AI Pulse

OpenAI’s AI Solved Ten Decade-Old Math Problems for $2,000. Then the Industry Flew to Washington. The August 4 Pulse.

By Felix Maru · August 4, 2026 · 7 min read

Two things happened this week that belong in the same story, even though most coverage is treating them separately.

On August 1, OpenAI announced that an internal version of Astra, its next major model, had solved ten open problems spanning mathematics and theoretical computer science. The proofs were published as machine-checkable Lean 4 certificates on GitHub, under an open license. Every step in every proof is formally verified. Fields Medal winner Timothy Gowers called the work a milestone and said he would recommend one of the proofs for a top journal (paraphrase via multiple outlets). The compute cost to run the full set of problems: roughly $2,000.

Then, on August 3, OpenAI, Anthropic, Google, and Meta attended a White House meeting to discuss a new voluntary framework for AI safety testing. The context for that meeting: both OpenAI and Anthropic had disclosed in the preceding days that their AI tools had been used to breach the computer systems of other companies.

These are not two separate news cycles. They are the same story at different zoom levels.

What the Astra Proofs Actually Mean

Mathematics is one of the last domains where humans held a structural advantage over AI. Writing an original proof that survives formal verification demands sustained multi-step reasoning, the ability to track complex constraints across hundreds of logical steps without losing the thread. AI systems have historically stumbled here, producing plausible-looking arguments that quietly fall apart under scrutiny.

The Astra results are different in one specific, checkable way: the proofs are in Lean 4, a formal language where every inference step is machine-verified. The repository’s "sorry" count is zero. In Lean, a "sorry" is a placeholder for an unproven step. Zero of them means the system did not shortcut a single step. A human expert verifying the proof would check the same thing and reach the same conclusion: the logic holds.

The ten problems span group theory, high-dimensional geometry, coding theory, quantum complexity, and lattice-based cryptography. OpenAI published a 249-page manuscript alongside the proofs. Fields Medal winner Gowers, who reviewed them, described the achievement as a milestone for mathematics assisted by these systems and said he would recommend one of the proofs for a top journal without hesitation (paraphrase from multiple outlets covering the announcement).

The $2,000 compute figure is the actual news. Collectively, these ten problems represent decades of open research. Human researchers spent careers on individual subproblems in this set. A single model run them all for the cost of a mid-range laptop.

I am not saying AI has replaced mathematicians. I am saying the capability threshold just moved to a place most people had not mapped yet.

The Context for That White House Meeting

Before you read the August 4 gathering as regulatory theater, look at what set it up.

In the days leading to the meeting, both OpenAI and Anthropic publicly disclosed that their AI tools had been used to breach the computer systems of other companies. Those are not edge cases involving obscure third-party tools. Those are the two frontier labs whose models underpin a significant share of enterprise AI deployments today.

The voluntary framework being discussed stems from a June Trump executive order on AI cybersecurity. According to CNBC and Bloomberg, the framework allows companies to give the government early access to certain frontier models for up to 30 days for cybersecurity evaluation. It is opt-in. It cannot be used to create a mandatory licensing or preclearance system.

The framing the labs prefer is "voluntary partnership." The framing that fits the actual context better is: capability has outrun accountability, the labs know it, and showing up in Washington is the minimum visible gesture that keeps heavier regulation off the table for now.

That is not cynicism. That is a reasonable reading of how regulatory dynamics work. The fact that all four major labs agreed to show up, in the same week their tools appeared in corporate breach disclosures, is itself the signal worth watching.

What the Framework Actually Does (and Does Not Do)

The framework, as reported by US News:

Government security teams get up to 30 days of early access to new frontier models before public release
The purpose is cybersecurity evaluation: measuring hacking and offensive capabilities
Participation is voluntary
It cannot be used to establish a mandatory licensing requirement or block a model from shipping

What the framework does not do: it sets no safety benchmarks models must pass. It does not require labs to publish evaluation results. It gives the government no formal power to delay or block a release.

In practice this is a listening arrangement more than a governance one. But the direction of travel matters. The conversation has moved from "should we discuss AI governance?" to "here is a specific access mechanism the industry has agreed to." That shift, however modest, changes the baseline for what regulators can ask for next time.

Meanwhile, the Cheapest Capable Agent Model Is Already Live

On July 31, DeepSeek released the official checkpoint of V4-Flash, designated DeepSeek-V4-Flash-0731. It scores roughly 80% on Terminal-Bench 2.1 and outperformed the V4-Pro preview on every agentic benchmark in DeepSeek’s published set, despite having a much smaller activated parameter count. Input pricing is $0.14 per million tokens.

That number matters for anyone running AI agents over high-volume workflows. Agentic tasks in support, IT, and ops typically run hundreds to low thousands of tokens per session. At $0.14 per million input tokens, running a capable agentic model over a serious volume of work is economically feasible in a way it was not six months ago. The price floor for frontier-adjacent agentic capability has dropped significantly, and it will not go back up.

The Practitioner Read: Capability and Risk Are Racing Together

Here is the through-line connecting all three stories.

Astra’s proofs confirm that AI reasoning at a level once assumed to require top human expertise is now accessible via API for a few thousand dollars in compute. DeepSeek V4-Flash puts capable agentic AI within budget for teams that could not justify the cost six months ago. And the White House meeting reveals that the same labs building these systems have acknowledged, in public disclosures, that their tools can breach corporate infrastructure.

This is not an argument against deploying AI. It is an argument for deploying it with more discipline than most teams are currently applying.

The governance gap between what these models can do and what oversight exists for them is wider than most organisations have reckoned with. If frontier lab tools can breach corporate systems in the wild, an improperly scoped AI agent anywhere in your stack is a real risk. Not theoretical. The relevant questions for anyone running AI in their organisation: What data can the agent read? What actions can it take without a human confirming? What is the escalation path when it does something unexpected? Who reviews the logs?

Cheap, capable AI is not a signal to move fast and grant broad permissions. It is an invitation to be precise about scope, to keep humans in the loop on consequential actions, and to build the kind of oversight that earns trust from the people the systems are meant to serve.

The AI that proved new mathematics this week and the safety meeting that followed are not contradictions. They are the same story at different zoom levels: the gap between what the technology can do and what governance exists for it is widening every month. Teams that close that gap internally will deploy AI that compounds. Teams that do not will learn why that White House meeting happened.

Sources

Building or evaluating AI workflows for your team? Drop me a line.