OpenAI and Anthropic try to slow down AI progress

what to know for now

🗣️ Amodei and Altman both spent the month arguing the industry needs to slow down, while the public kept souring anyway. Investor Gavin Baker claimed on the All-In podcast and on X that Amodei’s risk warnings fed the backlash against AI and data centers specifically, and that he’d “lost the argument” on regulation. Amodei’s reply was that his messaging has been “about equally balanced between risks and benefits,” that the real problem is how “ordinary people don’t trust companies, governments, or the tech industry and always suspect that we are cooking up some new way to screw them over,” and that the fix is delivering rather than promising, since “promising that AI will cure cancer is more a cliche than inspiring.” 52% of Americans now say they’re more concerned than excited about AI in daily life, up from 37% in 2021. Read more

🌶️ OpenAI published the first numbers for its own chip. Jalapeño is the inference ASIC OpenAI co-designed with Broadcom, announced back in June, and on August 25 the company put results behind it: 1.5 to 1.9 times more work per watt and 1.7 to 3.6 times lower latency than a comparable Nvidia GB200 system, on OpenAI’s own benchmarks. It isn’t a product you can buy, it runs inside OpenAI’s infrastructure serving ChatGPT, Codex, and the API, with deployment starting by year end and second and third generations already in development. Read more

✂️ OpenAI is pulling its models out of Cursor because Elon Musk bought it. SpaceX acquired Cursor in July, which triggered contractual cancellation rights, and OpenAI announced it will finish winding down by November 12. The stated reasoning is Musk’s track record on contracts, citing the $2 million annual Twitter deal he killed after buying the platform and xAI training on OpenAI outputs, and OpenAI accounts for only about 5% of Cursor’s model traffic. Musk’s response was “I couldn’t care less.” The neutral multi-model editor was a genuinely useful thing and it’s now a casualty of a grudge, so plan your toolchain accordingly. Read more

⚠️ A fourth plaintiff joined the lawsuit against xAI over Grok-generated abuse imagery. The woman, filing as Jane Doe 4, alleges her stepfather used Grok to turn a photo of her at age 11 into more than 7,000 explicit images, and that he was found dead of suicide two days after law enforcement found them in a raid. She joins three Tennessee teenagers who sued xAI, now part of SpaceX, for failing to build safeguards against generating explicit imagery of real people including minors, and the case seeks class action status. “It is taking everyday life and turning it into child sexual abuse,” she told TechCrunch. SpaceXAI has not commented. Read more

🐂 The mystery model topping OpenRouter turned out to be Chinese, open, and almost free. Developers spent a weekend trying to identify “Ox Alpha,” an anonymous free model pulling double the usage of second-place DeepSeek, before Z.ai claimed it as GLM-5.3-Flash and published the weights on Hugging Face. Artificial Analysis has it at 57 on the Intelligence Index at $0.045 per task, roughly 10x cheaper than similarly ranked rivals, and Z.ai says it runs entirely on Chinese-made chips. Read more

🧬 Claude designed proteins that work in a real lab, against 14 of 15 targets. Anthropic ran an autonomous design campaign with no human scientific guidance during execution, producing 354 confirmed binders out of 1,320 designs, and had Adaptyv Bio and Twist Bioscience build and test them independently. Hit rates landed at 26.7% for Mythos Preview and 22.6% for Opus 4.8 working all targets at once in a 48-hour session, rising to 35.1% when Mythos Preview took targets one at a time, against the 10% to 15% that Anthropic says is normal for the field. Some of the strongest designs bound several times more tightly than the best previously published result. Doubling the hit rate on the part of drug discovery that usually eats years is the kind of result that makes the chatbot framing look silly. Read more

💰 Anthropic booked $11.5 billion in one quarter and stopped setting money on fire. Bloomberg saw preliminary documents putting Q2 revenue at more than fourteen times the $787 million Anthropic did in the same quarter of 2025, and more than double the $4.73 billion it did in Q1. The company also reported positive adjusted operating income, which no frontier lab has managed before, though “adjusted” is carrying a lot of weight in that sentence and Anthropic declined to comment on any of the numbers. All of it points at an October IPO that reporting pegs around a $2 trillion valuation, so read this quarter as the pitch deck it is. Read more

🏚️ Fortune got inside DeepMind’s very bad year. Gemini 3.5 Pro blew three release deadlines, June, mid-July, and into August, and one engineer put the cause at Google chasing search traffic and chat while everyone else ran at coding and agents: “They dropped the ball a little bit.” The talent bleed runs both directions, with Noam Shazeer gone to OpenAI, Nobel laureate John Jumper and two other AlphaFold veterans gone to Anthropic, Jeff Dean out, and safety researcher Alex Turner resigning over the April Pentagon deal that 580-plus employees signed a letter against. Read more

💸 Stripe bought OpenRouter for $7.5 billion. Bloomberg reported the deal done in mid-August, roughly three months after OpenRouter raised at a $1.3 billion valuation, which is about a 5x markup on a company whose product decides which model answers your API call. OpenRouter claims 8 million users and access to more than 400 models, and the Wall Street Journal first surfaced the talks in July. Read more

🥇 Musk called Grok 4.6 “objectively No. 1.” Artificial Analysis has it third. The model scored 61 on that group’s Intelligence Index, behind Opus 5 at 63 and Fable 5 at 62, while beating Sol on coding and edging Fable on professional and legal work. Where the claim holds is the bill, at $2 per million input tokens and $6 per million output, roughly a 60% discount to the frontier, and it landed on Google’s Enterprise Agent Platform about a week after launch. Read more

🔍 Nvidia is in talks to back Perplexity at north of $30 billion. The round would be up more than 50% from the roughly $20 billion valuation a year ago, on annualized revenue that’s climbed past $750 million from under $250 million at the start of the year, much of it credited to Perplexity Computer. Nvidia has been an investor since 2023 and Perplexity committed in July to running its agent workloads on Nvidia’s Vera CPUs. Read more

📉 DeepSeek raised API prices by as much as 1,100%. The increases run from 51% to eleven-fold depending on the model, the token type, and the hour, and they took effect at 16:00 UTC on August 16. DeepSeek also introduced peak and off-peak billing, with off-peak at half of peak and 17 of every 24 hours qualifying, and said it’s doing this to “allocate resources more reasonably,” which is what a company says when demand outruns its capacity. Read more

🔏 Anthropic started signing Claude’s output with a watermark you can’t see. The company published a support page laying out plans to embed invisible marks in text, code, and file outputs, with files carrying C2PA provenance labels already used for AI images and video. The watermark survives copy and paste out of the app, models shipped after August 2 have it built in, and older ones get retrofitted. Read more

🧪 AI Research of the Week

Learning more about Claude’s mathematical capabilities
From Anthropic

Jake’s Take: An Anthropic staff engineer, Jarred Sumner, told an unreleased research version of Claude to take a real stab at the Riemann hypothesis (debatably the most famous unsolved problem in mathematics, posed in 1859), and let the model works the mathematics completely independently.

It didn’t solve it. And really, nobody thought it would. Anthropic said flatly that the techniques Claude used won’t get anyone there (yet).

What it did instead however, was the interesting part of this write up. The Riemann hypothesis claims that every interesting zero of a particular function sits on one specific line, and since nobody can prove that for all of them, mathematicians prove it for a fraction and spend careers pushing the fraction up. Brian Conrey got it to about 40% in 1989. A team of four nudged it to 41.6% around 2018. Here, Claude took it to 67.2% across two Claude Code sessions in roughly a day and a half, after 650 failed ideas, 60 subagents, 2,400 shell commands, 54 papers pulled off arXiv, and 31 million output tokens. The winning move was combining a technique from the 1970s with a Bombieri paper from 2000 that nobody had thought to put together. Claude then wrote the proof in Lean, a language that machine-checks mathematics, so the result stands on its own without anyone taking Claude’s word for it.

67.2% does not mean we’re two-thirds of the way to a proof, and it says nothing about where the other third of the zeros are. But analytic number theorists moved that constant something like 1.7 percentage points across 31 years, and Claude here moved it 25 points over a day and a half with a unique synthesis of solutions.

Alpöge and Furman checked the work inside Anthropic, Conrey and Goldston validated it from outside, and the Lean proof is sitting on GitHub (the link, for those of you who are much smarter than me).

what to know for later

🧒 OpenAI now guesses how old you are and puts teenagers somewhere different. ChatGPT for Teens began rolling out, automatically enrolling anyone who says they’re 13 to 17 or who the age-prediction model thinks is. The model reads account-level and behavioral signals, including topics discussed, time of day, and account age, and the teen tier bundles study tools with tighter content limits, parental controls, and a model spec that rules out romantic language and emotional dependence. Get misclassified and you verify through Persona to get full access back. Read more

👧 Meta is paying $18 billion over teen harm. The settlement resolves claims from dozens of states that Facebook and Instagram were built to keep kids engaged while the risks got downplayed, and teen accounts now get a 2-hour daily limit, an overnight shutdown, muted school-hour notifications, stronger age checks, and break prompts. Meta guaranteed about $12.7 billion of it, with the remaining $5.3 billion contingent on TikTok and YouTube adopting comparable rules and paying around $6 billion each, a condition Meta has been pushing via full-page newspaper ads and an open letter. Read more