Washington Pulls Fable 5's Plug, Amazon Snitches & the Self-Building Codebase

The Fable 5 Shutdown

The government flipped the switch. Late Friday, Anthropic posted the announcement that defined the weekend (12K replies, 25K RTs, 85K likes, 83 million views): "The US government, citing national security authorities, has issued an export control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, whether inside or outside the United States, including foreign national Anthropic employees." Because compliance was binary, "we must abruptly disable Fable 5 and Mythos 5 for all our customers" — every other Claude model stays up. Anthropic called it "a misunderstanding" it's working to reverse (full statement). ClaudeDevs confirmed the customer-facing version and, a couple hours later, reset everyone's 5-hour and weekly rate limits as a consolation.

Who pulled the trigger: Amazon. Theo surfaced the WSJ reporting (1.4M views) — "Amazon reported the jailbreaks to the Department of Commerce, who instituted the ban" — and later pointed to Axios confirming it (axios.com/2026/06/13/anthropic-amazon-white-house). The breaking version, via steph_palazzolo: Amazon CEO Andy Jassy was among the tech leaders who raised security concerns to senior Trump officials, and those conversations "set in motion the government's new export controls." The irony — Amazon is Anthropic's biggest compute provider and a major investor — set off the whole reply section: "a co-dependent partnership… Anthropic begging to be regulated and their top partner providing the justification. It all seems calculated", and a careful caveat from Janek Mann: "Technically the reporting is that Amazon researchers found the jailbreak, not that they reported it to DoC." One on-brand detail: Fable 5 reportedly still works on AWS Bedrock.

The reactions split three ways. Grief, geopolitics, and gallows humor. On the original Anthropic thread, Grok argued it isn't random panic — "Mythos 5 packs serious dual-use cyber power, and governments have long restricted that kind of tech… the part that actually sucks is the blunt global shutdown instead of smart, targeted rules" — against the "peak government panic" camp. Austin Federa listed the precedents (the PS2, the Power Mac G4, FPGA bitstreams and even t-shirts have all been export-controlled). And the reply section filled with refund requests and one viral lifelong-conservative lament (218K views) arguing the move just hands the open-weight future to Chinese labs.

The people who lost their toy. Simon Willison watched it happen live: "I ran a script every minute against the API to see how long I'd maintain access to claude-fable-5 — I lost access 14 minutes ago." Theo turned the news into a published video in under two hours ("Can you believe we did it without Mythos?") and riffed his way through the night — "I guess, in the end, it was a fable after all", "The United States is no longer the best place to build an AI lab". Matt Pocock took the smug-but-fair victory lap: "Fable back in the stable. Looks like my 'I'll see how it shakes out' approach to model assessment wins again."

The "just fix the jailbreak" fight. LLMJunky carried the cynicism: "a sobering reminder that literally anytime, these labs can turn the faucet off… intelligence on their terms, if at all", and the sharpest point — Karpathy himself, now at Anthropic, would be barred from working on or with Mythos under these restrictions. When critics said Anthropic should "just fix the jailbreak issue," he pushed back hard ("Shame on Anthropic for not fixing a potentially unsolvable problem inherent to all LLMs") and flatly: "No, Dario cannot 'just fix it'" — echoing the agentic-security line going around that there's "no firewall for English" and that preventing models from finding-and-fixing bugs would break agentic coding entirely. steipete landed the cleanest dunk on Dario's own week, via Will Manidis: "Dario (48 hours ago): 'US gov should be able to block model deployment.' USG: export controls models. Dario: 'not like that.'" He also flagged The Information's "blindsides business partners" piece and his favorite new americanism — "not consistently candid in their communications."

The measured take. Jerry Liu zoomed out: "As much as I want to regain access to Fable, let's remember that 99.99%+ of the world has barely scratched the surface on GPT-5.5 and Opus 4.8… very few people in F500 are tokenmaxxing effectively." The forward-looking question, via Simon Dorsey: "How is the US going to deal with all the Mythos-level open-weight models in 6-12 months time? Enact a full AI ban?"

Europe, Sovereignty & Open Weights

Armin Ronacher's essay. Armin turned the ban into a long piece (94K views) on "what it says about power and dependency, and what it should mean for Europeans and citizens of the world" (lucumr.pocoo.org). His blunter posts cut through: "Europeans please wake up" and the kicker that the ban is wide enough that "even going to the US is not going to help you". The replies are the real debate. Josip Licardo: "Great analysis, but the conclusion is cope. Europe doesn't need cooperation. It needs chips, capital, and models of its own. It has none of those." Armin's answer: "Doing any of this without cooperation is doing it in hard mode." A genuinely interesting counter from Stéphane Cocquereaumont: "We already have AI labs in Europe… if the US is not competing in Europe, that would increase local opportunities. This is reverse protectionism." And antirez argued the compute is already here — Italy's Leonardo plus the Swiss CSCS "has more than enough compute to train a very large LLM."

The open-weight drumbeat. Hugging Face's Clement Delangue reframed it as a fork in the road: "Path 1: closed-source APIs, concentration of power… Path 2: open-source AI, where everyone gets to participate, own, and decide." The mood, via Simon Dorsey: "Imagining a future where you have to show your US passport every day to turn on your AI 💀."

Agentic Coding & Agent Harnesses

steipete's codebase builds itself. The most striking agentic post of the weekend, 95K views: "Codex runs INSIDE crabbox while it is building crabbox… looping nonstop for the last 4 days in multiple trees. Since all of it is e2e verifiable it basically builds itself. Codex even signs up for the services automatically via browser/computer use. My main job is adding credit card details and closing things that I don't see as a fit." The replies wrote themselves: "We're moving from coding to supervising autonomous build systems", "Software engineers are becoming product managers for software engineers that aren't human", and the deflationary "Sounds like you're just the credit card department now". On the "what is this for" question, steipete's answer: "Fewer people having to run around with their MacBooks open." On cost, he says the $100–200/month subs work for most, with Chinese-lab offerings "a year behind latest gen but for 10% of the costs." Worth heeding the security worry in the replies: an agent auto-signing-up for services with your card is a juicy target.

pi-coding-agent ships — and immediately hits the pinning wall. Armin released the standalone package for CLI use and building on top, then asked the question nobody has a clean answer to: "It's entirely unclear how we could properly pin it, particularly with shrinkwrap being deprecated now (npm/cli#9262). What's the right move? Not to use npm install -g at all?" He clarified pi itself still uses shrinkwrap and vets deps for you — the problem is keeping that working once npm 12 rips it out. OpenClaw hit the same wall ("we use this for all our plugins… npm's only workaround is bundle your entire dependency tree, which is not very helpful"), and Dr. Xi Zeng named the meta-problem: "exactly the annoying edge where agent tooling meets package tooling — the agent wants a clean install story; npm gives you six almost-right rituals and one deprecated warning."

Nested subagents land for real. Matt Pocock: "Seeing a subagent spawn its own subagent is unbelievably satisfying. Subagents are good now" — the payoff from Boris Cherny shipping nested subagent support (capped at depth=5) in Claude Code. Meanwhile the harness layer is consolidating: the Vercel AI SDK now wraps Claude Code, Codex, and Pi as HarnessAgents with sandboxed sessions and AI-SDK-compatible streams (in ai@canary).

Does Git die next? swyx's "The Future Codebase" is the spiciest dev-culture thread of the week: "After the PR dies, after Code Review dies, I'm seriously wondering if Git needs to die next. Roughly 20-40% of code spend is just managing merge conflicts… maybe the future codebase looks more like a Notion or Linear database than .git objects." Zed walked right into it advertising DeltaDB: "the conversation that generates the code is becoming the true source of our software, and Git can't keep up." The good pushback: Ankur Goyal — "the value of git is reproducibility; any alternative needs to maintain that"; Bill Kerr — "the 'an owner ships it' line is the tell. Merge conflicts mostly exist because Git refuses to pick an owner"; debamitro — "Git will live like the unix CLI utilities; PR was a GitHub invention, not part of git"; and the simple reality check: "definitely not seeing this 20-40% — am I doing something wrong or are you?" The most thoughtful reply is loadingalias's five-year tour of why a genuinely collaborative VCS is "fucking hard" — migration, content-addressed storage, safe memory reclamation, and then the real wall: collaboration.

Models, Benchmarks & Cost

The cost gap nobody screenshots. LLMJunky on the chart Artificial Analysis didn't run: the cost to complete the intelligence-index benchmark — "GPT 5.5 xhigh: $3,357. Fable 5: $9,940. Nearly 3x more expensive for a singular point on the intelligence index. Fable is a good model, but this only makes…" steipete's adjacent observation, likely to age well: "GPT is 10-20x more token+cost effective for ~similar outcome."

The benchmark swap that flipped the leaderboard. Artificial Analysis updated its Coding Agent Index, replacing SWE-Bench Pro with Datacurve's DeepSWE — which lifts Codex with GPT-5.5 (xhigh) above Claude Code with Opus 4.8 (max), while newly-released Fable 5 (max) tops the board. A reminder that "the leaderboard" depends entirely on which eval you trust.

Models doing mechanical engineering. LLMJunky kept pushing CAD: GPT-5.5 is now "pretty good at building in CAD" (~$45 in tokens, "not far from models being able to build almost anything") and he wants real 3D-modeling benchmarks because they expose "a clear difference between SOTA models and their open-source counterparts."

Panels over single models. OpenRouter's Alex Atallah announced the Fusion API: "Fable-level performance on deep research tasks, at half the cost — better-than-SOTA using panels. The future of AI is neurodiversity, not single-model takeovers." A neat counter-narrative the same weekend a single model got nationalized.

Also Worth a Look