There is a moment in a race when the pack splits and you realize two different races are happening on the same track. Tuesday was that moment. Anthropic opened its Cyber Verification Program to broader access, attached a $35 million fund, and queued Mythos-class capabilities behind it. Google released Gemini 4 Argon to cybersecurity partners before anyone else. Mistral, on a stage in Abu Dhabi, announced a 1.05 trillion-parameter open-weight model and told the room it beats rivals outside China, "including cyber."
Three labs. One message: the frontier is now a security test as much as a capabilities race.
And then there is OpenAI, which published 722 mathematical manuscripts from an internal model that has never been on sale. The mathematicians are already arguing about whether it qualifies as research. The argument is worth having.
Start with the mechanism. Anthropic's Cyber Verification Program existed before today, but it was structured for a narrow set of red-team researchers. The October 6 expansion is different in scope: broader dual-use capabilities on Opus and Sonnet, opening now, with Mythos-class access to follow in the coming weeks. Alongside that announcement, Anthropic introduced the Defender Advantage Fund, 0xDAF, a $35 million pool of Claude usage credits for open-source software developers and critical infrastructure operators. That is not a press release. That is a bet that the security community, given real access to Claude's full dual-use capability set, will find and close more holes than bad actors will open.
Google's move was different in character but identical in intent. Gemini 4 Argon, the lab's next frontier model, went to cybersecurity partners before it went anywhere else, at $2 and $10 per million input and output tokens. The pricing is competitive with the frontier floor. The choice of customer is the point: Google decided that verified defenders are the right stress test for a new model before a wider rollout.
Mistral arrived at AI Everything Abu Dhabi with a 1.05 trillion-parameter model called Large 4, nicknamed Le Chonk, and CEO Arthur Mensch told the room it outperforms rivals outside China, "including cyber." The model is live on the API now. Open weights come October 27. Before then, Mensch said, a preview release with fewer safety constraints goes to specific security researchers and state authorities for testing. The cyber claim was delivered without a supporting benchmark, which is either strategically shrewd or something that will require revisiting.
The blast radius. Every serious security research organization, CISO office, and national cyber agency now has three separate program applications to file before any of them expire. The Anthropic CVP has a formal review process. Google's cyber program has vetting. Mistral's preview period has an access list. Three lanes opened simultaneously.
The pattern here is not accidental. The labs have watched AI-enabled cyberattacks mature long enough to know that the asymmetry between offense and defense is growing. A model that can draft a working exploit can also patch the vulnerability it identified. The question is who gets to the model first and with what intent. All three labs answered Tuesday with the same instinct: the verified defender gets priority.
The read. The frontier stopped treating cybersecurity as a vertical market. It is now treating it as a trust-building infrastructure test. Every lab that puts a frontier model in front of a verified defender and asks "what can you do with this" is collecting real data about the gap between capability and harm that no benchmark evaluates cleanly. This is live safety research at production scale. The labs have decided they need genuine adversarial conditions to close the gap between what their models can do in a lab and what they do in the world.
The builder's move. If you do legitimate security research, Tuesday opened three lanes. Apply to the Anthropic CVP at anthropic.com. Watch the Google Gemini 4 Argon cybersecurity partner program. Mistral Large 4 is live on their API now, no waitlist, at $1.36 per million input tokens.
On October 6, OpenAI published 722 mathematical manuscripts produced by an unreleased internal frontier model, organized into 372 problem families, in a public GitHub repository alongside Lean proof formalizations and reasoning summaries.
The scale is worth sitting with. This is not ten results from a named flagship model, the way August's Astra announcement was. This is a dump of 722 papers from a model OpenAI has not released, spanning pure mathematics, theoretical computer science, and mathematical physics. The problems addressed are ones that have been open for substantial periods. A "family" groups a principal result, companion arguments, consequences, and alternative proofs into a single entry.
The Lean proofs matter as much as the results. Lean is a programming language that lets a computer verify every logical step in a mathematical proof without taking any claim on faith. Many of the 722 manuscripts include Lean certificates. Not all of them do. OpenAI's own README flags that the unformalized results may have issues, and the lab says it will work to add formalizations as they are obtained.
Scientific American's coverage quotes mathematicians raising concerns about research process and academic norms. The tension is genuine: these results were not submitted through standard journal review, not co-authored with the mathematical community in the structured way OpenAI's earlier Astra releases were, and a meaningful portion lack the computer-checkable verification that would let the community confirm the claims independently. Whether that makes this a breakthrough in openness or a stress test on peer review is a question academic mathematics will spend the next year answering.
The contrast against the security story above is exact. Anthropic is expanding access with gates, with a funded mechanism, with explicit review of who gets what capability level. OpenAI published 722 manuscripts from an unreleased model onto GitHub and asked the mathematical community to sort it out. Both approaches have a logic. OpenAI's logic is speed and radical openness. Anthropic's logic is care and verification. They are not the same instinct, and the communities on the receiving end did not ask for the same thing.
The builder's move. If you work in formal methods, pure mathematics, or theoretical CS, the repository is public on OpenAI's GitHub. The README carries the caveat. Treat unformalized results as claims that need independent verification, not as settled conclusions.
The architecture in detail: 1.05 trillion parameters total, 49 billion active per token, Mixture of Experts with a 1.6 billion vision encoder built in. A 1 million-token context window. Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's European datacenters, on data covering 160 languages including all official EU languages.
The API is live at $1.36 per million input tokens and $4.18 per million output tokens, which puts Mistral below the pricing floor set by the OpenAI and Google flagships. Open weights ship October 27. Before that, a preview release with fewer safety constraints goes to vetted security researchers and state authorities for testing, which tracks the security-first access pattern every other lab adopted today.
What Mistral is asserting: that a European lab can build a 1 trillion-parameter open-weight multimodal model, train it on European infrastructure, and remain cost-competitive with the closed-weight giants. If the October 27 weight release validates the API benchmarks, it will be a significant open-weight milestone.
Two items from xAI in the window. First: grok-voice-transcribe-1.0 reached end of life on October 2 and is now routing all requests to grok-voice-transcribe-2.0 at the same price with higher accuracy. The cutover is silent. If you were calling 1.0, you are now on 2.0 without touching a line of code.
Second: Grok 4.7, released September 21 as xAI's current frontier model, ran a 50 percent discount in Ramp Router through October 6. That window closes today. If you have an active Ramp subscription and wanted to test Grok 4.7 at half price under real workloads, the deadline is now.
Anthropic Sonnet 5.5 is expected in the coming weeks. Anthropic confirmed its existence on September 22 and said it would follow Opus 5.5 "in the coming weeks." No announced date, no price, no benchmarks yet. Several September sources placed the window at late October. The CVP expansion and Defender Fund suggest Anthropic's attention is in security territory this week, but a Sonnet release can drop independently of the policy calendar.
OpenAI's unnamed internal model produced today's 722 manuscripts. It is not Astra, the model that produced the August "Ten Advances" results. A second, apparently more capable internal model appears to have been in training since at least late August. What gets released commercially from it, and when, is the open question. Given that OpenAI has now published results from two unreleased internal models in three months, the release cadence of whatever comes next is probably shorter than the usual cycle.
Anthropic IPO: Pre-IPO investor day is scheduled for October 14. Anthropic is targeting a November listing, per Bloomberg reporting from October 1.
Regulatory: The FTC confirmed it is investigating OpenAI, Anthropic, and other labs and plans to compel executives to testify. No timeline given.
Daily digest at 9 PM ET. Weekly magazine every Friday morning. Six labs, one feed. No spam, one-click unsubscribe.