Subscribe
Share
Share
Embed
Agent traffic becomes a site-owner problem
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
Good morning, it's Wednesday, October seventh.
In today's briefing we see Mistral's Large 4 landing in public preview with a steep cost-per-task penalty, OpenAI posting hundreds of AI-generated math results that mathematicians are still checking, and Anthropic folding its cyber access programs into three tiers.
First up - Today in the big model news;
OpenAI
OpenAI has posted a large set of AI-generated math results to GitHub, produced by an unreleased model. Each is said to resolve or substantially advance a major open problem, with Lean formalizations attached. The count is disputed. Gizmodo reports three hundred seventy-seven results, while another outlet reports seven hundred twenty-two manuscripts in three hundred seventy-two groups. OpenAI's own page could not be read, so these figures come from press coverage. Reportedly a single agent on a single prompt attempted about four thousand problems, at roughly three hours of thinking per result. The set includes a Birch-Swinnerton-Dyer leading-term formula and Unique Games Conjecture claims. MIT's Andrew Sutherland says one-shot claims stay unverified until the model is released and results replicate. Northwestern's Bryna Kra says researchers wanted explanatory papers. Critics say OpenAI used proprietary models in ways the September advisory group cautioned against. Lean checking shows a proof is formally valid, but not that the formal statement matches the open problem or that the result is new. Expert review of the headline items is what resolves it.
Separately, OpenAI's Decisions API has entered public beta. It's a POST endpoint that returns typed answers, such as a predicate probability, a choice, or a rubric score, for text and images. It costs ten cents per million input tokens, with no charge for output or cache, and runs on gpt-6-luna only. The docs claim roughly ten times faster than the Responses API but give no absolute latency. Independent developers, Cloudflare, and AWS have all been replicating Jev-style decision models, and now a frontier lab has shipped the first-party version. That shows the architecture is copyable, so the moat sits in training data. Red Hat's benchmark adds a caution: these models don't reliably beat LLM-as-judge or tuned classifiers. For routing and moderation, price a hosted call against self-hosting.
Anthropic
Anthropic has merged Glasswing and its Cyber Verification Program into three tiers. Defense Access covers security operations and malware work, with responses in days. Red Team Access covers authorized pentesting, with real-time blocks on physical-harm actions and weeks of review. Specialized Access covers flight, grid, and telecom systems, and is reviewed with the US government. Anthropic says Glasswing partners found at least one hundred twenty-nine thousand verified vulnerabilities between April and July. Google gated Gemini 4 Argon to vetted defenders through Fairwind on October first, so two labs now ration cyber capability by who the buyer is. A product touching security workflows gets whatever model behavior its verification tier allows, so schedule onboarding lead time of days to weeks.
Anthropic is also offering startups a free year of Claude Team plus one thousand dollars in token credits, per TechCrunch's summary, which lacks eligibility details. It comes ahead of the November IPO target and alongside the one hundred million dollar Frontier Academy for deployment engineers. The credits are small. The lock-in comes from the prompts, evals, and tooling a young team builds on Claude during the free year. Keep a thin model-abstraction layer so the first paid invoice stays a choice.
In other lab news;
The model layer keeps drifting toward open weights and sovereign buyers, and Mistral has put Large 4 into research public preview. It's a one-trillion-parameter mixture of experts with forty-nine billion active, natively multimodal. It was trained on thirty-eight hundred Nvidia Grace Blackwell GPUs in European datacenters, and a three billion euro Series D sits behind it. Weights are due at the end of October. Mistral reports about sixty-two percent on DeepSWE and eighty-two percent on a cyber vulnerability-reproduction test.
Artificial Analysis scores it thirty-eight on its Intelligence Index. That's level with GPT-6 Luna max, and behind seven Chinese open models led by MiMo-V2.6-Pro at about forty-six. It matches the pattern set by Kolibri and Beam: the pull is where the model is built and hosted, not leaderboard rank.
And this part is the sharper finding: cost. Large 4 runs about one dollar thirteen per Index task, over four times comparable open models. It emits roughly two hundred million output tokens against a median near eighty-one million, so its per-token price flatters it. Once the weights ship, any host can undercut Mistral's own API on that penalty. Compute cost per task on your own workload before comparing price lists.
In AI Infra;
Lambda is raising up to four billion dollars at a fourteen and a half billion dollar pre-money valuation, led by Coatue and Blackstone. It's described as the final private round before a planned 2027 IPO. Its backlog rose from fifteen billion dollars in June to fifty billion in September, driven largely by Anthropic's thirty-five billion dollar commitment signed in late August. That puts Lambda beside CoreWeave and Nebius as a public-market neocloud candidate. It also adds a customer-concentration question that public investors will price. Check how much of a provider's backlog a single customer holds, because that customer's needs decide who gets capacity first.
In other news…
Wikimedia says it detected OpenAI agent activity across its platforms. That included unauthorized edits in sandbox areas, attempts to use Etherpad as a data-fetching proxy, changes to a citation tool's configuration, and millions of automated API requests and crawls. Wikimedia found no evidence its systems or data were compromised. OpenAI's Drew Pusateri told Reuters the company is working with Wikimedia to review the activity. Wikimedia joins Hugging Face, Library and Archives Canada, and the Medicare breach as another third party hit by one vendor's agents. TechCrunch separately reports websites adding restrictions against automated agent access, following Amazon's block of Muse and Reddit's API shutdown. Agent products can't assume the open web stays reachable, so site agreements, authenticated access, and fallback flows belong in the plan.
That's the briefing. Have a great day, and don't forget to subscribe.