The IMO moment: how OpenAI beat DeepMind on comms — and what it tells us about AI research, Figma's IPO math, and the energy ceiling

At 1 a.m. on Saturday, 19 July 2025, OpenAI researcher Alexander Wei published a short note claiming his team had reached gold-medal-level performance on the 2025 International Mathematical Olympiad. By Monday morning, Google DeepMind's co-founder and CEO Dennis Hassabis was on the record too: his group had solved five of six IMO problems for 35 of 42 points — comfortably over the gold threshold. Both labs had, on the same weekend, produced the first credible AI performances at the level reached by elite pre-college prodigies.

The technical achievement is genuinely new. According to OpenAI's Noam Brown, the model that produced the IMO proofs is a general-purpose reasoning LLM — no Lean formal verifier, no Python interpreter, no internet, no code interpreter — operating under the same 4.5-hour constraint a human contestant faces. Brown framed the result explicitly: "We achieved a milestone that many considered years away, gold-level performance. Typically, for these AI results, like in Go, Dota, Poker, Diplomacy, researchers spend years making an AI that masters one narrow domain and does little else. But this isn't an IMO-specific model." It is, he emphasised, a technique for scaling reinforcement learning on hard-to-verify tasks — proofs that take human experts hours to grade — and the model itself thinks for hours.

The DeepMind system, by Hassabis's own account, scored 35/42 and was held back from publication at the request of the IMO board until after official verification and the student closing ceremony. That is the polite move. It is also the move that lost DeepMind the narrative. As one TBPN host put it: "Speed is greater than bureaucracy. Miss the moment, lose the narrative." For a roughly 36-hour stretch over the 19–21 July 2025 weekend, OpenAI owned the frame. DeepMind's Monday-morning announcement was received as confirmation rather than revelation — the second team to cross the tape rather than the first.

What the result actually proves

The substance is worth slowing down for. Brown's distinction matters because it is the difference between an IMO demo and a research breakthrough. Single-domain reinforcement-learning triumphs — Go, Dota, Poker, Diplomacy — involve years of work on one task the AI masters and almost nothing else. The OpenAI result, on Brown's framing, is general: the same reasoning machinery should transfer to other hard-to-verify problems, including competitive programming at the level of AT Coder.

Terence Tao, the Fields medalist, spent a long X thread after the announcement warning that direct AI-vs-AI comparisons on IMO depend heavily on the scaffolding permitted. Time acceleration, calculator access, team-of-six collaboration, leader prompts — any of these change the difficulty. Tao's 2025 gold threshold of 35 of 42 effectively requires solving five of six problems near-perfectly. Self-selected test methodology, he argued, makes apples-to-apples comparison unreliable. That caveat is fair, and it complicates the easy takeaway that one lab is "ahead." Both labs claim gold. Neither has yet published the kind of reproducible evaluation framework that would let outsiders adjudicate.

Still, the technical signal is clear: reasoning models that can think for hours, using chain-of-thought extended over long horizons, are now operating at IMO-gold level without a verifier in the loop. One industry analyst quoted on TBPN pegged the next-generation reasoning model that could do this work at roughly $20,000 per month per user. Will Depue pushed back on framing that as disqualifying: "Capability per dollar will drop 100X a year. The 3K task arc AGI 80% could probably be $30 if we cared to optimize it. Repeat after me, all that matters is top line intelligence." Whether you take Depue's exponential seriously or treat it as venture-marketing, the cost curve on these models is moving in one direction only.

The DeepMind approach leans on Lean. The OpenAI approach does not. That is a real technical disagreement about how to certify machine-generated proofs. It is also, in 2026, a live debate about whether formal-verification scaffolding is the future of mathematical AI or a transitional crutch.

News velocity as competitive moat

The comms gap is the more lasting lesson. OpenAI shipped its result on Saturday morning, before the IMO board had formally closed the contest for human students. DeepMind waited. By the time Hassabis posted, social media had already moved on from "an AI did it" to "did it without tools" to "what does this mean for AGI timelines." The DeepMind announcement was still impressive — but it was processed as confirmation rather than news.

Consider what that means in practical terms. Researcher Alexander Wei's Saturday post functioned as a one-line embargo-breaker. Brown followed with a thread. By the time DeepMind posted on Monday, OpenAI's framing had already entered the discourse: tool-free, general-purpose, hours-of-thinking. DeepMind's framing — respect for the IMO board, alignment with the math community — never got comparable oxygen. It will be remembered as the polite runner-up.

This is not unique to AI labs. It is the same comms discipline that decides which IPO gets the cover story and which gets the back-page mention. Whoever moves first sets the rubric; everyone else is judged against it. The IMO result illustrates, in miniature, a structural fact about 2026 tech coverage: speed is the moat. The research lead may belong to DeepMind on some metric, but the perception lead now belongs to OpenAI.

Figma, Polymarket, and the broader IPO tape

The same morning on TBPN also brought two other market-moving stories that belong in the same frame. Figma filed its S-1. Jamin Ball of Altimeter Capital walked through the numbers: roughly 49% revenue growth, roughly 28% free cash flow margins, on a near-$1B run rate. Add the two and you get a rule-of-40 score of approximately 77 — the highest among roughly 80 public software companies Ball tracks. That is a market the public side has not seen since ServiceTitan.

Figma's secret weapon is not the growth rate. It is the roughly $1B breakup fee Adobe paid when the 2023 acquisition collapsed. That cash sits on the balance sheet as a non-dilutive war chest, which is exactly what gives the company pricing power and margin durability the street is now rewarding. Pricing was expected Wednesday, trading Thursday, on the schedule discussed 28 June 2026.

Polymarket, for its part, announced it acquired QCEX for $112 million to re-enter the US market after years of being inaccessible to American users. The platform has been a staple on the TBPN ticker; the deal is the legal mechanism for operating inside the United States after founder Shane Coplane's prior DOJ exposure. The significance is bigger than the sportsbook crowd. A US-legal Polymarket becomes a real-time sentiment data layer for business and political decision-making — institutional money will treat it as such.

The two stories together sketch a market pattern. Quality software assets are repricing on rule-of-40 excellence; prediction markets are recovering legal access to the US; and AI labs are competing on the velocity of their disclosure cycles, not only on the underlying benchmarks. Each of these is a separate story. Together they describe the same animal: a 2026 capital market that rewards speed, scarcity, and operating leverage simultaneously.

Rare earths, energy, and the binding constraint

The week's other big tell came from the supply side. Chris Buskirk of 1789 Capital argued on TBPN that China played its rare-earth export card "too early" during the tariff negotiations, opening a window for US onshoring that Apple has already begun to use. Apple signed a roughly $500M deal with MP Materials for US-produced rare earths, per Wall Street Journal reporting cited on the broadcast. Vulcan Elements is another US rare-earth startup Buskirk's firm is backing.

The structural read is that China spent maximum leverage on a tariff negotiation. Had Beijing waited until a Taiwan contingency, rare earths would have been a far more effective chokepoint. Instead, the US now has time and capital to build ecosystem capacity. Mariana Minerals, out of stealth on the same broadcast, raised a $65M Series A led by Coastal Ventures and Breakthrough Energy Ventures to attack the subscale gap between junior miners and majors — a software-first, vertically integrated approach to the post-exploration-to-production handoff.

Buskirk's broader thesis: "The civilization that has access to the most abundant, cheapest energy wins, period, full stop." US AI compute demand by 2035, he argues, will require more power than the country currently produces. Energy is the binding constraint on AI, crypto, defense, robotics, and space simultaneously. China, in his framing, is "just copy and pasting nuclear reactors" while US startups are shackled by permitting.

That may be a partisan read of the regulatory landscape. It is, however, a structural read of where the bottleneck is. The labs that hit IMO gold this month did so by burning serious compute on long-horizon reasoning. The next plateau — AGI, as one TBPN guest defined it, "an AI that can learn to run a gas station for a year without a team of scientists collecting the gas station data set" — will burn more. Whoever solves the energy side first solves the AI side first.

What the comms race really signals

Strip out the noise and the IMO weekend tells a clean story about how AI labs now compete. Three layers are visible.

First, the technical layer. Both OpenAI and DeepMind can put a general-purpose reasoning model over the IMO gold bar. That bar is higher than AIME and harder to game. Brown's claim that this is not an IMO-specific model — that the technique generalises to hard-to-verify domains — is the more important technical statement of the week, and OpenAI got to make it first.

Second, the disclosure layer. DeepMind respected the IMO board's request and lost the narrative. OpenAI did not wait. In a market where the difference between "first" and "also-ran" is roughly 36 hours of social media attention, that is a meaningful strategic cost. As one TBPN host summed it up: "It's hard to overstate the significance of this. It may end up looking like a moon landing moment for AI. A next-word-prediction machine — because that's really what it is here — no tools, no nothing, just produced genuinely creative proofs for hard novel math problems at a level reached by only an elite handful of pre-college prodigies."

Third, the perception layer. The perception of who leads in frontier AI research is now a moving target driven by news velocity as much as by benchmarks. DeepMind's bureaucratic respect for the IMO board is defensible on the merits and costly on the frame. The next time both labs cross a tape on the same weekend, the question will not be whether the result is real. It will be who got to say so first.

That is the new research lead. It looks suspiciously like a comms lead. And the two are now inseparable.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://www.youtube.com/watch?v=7-DCCfzTqEw