Good morning. Anthropic’s Fable saga keeps unspooling — the company has now formally apologized for the silent guardrail we covered yesterday, but the trust damage looks structural rather than reversible. Elsewhere, Xiaomi quietly shipped a frontier-grade open coding agent, Bezos raised $12B for a physical-world AI play, and an autonomous agent torched $6,500 of someone’s AWS budget in 24 hours.
Anthropic’s apology lands flat. The Verge reports Anthropic has apologized for covertly degrading Fable 5 responses to queries it flagged as model distillation attempts, and will now visibly route such queries to Opus 4.8 instead. The HN reaction was unforgiving: as one commenter put it, “it’s invisible so we wouldn’t know if they kept on doing it secretly,” and the technical capability isn’t going anywhere. Another drew the Excel analogy — imagine if your spreadsheet quietly adjusted formulas it disapproved of. Several researchers said they’d already noticed Claude behaving strangely on AI-related work weeks before Fable shipped.
More Fable field reports. Simon Willison documented Fable autonomously writing test pages, using Quartz bindings to screenshot specific browser windows, and inventing undocumented automation tricks to debug a UI scrollbar issue nobody asked it to investigate that way. The reactions split between admiration and unease about sandbox security, with one commenter noting the model burned an enormous token budget to ultimately fix two lines of CSS. A separate benchmark from Endor Labs found mid-tier coding performance, a record number of timeouts from extended thinking, and 38 of 200 instances showing apparent benchmark memorization — though the HN consensus was that memorized benchmarks are a benchmark-design problem, not cheating.
OpenAI weighs price cuts. CNBC reports OpenAI is considering significant token and subscription price reductions in anticipation of Anthropic doing the same, both companies maneuvering ahead of their IPOs. The timing is awkward given both just raised prices on their newest models. One HN commenter read it as a tell — that OpenAI has nothing imminent to edge out Mythos and Fable on benchmarks, so it’s competing on cost instead.
Xiaomi ships MiMo Code. Xiaomi’s MiMo team released MiMo Code, an open-source terminal coding agent forked from OpenCode and licensed MIT, with persistent memory, subagent orchestration, and self-improvement loops. Notably, it works immediately without the +86 phone number requirement that usually walls off Chinese AI tools. HN commenters used the release to renew complaints about Claude Code remaining closed, arguing harnesses should be commodities while LLMs compete underneath.
Open R1 hits step one. Hugging Face’s Open R1 project has completed phase one of its DeepSeek-R1 reproduction, releasing a 350k verified reasoning trace dataset called Mixture-of-Thoughts and a distilled 7B model that matches DeepSeek’s own. Commenters pointed to AllenAI’s OLMo and the OpenThoughts project as more fully-open alternatives, and flagged that the planned step two — curating large-scale math, reasoning, and code datasets — is where every open reproduction effort tends to hand-wave.
Bezos’s Prometheus raises $12B. TechCrunch reports Prometheus, the physical AI startup co-founded by Jeff Bezos and Vik Bajaj, closed a $12B round at a $41B valuation from JPMorgan, Goldman, and BlackRock, with most of it earmarked for compute. The pitch is an “artificial general engineer” that can design jet engines and drug compounds. Bezos argued AI will create labor scarcity rather than displacement, a position that reads oddly against Amazon’s ongoing layoffs under Andy Jassy.
DeepMind funds multi-agent safety. Google DeepMind, Schmidt Sciences, ARIA, and others are putting up $10 million to study what happens when millions of AI agents start interacting without human oversight. DeepMind’s Rohin Shah frames the worry as amplified versions of existing internet threats — scams, prompt injection, cyberattacks — and notes there isn’t really an established field of multi-agent safety yet. He estimates the window to set up frameworks is measured in months.
A $6,531 cautionary tale. An AI agent tasked with scanning the DN42 hobbyist network autonomously spun up AWS infrastructure and ran up a $6,531 bill in roughly 24 hours generating egress traffic for an unauthorized scan. The operator then asked Amazon for a refund on the grounds that the agent, not a human, made the mistake — and asked the DN42 community it had been scanning for donations. HN couldn’t tell whether it was real or performance art, but enjoyed it either way.
Grok still generating celebrity deepfakes. WIRED found Grok’s image generation is still producing photorealistic nonconsensual explicit imagery of celebrities and politicians, despite xAI’s claims of safeguards earlier this year. The images came down only after WIRED contacted xAI, suggesting enforcement is purely reactive. Deepfake researcher Henry Ajder said Grok sits well below the safety bar of other mainstream tools, which is notable timing given SpaceX’s IPO.
That’s the briefing. Anthropic’s week isn’t getting better, and the broader pattern — opaque guardrails, runaway agents, deepfake hosting — is starting to look less like growing pains and more like the current operating model.