Anthropic Apologizes: the Sabotage Walkback, Dario's Policy Essay & the End of Tokenmaxxing

The Safeguard Walkback

Anthropic apologizes — sort of. The big follow-up to Tuesday's backlash: Anthropic told Wired (scoop by Maxwell Zeff) that "We're changing Fable 5's safeguards for frontier LLM development to make them visible… We made the wrong tradeoff and we apologize for not getting the balance right." Simon Willison is "very pleased" (his blog post) — the policy, tucked away in the system card, had Claude identifying "requests targeting frontier LLM development" and silently limiting effectiveness.

But read the fine print. Simon's replies are a wall of skepticism worth scrolling: Kristoph: "They didn't walk it back. It will now refuse to perform the task and tell you why"; akira: "They just won't silently degrade perf"; Leon Zhu: "no longer secretly assassinating your research, but assaulting your research with a machine gun from the front door"; and kache's three-word reply: "Did they though?". Simon's own caveat: "make them visible" means they're undoing the truly egregious part — the silent degradation — not the safeguards themselves. On the trust question, his answer is that we have to trust Anthropic wouldn't risk getting caught lying about it; judging by the replies, that trust is not in great supply.

The pressure that worked. Before the reversal: antirez called the gating "deeply wrong" (via simonw RT), Hugging Face's Clément Delangue appealed in good faith for a course correction, and Dean Ball argued the secret-sabotage policy undermines good safety policy (via steipete RT) — it's plausibly anti-competitive conduct, which hands ammunition to people who want to dismiss all safety work as market protectionism.

Policy & Openness

Dario posts on main. Dario Amodei published "Policy on the AI Exponential" (2.8M views, 9.2K likes) — full essay. The framing: AI policy is Treebeard and AI progress is the Hobbits — government moves at tree-speed while models went from barely writing code to writing most of the code at major AI labs in four years. He points to Mythos Preview's cybersecurity implications as the moment the risks became undeniable, and argues the preserve-optionality era of policy (transparency rules, chip export controls) is no longer enough. As Theo put it: "We got Dario posting on main before GTA 6."

The reception is rough. The ratio-ing replies are the story: Conor Grogan's TLDR (798 likes) — "Declare AI too dangerous for ordinary competition so you propose a regulatory regime where only the largest incumbents can survive… Warn about corporate power while sketching a corporate-state cartel over compute"; tinygrad: "Historically, what happens to cult leaders when the rapture doesn't come?"; several note the timing two days after the IPO filings and days after shipping a deliberately nerfed Fable. The most-viewed reply is simply "Now make competing with us illegal" — the meme.

Armin Ronacher: "Gaslighting Openness." Mitsuhiko's short essay (post, RT'd by steipete) ties the threads together: opinion makers increasingly frame access as irresponsibility. Apple wrapping EU delays in safety language and Anthropic wrapping Mythos/Fable restrictions in security language are the same move — "They trained their models on public works, then block Open Source attempts to learn from and distill these systems." His follow-up quip lands the same point: "if you know enough history to cite the East India Company as a negative example, you know enough history to avoid becoming the East India Company entirely without governmental intervention." Related: antirez asks where Europe's AI scene went (via mitsuhiko RT) — beyond a few unicorns, the continent needs to "recover our industrial ethics."

The Token-Burn Economy

Tokenmaxxing hits a wall. Jerry Liu's thread is the clearest writeup of the new economics: his whole MTS team ran on a tokenmaxx-everything philosophy on Claude Max plans — until Fable. One engineer hit his limit 3 times in a day, burning the equivalent of $1.5k in 10 hours; half the team has hit quota on eng work. "If every engineer starts spending tokens at levels equivalent to headcount costs, our burn rate will meaningfully increase… we will have to start thinking about model routing in our core engineering usage." Good replies: the napkin math that a $350K engineer is ~$175/hr, so the spend can still pencil out (Jerry's pushback: tokenmaxxing means most tokens are wasted by design), use Fable as planner/verifier orchestrating Sonnet and Opus workers, and a lot of the waste is bad loops — rereading context, vague self-checks, retries without new evidence. Han Xiao's take: tokenmaxx while you can — "we need small, local and vertical models to be ready."

Theo's stress test. He maxed out two accounts in a day, and shared useful limit math: 93% of a 5-hour limit ≈ 25% of the weekly limit, so you get just under four "5-hour max-outs" per week. His verdict on the lineup: "Mythos is the new Sonnet, Opus is the new Haiku", and "holy shit this model loves to light tokens on fire". Nostalgia is already setting in: "Remember the days where we used GPT-5 and Sonnet 4 on $20/month plans and it felt nearly unlimited?"

The structural take. Jerry agrees with the divergence framing: cost management wasn't a priority before ~Opus 4.5, but frontier labs are now optimizing for something different than production ROI. Shensi (via Jerry's RT): "the cost of fable is going to make smart model routing impossible to ignore." LLMJunky polled whether anyone will pay API prices for Mythos after June 22, and a reply in Jerry's thread maps the subsidy sunset: Codex's crazy plans ended last week, claude -p moves to API billing, Fable leaves subscriptions on the 22nd — "It's been fun but was bound to end eventually."

Agentic Coding & Agent Harnesses

How Fable edited its own launch video. Thariq's video (444K views) on the workflow behind the Fable 5 launch video he cut without ever opening a video editor: Claude Code wrote code and tool calls for transcription services, ffmpeg cutting, colorgrading, the Figma MCP, and Remotion UI rendering. The deck is public (generated by Claude from his session transcripts in about an hour), and the original launch video is here. Don't skip the colorist's reply: the "color graded" footage applied the wrong gamma curve to S-Log3 — "if you don't understand every aspect of what the model is doing, you can't reliably notice when it's doing something wrong." Thariq took it well: being more in the loop would have caught it.

HTML plans go mainstream. Theo is converted: "HTML plans are so good. Thank you @trq212 for putting me on. Mythos is so good at them" — he got a full monetization breakdown as a styled HTML doc from a one-line request. (Karpathy was on this in May: "structure your response as HTML" and open it in a browser.)

/teach grows up. Matt Pocock may move /teach to its own repo — "It's not really about engineering" — with 100 example prompts and docs on the underlying principles. It now recommends primary-source reading in every lesson: "The goal isn't to keep you shackled to the agent, it's to get you confident enough to read the sources for yourself." His junior-to-senior pitch is npx skills add mattpocock/skills --skill teach. Best reply: a /baton skill — a self-handoff note (next move, branch, files, failed attempts) that survives compaction.

The counterpoint to loop-design. Via Jerry Liu's RT, Nick Kango on the "design loops that prompt your agents" meme: "The model subsidies will eventually end and this workflow… will result in massive amounts of code that's not well understood that you will have to pay loads to maintain." A useful tension with steipete's loops and Pocock's queues from earlier this week.

Claude Code field notes from Theo. Warning for account switchers: multi-account on the desktop app is "NOT VIABLE" — hard logout loses all sessions, while the CLI handles /login fine. And he's hitting safety flags mid-task for the first time: a big pass on a new auth feature "randomly degraded like 3/4ths of the way through."

Models, Benchmarks & the Race

Theo's Fable verdict. The breakdown video is out: "Mythos is here. It's called Fable 5. I spent the last 24 hours spending thousands of dollars in tokens to give you a realistic breakdown." Notable from the replies: ~$100 of compute for an 8-minute task is the real price tag; a generational leap over Opus but maybe only ~10% over GPT-5.5; plan-following is "next level" compared to Opus. He also wants more niche benchmarks: "We need ios-bench. We need ts-bench. We need baseball-bench."

GPT-5.6 watch. LLMJunky strongly suspects GPT-5.6 drops today — "I've heard from a couple sources that it is an extremely strong model" — and is polling whether it matches, beats, or falls short of Fable 5. The same question is all over Theo's and Jerry's replies: if Fable goes API-billing and GPT-5.6 doesn't, does dev market share shift?

Model Labs vs Agent Labs. Latent Space's new AINews essay with Sarah Guo (post, via swyx RT) — swyx calls it "probably the most important framework" for where AI value lives right now: model labs compete on capability, agent labs on harnesses, context, and distribution. A good TLDR thread here. Related from the same feed: the Satya Nadella crossover interview (No Priors × Latent Space at MS Build 2026).

Poetic launches. Markie Wagner introduced PoeticHQ (770K views): $50M at $500M from Kleiner Perkins and Founders Fund, claiming multi-hour tasks at 99%+ accuracy with 10× fewer tokens than agents. The architecture is the interesting bit: "When the world stays the same, Poetic runs fixed code: fast, cheap, identical every time. When the world changes, Poetic uses AI to regenerate its approach" — agents as code-generators rather than executors, aimed at AML, fraud investigation, and underwriting (AIG, SoFi, Chime as customers; eight-figure run rate with a team of four).

Other Bits