Fable 5 Comes Back Nerfed, Sonnet 5 Rewrites the Token Math
Fable 5 Returns — With an Asterisk
After roughly nineteen days behind glass — Fable 5 and its safeguard-free sibling Mythos 5 had been export-control-frozen since mid-June — the ban lifted overnight, and the mood went from euphoria to buyer's remorse before most of the US had woken up.
The unlock. Late on the 30th, Anthropic confirmed the reversal: "We've received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow." Jerry Liu spoke for the timeline — Fable 5 back, LeBron possibly to the Warriors, AIE World's Fair: "it's a good week to be in SF" — and, more succinctly, "We are SO back." RhysSullivan captured the developer version: "fable coming back to my codebase after 19 days of opus."
The asterisk. Then came the full redeployment post (1,624 replies, 3,454 RTs, 20,643 likes, 2.49M views): "Claude Fable 5 will be available again globally tomorrow. After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8." The post went further than a product note — Anthropic said it's drafting a "consensus framework — with Amazon, Microsoft, Google, and other Glasswing partners — for assessing the severity of AI jailbreaks," and scaling up collaboration with the US government on model testing, including pre-release access to models and safeguards and joint research. In other words: the export fight ended with Anthropic pulling the government closer, not pushing it away.
The audience read the classifier line as a bait-and-switch. For a crowd that lives in Claude Code, "routine tasks like coding and debugging will fall back to Opus" is not a footnote — it's the whole product. Tim Sweeney (331 likes): "'Routine tasks like coding and debugging' lol wat?" dogukanbuilds: "What will be the point of Fable 5 then, am I missing something?" newmike_io (236 likes): "so basically tunneling to opus whenever u want and charging for fable? nice." Alex.Mhz supplied the cynical gloss: "basically guys we've nerfed it, but it will be less nerfed the more you all spend the tokens and we can dig through your data." And modomango refused the framing outright: "No one is 'getting back access' to it. This is a heavily downgraded, heavily modified and censored Frankenstein of a model that apparently cannot even be used for coding."
The fine print made it worse. Fix (822 likes, 88.7K views) itemized the sub math: "Fable 5 will only be included for up to 50% of your weekly usage limits 😶 And it's only available on plans for 6 days before being removed on July 7, then it becomes credits-only." MidnightIgnite quoted the blog's escape hatch — "After this point — when sufficient capacity allows us to do so — we aim to restore Fable 5 as a standard part of subscription plans" — but the damage was done. erawrlyne (703 likes) delivered the churn thesis: "So we basically have Fable on our sub for less time than originally planned, for less usage than allowed, and it also can't be used for coding tasks during the time we CAN use it? Why would anyone even stay subbed at this point?" Theo, who'd flagged the usage cap the night before ("Fable 5 will be included for up to 50% of weekly usage limits through July 7… 🪦"), did the arithmetic on what subscribers actually got versus what was promised: "We were originally going to get 100% usage from June 9th to June 23rd (15 days). Instead we got 100% usage for 3 days and 50% usage for 7 😭."
Thariq worked the reply section like an on-call engineer. His clarification thread (311 replies, 1,168 likes, 152K views) — "As with the original classifiers, a small fraction of routine coding and debugging tasks will be flagged and fall back to Opus. We're excited for guys to get access back tomorrow" — turned into a live Q&A. Will the fallback be visible? Yes, transparent. Billed at which rate? "yes you will know and yes billed at opus rate, with prompt cache miss paid for." The instant-classic exchange came from dax: "a lot of my prompts were being downgraded and when i checked logs it said TOO_DUMB_TO_NEED_FABLE what was happening there?" — to which Thariq, caught out (146 likes), replied: "didnt think you would read the logs tbh." And when a user asked whether it'd be live in time to stay up for, the answer (149 likes) was simply: "you should sleep."
The sharpest note on the whole rollout was about messaging, from a fan. LLMJunky: "i cannot believe your lawyers approved that post like that. that is the worst wording you could have possibly used. 'Some coding and debugging tasks may get inadvertently flagged and routed to Opus 4.8.' It's not hard." Earlier he'd panicked in real time — "YOU'RE TROLLING FABLE CANNOT BE USED FOR CODING 😭" — before landing the accurate version: "You can use Fable for coding. It's just much more likely to trip the classifier. Going to be strict, I presume." The story of the day, in one line: the export controls came off, and Anthropic replaced them with a self-imposed filter that happens to bite the exact users cheering loudest.
Sonnet 5 and the Price-Per-Token Reckoning
Overshadowed by the Fable circus but arguably more consequential for anyone paying an API bill: Claude Sonnet 5 shipped, and the interesting story isn't the benchmarks — it's the tokenizer.
Simon Willison did the accounting. In his notes (and a pelican) on Sonnet 5, written up here, the new tokenizer emits roughly 30% more tokens than Sonnet 4.6 for the same text — so even though the per-token price is unchanged ($3/$15 per million, with intro rates of $2/$10 through August 31), effective cost rises sharply and unevenly by language: English 1.42× more expensive, Spanish ~1.33×, Python code ~1.27×, and Simplified Mandarin essentially flat (1.01×). Anthropic positions the model as "close to Opus 4.8, but at lower prices," with a 1M-token context, 128K max output, adaptive thinking on by default — and, notably, no more temperature / top_p / top_k sampling params. The pelican-SVG tradition survives; the pricing predictability does not.
That reframed a debate that had been simmering all week. Steipete distilled it: "Price per token != cost per task." Theo, never one to miss an efficiency angle, pushed the point and posted a video on why OpenAI's models are so efficient — "With Sonnet 5's insane inefficiencies, feels like a good time to post it :)" — and amplified the take that "you can't compare models token to token — needs to be outcome-based pricing" plus the claim that OpenAI's token efficiency is a deliberate, years-long focus. He also couldn't resist a jab at the marketing math — "'A fraction of the price' — I guess 130/100 is technically a fraction?"
And Theo tried to lower the temperature with a video. His pinned Sonnet 5 explainer (45 replies, 335 likes, 22.7K views): "Big day for Claude fans! We now have Sonnet 5 and Fable is back tomorrow 🙏 I saw a lot of misunderstandings about Sonnet 5, so I rushed out a video to try and break it down." The replies proved the misunderstandings were mostly disappointment — the same "coding and debugging fall back to Opus" line bleeding over from the Fable thread. The through-line for the day is clean: a per-token headline price is now actively misleading, and the people who care most are the first to notice.
Speed Is the Next Unlock
If Sonnet 5 made tokens the topic, a parallel thread argued the real bottleneck is latency, and that the next step-change won't be smarter models but faster ones.
LLMJunky made the case from the trenches. Quoting a note that "gpt 5.6 at 750 tok/s doing computer use is going to be a little scary" (96 likes): "this is the next major unlock for agents. Right now computer use requires agents work in a loop — make a small change… check state… send to LLM… next step… painfully slow for any complex task. What currently takes my agents 90 minutes, I can do by hand in less than 8 — and the overwhelming majority of that latency is waiting on the model. There will come a time when we look back on these days like we do for dial-up internet." He's already running fast models in preview: "changes the shape of the products you can build… akin to realtime models, but smarter & more capable." A skeptic warned that "beyond 500 tok/sec you can't even keep up with what the model's doing" — LLMJunky's answer: "i absolutely want that. It is the way of the future. You need to prompt better."
The hardware bet showed up on cue. Karpathy amplified Etched's stealth-exit: "We're coming out of stealth. We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800M raised. Early customer tests show us achieving SOTA throughput, latency, and power efficiency on inference workloads. Our first racks ship…" A transformer-specialized inference chip landing the same week the discourse turns to tokens-per-second is not a coincidence — it's the supply side of LLMJunky's dial-up analogy.
Dispatches from the World's Fair
AIE World's Fair rolled into Day 2 in SF, and the loop crowd took the stage. Steipete keynoted and leaned into the running gag that "we didn't talk enough about workflows loops yet" — the workflows-are-out-loops-are-in framing that's threaded through pi.dev's releases all month — earning a "great talk by steipete on building loops" from the floor. Amid the model drama he also found the reset button.
The color was as good as the content. swyx flagged an unexpected favorite detail: "the daily conference newspaper. Nice to flip through something with a hard stop, no algorithm trying to turn five minutes into an hour" — and teed up poster sessions turning into "poaster sessions" tomorrow ("submit your hottest tweets to @vibhuuuus for printing"). Context engineering got its own track this year. And Thariq, fresh off the Fable all-nighter, was still tweaking his deck: "making some last minute changes — see you tomorrow at AIE!"
Also Worth a Look
Claude Desktop lands on Linux. Boris Cherny: "You asked, we listened. Claude Desktop on Linux is here!" Per ClaudeDevs, it's a beta on Ubuntu and Debian, and it's a first-class desktop experience — Claude Code, Claude Cowork, and chat — "alongside the browser and terminal," on all paid plans. The terminal-native crowd finally gets a GUI without leaving their OS.
Matt Pocock ships a /wizard skill. Pocock: "Getting sick of setting up third-party services, so I built a skill for it. /wizard builds you an interactive CLI for the task you're currently doing, and takes as much work off your hands as possible." The skill is open on GitHub. It pairs with his ongoing effort to make his /review skill smarter by feeding it Martin Fowler's catalog of code smells — "Mysterious Name, Duplicated Code, Feature Envy, Data Clumps, Primitive Obsession, Repeated Switches, Shotgun Surgery, Divergent Change… This stuff is catnip for LLMs."
shot-scraper learns to record video. Simon Willison: "I've added video support to my shot-scraper browser automation tool — you (or your coding agent) can now create a storyboard YAML file and use that to record a video demo of new web application features." Write-up here. A neat primitive for the "agent ships a feature and produces its own demo reel" workflow — his own example was scripted by Codex Desktop and GPT-5.5 xhigh.
Jerry Liu on the shape of document parsing. A useful frame for anyone building RAG-over-docs: fully solving document parsing means covering the whole Pareto curve of accuracy, cost, and latency — high-accuracy 99%+ parsing for regulated industries at one extreme, low-cost/high-throughput at the other — rather than shipping a single one-size-fits-all parser.