AI Companies Still Haven’t Delivered on Their Biggest Promises

Dario Amodei publicly disputed a claim made by investor Gavin Baker that sources inside Anthropic told him Amodei believed Anthropic might one day be one of only a few entities left alongside governments. Anthropic spokesperson Schulteau Douglas called the claim completely false and said whoever said it is lying, adding that economic concentration of power is one of Anthropic's primary concerns. Baker said he would discourage Amodei from ever repeating the comment, and David Sacks compared the alleged remark to SBF territory for its hubris.

Amodei argued that the equation of regulation with regulatory capture with concentration of power is overly simplified, and that fair institutional processes have decentralizing power that the capture frame underrates. He cited California SB 53 and SB 1047 as examples of policies that exempt companies below certain revenue or model training cost thresholds, and pointed to testing frameworks advocated at CAISI and the White House that impose more rigorous requirements on frontier models than on smaller competitors. Baker countered that Amodei's pro-regulatory messaging has aided efforts to ban data centers in America, that anti-data center advocacy groups may use clips of Amodei warning about AI dangers, and that the main risk to Anthropic a few months ago was nationalization driven by Amodei's own rhetoric. Amodei disputed that his preferred regulatory path has failed, citing the Trump administration's reported pre-deployment testing approach for frontier models as aligned with his views, though he said he must see the details before fully endorsing it.

Amodei said the most accurate criticism of AI companies including Anthropic is that they have not yet delivered on their big promises, and that saying AI will cure cancer has become a cliche that most people find deceptive rather than inspiring. He attributed public distrust of AI to decades of eroding confidence in companies, governments, and the tech industry rather than to AI leaders warning about risks. He said Anthropic is ramping up efforts in biology and medicine and hopes to have early results in coming months and significant results in coming years. Commentators were divided on his response. Jessica Lessin said two posts changed the narrative about him and gave the tech world a more accessible message. Lulu Cheng Meservey observed that Amodei rebutted a qualitative accusation with a quantitative word-count argument, causing him and critics to talk past each other. The host argued that public perception is not shaped by essay word counts and that Amodei cannot simultaneously claim to understand that social media clips the most negative material and then give interviews filled with easily soundbitable negative statistics.

Angel Broden argues that the assumption any AI lab can build public trust through a spectacular breakthrough such as curing cancer is an oversimplification. She points to the pharmaceutical industry as evidence that delivering major health improvements does not translate into public trust. Broden contends that people ethically opposed to AI are not skeptical of its potential but are questioning whether the companies building it have their best interests at heart. She argues people will judge AI companies on pricing, access, lobbying, opacity, distribution of economic gains, and who holds power. She finds Amodei's diagnosis that institutional distrust is the root cause insufficient, arguing that asking people to place more trust in a small number of powerful institutions to steward AI does not resolve that problem.

Anthropic's second risk report disclosed three significant unreleased models as of mid-July: Opus 5, a model called Model 1 with capabilities broadly in line with Mythos 5, and a model called Model 2 described as somewhat more capable than Mythos 5 but not a capability jump comparable to the leap from Claude Opus 4-6 to Mythos Preview. Anthropic stated it has no plans to release Model 2 publicly and has not run it through their usual evaluation suite. The gap between publicly available models and what labs hold internally is described as wider than it has been in the past.

Anthropic has begun meeting with investors and investment banks ahead of a potential IPO. Revenue reached 11.5 billion in Q2, up 14 times compared to a year ago, annualizing to approximately 46 billion. Investors expect a 2 trillion dollar valuation, exceeding SpaceX's 1.7 trillion valuation and more than doubling Anthropic's May fundraising valuation. Investors expect revenue of 100 to 120 billion by end of this year, while Anthropic's own forecast targets 190 to 200 billion by 2028. Fortune noted that public stocks trade on earnings multiples, and a 2 trillion dollar Anthropic would need annual profits of 59 to 79 billion to justify that valuation.

ZAI released GLM 5.3, built on the same base model as GLM 5.2 with performance advances attributed entirely to scaling reinforcement learning. It scores 28.3 percent on Terminal Bench 3.0, roughly 5 points behind Fable 5 and GPT 5.6 Sole, and achieved state-of-the-art results on Automation Bench and GDPVAL for agentic tasks, slightly edging out US frontier models. It costs less than one tenth the per-token price of Fable or 5.6 Sole. Nathan Lambert argued that strong performance from Chinese labs should no longer be surprising and that dismissing results as distillation or benchmark gaming leads to underestimating their capabilities. Wall Street analysts are updating their view that Chinese labs are not simply selling frontier intelligence for pennies on the dollar, as the pricing gap with cheaper US models has contracted substantially.