Anthropic IPO at Risk, Meta's Muse Pop, Token Prices Fall, Open Source Gains Share, Alignment Fails

Anthropic’s IPO prospects appear increasingly uncertain as investors weigh Dario Amodei’s public warnings about existential risk against the company’s growth ambitions. Amodei has said the chance of human extinction exceeds 10% and that AI safety remains unsolved; one speaker argued those statements undermine the IPO, while David Sacks said Amodei should remain CEO because of the culture he built and Anthropic’s ability to compete. Anthropic was reportedly targeting a $2 trillion valuation, compared with OpenAI’s reported $1.2 trillion target, but both offerings could be delayed. Anthropic had been expected to file and go public in October; reports variously pointed to November or May, with further delay possible. Polymarket’s probability of an Anthropic IPO in 2026 fell from a 96% peak to 76%.

The valuation debate centers on whether Anthropic can sustain premium pricing as open-weight models improve. One rough example suggested that even a $100 billion run rate might not justify a $2 trillion IPO valuation if investors discounted the business to $1 trillion or less for competition and customer-concentration risks. Speakers said the S-1 may need to address employee departures over existential concerns, the company’s calls for regulation, and its biology work. David Friedberg identified customer concentration as the central business risk, while also saying Anthropic’s life-sciences models outperform alternatives and that its recruiting appeal remains unusually strong.

A surge of open-weight releases is challenging the assumption that frontier labs can preserve a wide lead. Speakers cited Alibaba’s Coin 2.1 as outperforming Google’s Nano Banana 2 while running locally for free, Xiaomi’s 309-billion-parameter MIMO Pro as comparable to Opus 5 on a user’s server, and Bonsai 2—a 27-billion-parameter Coin fork—as reaching 98% of the larger model’s performance in 5.9 GB. One speaker said token use had shifted in 12 weeks from 80% closed and 20% open to 80% open and 20% closed. Anthropic and OpenAI cut token prices by 50% that week. Some self-hosted models cost under $0.10 per million tokens before compute and related expenses, and smaller models may handle 80%–90% of corporate tasks.

The commercial outlook depends on how much work truly requires frontier intelligence. Speakers said a small number of customers account for much of the most expensive token use, making labs vulnerable if major buyers switch to cheaper models or self-hosted alternatives. Friedberg argued that premium models may retain demanding engineering, math, and biology tasks while open-weight models take routine coding and enterprise work. He said frontier revenue should be discounted if more than 60%–70% of token use is interchangeable with open source, but could remain strong if only 10% is. Sacks countered that 10%–30% of the market may pay a large premium for frontier intelligence and said about 60% of worldwide compute capacity being added over the next year or so is for Anthropic and OpenAI. He put their lead over commodity models at only six to 12 months, contingent on continued gains.

Meta’s Muse made a rapid consumer splash, reaching number one in the App Store with three million downloads; one speaker said Meta stock rose 10% after the release. Muse is free, connects to users’ information, and was described as useful for practical work such as inbox triage and arranging travel. Speakers said it completed inbox triage “relatively flawlessly,” while another described it as less free-wheeling than Grok Bot but effective at tasks. Meta reportedly delayed the rollout by a couple of months to address safety and avoid enabling harmful capabilities, and said Muse’s design drew heavily on OpenClaw. The broader claim was that users often care more about the product and its tools than which model powers it; simple agent interfaces could make AI’s practical value visible to tens of millions of people.

The discussion exposed competing definitions of alignment and safety. Sacks said alignment should mean predictable, reliable, safe products that respond to users, not models optimized for abstract notions of humanity’s preferences or the median voter’s. He said Anthropic’s constitution generally tells Claude to trust Anthropic over operators and users, while permitting it to challenge Anthropic and refuse requests it considers unethical; another speaker warned that training models to exercise moral judgment and resist their creators could make them less predictable. Speakers also discussed secondhand reports that AI firms sought a Section 230-like liability shield, while emphasizing that a separate equity-for-relief rumor was unconfirmed and not established policy. They said administration officials rejected liability waivers and cited potential legal enforcement. Anthropic’s biology research also drew scrutiny: its preprint described agents identifying previously uncharacterized proteins and enzymes, with lab testing of predictions in an automated biosafety level 1 and 2 facility, not a site for making pathogens or conducting gain-of-function research.