Rumor Check: Did Games and Simulators Make OpenAI and Anthropic Internal Models Better at Math?

On October 28, 2026, an X account posted what it called rumors: that more aggressive reinforcement learning (RL) on games and simulated environments explains why internal models at OpenAI and Anthropic produced results on problems that had resisted decades of work, and that Tesla's Digital Optimus hiring shows the same bet. The post had about 5,700 views when we read it, and the replies were mostly polite requests for evidence.

Part of the thread is solid: both labs did publish headline results this week. The part that matters for readers, why those models got better, is not in any public document we could find. This post sorts the thread claim by claim.

A magnifier over a claim bubble marked with a green check, illustrating a claim-by-claim rumor check

Claimed vs verified #

What is solid: the ETH results #

"ETH" in the thread is the Exponential Time Hypothesis from complexity theory, not ETH Zurich. It says some problems, including 3-SAT, cannot be solved much faster than brute force.

• OpenAI announced on October 26, 2026 that an internal model refuted it, publishing a paper and Lean code in itsopenai/math repository . Our earlier write-up ishere .
• Anthropic followed within hours. Itsformal-math repository says the result "was discovered by an internal research Claude model" and that the Lean formalization "was written by AI models." Secondary coverage reports the algorithm was found with about 14 million output tokens and no human input, per the paper, which we could not confirm on the repository page itself.
• Caveats both sides share. Per aYahoo-syndicated report , OpenAI's model "remains proprietary," and the prompts, compute time and chain-of-thought outputs are not released. Russell Impagliazzo, Ramamohan Paturi and Ryan Williams are organizing an outside review, and wrote that their involvement should not be taken as certification.
• Practical impact is nil today. One analysis of the Anthropic bound says it beats brute force only for astronomically large formulas, and P versus NP stays open.

So "the internal models did something big on ETH" holds up as a claim. "Both labs found it independently the same week" is also stated by Anthropic's paper. Neither says anything about training.

The Navier-Stokes line #

The thread cites Navier-Stokes alongside ETH as proof of scale. As we explained in No, AI didn't just solve Navier-Stokes, the claim is a narrow forced blow-up result, not the Millennium Prize problem, and it has not been independently peer-reviewed. It is evidence that labs are producing formal math results. It is not evidence about method.

The core rumor: RL on games explains the jump #

Here the thread moves from reporting to explaining, and the explanation has no public source. What we checked:

1. Anthropic's ETH page says nothing about training. The only other model reference is that Claude Opus 5.5 at high effort can rederive the refutation from a condensed description, and that description is about tensor networks.
2. OpenAI's materials describe outputs and verification. Training, prompts and compute are not disclosed.
3. A search for any lab statement tying RL-on-games to these results found none.

That does not make the idea absurd. RL with verifiable rewards is a well-known driver of reasoning gains, and math and code are domains where answers can be checked automatically. A reasonable person could guess that simulated environments help. But a guess that fits the trend is not a leak. The thread offers no document, no named source and no number, and a few replies say so: one asks what public evidence separates simulator-driven reasoning from "better data and longer training," and another says gains on benchmarks do not prove transfer from games.

That reply names the real difficulty. Several things changed at once inside these labs, and a result on one hard problem cannot tell you which change mattered.

The hiring-ad evidence #

The thread's strongest-sounding clue is a job ad. We checked each one.

OpenAI, "Researcher, Synthetic RL." The listing exists, in OpenAI's careers pages and on investor job boards. Mirrors describe a team working on reinforcement learning with synthetic data, environments and feedback, exploring self-play, simulators and other synthetic evaluations. The listing appears to date from early 2026, per job-board copies, so it predates the October results by months.

Anthropic, "Research Engineer, Universes." The listing says the Universes team trains models to perform complex, long-horizon agentic tasks in ultra-realistic settings. We did not find the math-reasoning wording in the Universes ad. It appears in postings for other RL teams, so the thread blends two or more listings.

What a job ad can show: a lab is investing in that area. What it cannot show: that a given model was trained that way, or that the training caused a specific result. Labs also hire in many directions at once, and an ad describes a hope for the work, not an outcome.

A green ball follows a curved track toward a small cream cup, illustrating reinforcement learning

Digital Optimus is a different thing #

The thread adds that Musk's Digital Optimus hiring "explicitly calls for measuring transfer from games into general computer use." Secondary reports describe a Palo Alto machine learning role on the project's gaming track where engineers would work on memory, planning and real-time decisions and test whether game-learned skills carry over to other software. We could not read the original text, so treat the wording as reported, not confirmed.

Two points keep this honest. First, Digital Optimus is a pixels-in, keyboard-and-mouse-out computer-use agent, a different system from the math-proving internal models. Linking them is the poster's frame. Second, the ad lists transfer as something to measure, which is the opposite of saying it has been shown. Our own check of Musk's Diablo claim, halfway through the campaign, no video or scores shared, found the same gap between statement and evidence.

What real evidence would look like #

Skip the rumor mill and look for these, from the labs themselves:

• A system card or technical report that states the training recipe or at least the environments used.
• An ablation comparing the same base model with and without the simulator or game stage.
• Held-out transfer tests on tasks the training could not have contained, with controls, which is what one reply to the thread asked for.
• Independent replication , ideally an open model trained the same way showing the same gain.
• Independent certification of the math , since Lean checks the proof but not that the formal statement matches the claim. Anthropic's own page notes that no machine checks that the file says what the paper says.

For how mathematicians are handling the verification side, see our summary of OpenAI's math results and expert reactions and what to check in a mass release of AI math.

Why the rumor is easy to believe #

The thread has the shape of a good story: two rival labs, the same week, one mechanism, and a pair of job ads that seem to match. That fit is what should make a reader slow down. The ETH results arrived within hours of each other, which invites a shared-cause story, but Anthropic says its own model found the algorithm, and nothing in either release says the two models share a training method. Coincidence of timing, convergent research, or a shared trend in the field would produce the same headlines.

The related Anthropic news in the replies, six drug-discovery partners and a one-person biotech company, is a separate announcement. We cover what is and is not established there in Anthropic drug development: what is preclinical. It shows Claude being used in science, not how it was trained.

What this means if you build with AI #

• Do not design strategy around an unconfirmed training claim. If your roadmap depends on "RL on environments transfers to general reasoning," test it on your own tasks with held-out cases.
• Read job ads as signals of direction, not results. They tell you where a lab is spending money.
• Ask for the denominator. One headline result tells you little about how often a model succeeds on comparable problems.
• Expect more of this. With labs racing on math announcements, rumors about method will keep outrunning documents. SeeAnthropic's formal verification work andthe Claude 3SUM claim check for earlier examples where we checked claims against artifacts.

What to watch #

• Whether OpenAI or Anthropic publishes a system card or training note for the models behind the ETH results.
• The outside review of the ETH refutation and any statement on whether it is correct.
• Any controlled transfer study, from a lab or an academic group, on game-trained models and non-game tasks.
• Whether the Digital Optimus team shares transfer numbers instead of a hiring ad.

Related reading #

Details are as of October 28, 2026 and rest on public posts, lab repositories and press reports; the thread's claims about training are rumors.