Annals paper: https://annals.math.princeton.edu/2021/193-2/p04