Good morning. The Fable 5 saga has a resolution, sort of — Anthropic is back in the Commerce Department’s good graces, but the terms are murky and the goodwill damage looks lasting. Also today: Cloudflare tightens the screws on AI crawlers, Claude helped a researcher find a festival-ticket exploit, and Meta wants to rent out its GPUs.
Fable 5 returns, with strings attached. Commerce lifted its export controls on Fable 5 and Mythos 5 after roughly 2.5 weeks, and Anthropic announced the model’s return — capped at 50% of weekly usage limits and free only through July 7. Wired reports the concession was narrow: an existing guardrail extended to route certain code-fixing requests to Opus 4.8, a behavior cybersecurity researchers described as relatively benign. Defense Secretary Pete Hegseth reportedly still sees no path to lifting his February supply-chain-risk order, so Anthropic’s Pentagon problem persists.
The HN reaction is bleaker than the headline. The top comment on the restoration thread captures the mood: “The damage is done.” Several commenters said they’re migrating to open-source models or Codex CLI, citing regulatory unpredictability and, separately, this week’s revelation of steganographic markers in Claude Code. Others zeroed in on the vague promise that Anthropic will “proactively detect and address security risks” and report malicious activity — with no definition of what that means. One user tested Fable 5 on a book manuscript about human consciousness and got it flagged.
Cloudflare turns the screws on AI crawlers. Starting September 15, Cloudflare will block “mixed-use” crawlers by default from ad-supported pages — meaning bots that combine search indexing, AI agent traffic, and training scraping will need to separate those functions or pay through Cloudflare’s Pay Per Crawl marketplace. The policy hits new customers, new sites, and all free-tier existing customers. CEO Matthew Prince noted bot traffic has already passed human traffic online, ahead of his own projections.
Claude found a hole in nearly every US festival’s ticketing. Security researcher Ian Carroll used Claude Opus 4.7 to uncover a vulnerability in Front Gate Tickets that would have let him mint unlimited free VIP tickets to almost every major US music festival. He reported it responsibly; Front Gate patched within 24 hours. Carroll thinks Claude could plausibly have run the exploit end-to-end on its own — which is either a good argument for the classifiers Anthropic just added, or a bad one, depending on your priors.
Meta wants to rent out its GPUs. Meta is reportedly building “Meta Compute”, a cloud business that would resell AI compute and hosted models — including the new closed-weight Muse Spark — in direct competition with AWS, Azure, and Google Cloud. It mirrors what xAI recently did with SpaceX’s excess capacity. Meta has committed $182.9 billion to AI infrastructure with limited direct product revenue, so the “sell picks and shovels” pivot has some fiscal logic, assuming the depreciation math works out.
Gemini Spark lands on Mac. Google’s agentic assistant is now available in macOS beta for AI Ultra subscribers, with file management, MCP support, and integrations spanning Google Tasks, Keep, Canva, Dropbox, Instacart, and OpenTable. It puts Google in direct competition with Claude Desktop and Copilot on the desktop agent front. Google also added the Keep integration that critics called out as a glaring omission at launch last month.
Venice AI hits unicorn status on the privacy angle. Venice AI raised $65M Series A at a $1B valuation, led by Dragonfly with Coinbase Ventures participating. The platform offers 200+ models with client-side encryption and no data storage, and has hit $70M in annualized revenue with 3 million users in two years. CEO Erik Voorhees, a longtime bitcoin figure, pitches “uncensored” AI access as a neutrality principle — a framing that will land differently depending on which chatbot-safety story you read last.
A small ML research note worth flagging. Researchers funded by Paradigm showed that applying matrix orthogonalization during memory reads in mLSTM models — using Newton-Schulz iterations borrowed from the Muon optimizer — meaningfully improves noisy associative recall. The motivation is practical: long-horizon RL where transformer attention’s quadratic cost isn’t viable. Worth a read if you’re following the recurrent-model revival.
That’s the briefing. If you’re a Claude subscriber, you have until July 7 to figure out whether Fable 5’s new pricing makes sense for you — and whether the trust math still adds up.