The Internet Is Furious at Anthropic After Claude Fable 5 Release

Anthropic dropped its most powerful public model on Tuesday, and by Wednesday, a significant chunk of the AI community wished it hadn’t.

The consensus around Claude Fable 5—the first publicly accessible version of the company’s restricted Mythos-class technology—appears to be that it’s pretty good at coding, and produces amazing results in everyday sessions. But it launched with some heavy complaints attached: It burns tokens at a ruinous rate; it secretly self-sabbotages for certain research tasks; and it forces every user into a 30-day data retention policy with no exceptions.

The backlash was immediate and loud, cutting across researchers, developers, founders, and open-source advocates. Not a normal launch-day grumble. Something closer to a reckoning.

The token furnace

The first thing users noticed had nothing to do with safety. Fable 5 costs $10 per million input tokens and $50 per million output tokens—double what Claude Opus 4.8 runs.

That pricing is aggressive enough on its own, but the real pain for users is how the model behaves inside subscription plans. Fable 5 counts double against usage limits compared to Opus, meaning the same work on Fable drains your plan allowance twice as fast before you’ve paid a cent in API fees.

In practice, things got worse. In our own quick test, Fable consumed our daily quota in a single prompt. Things don’t get any better if you are one of those clients with deep pockets. Bleeping Computer also tested Fable and found it drained a $100 Max subscription’s daily allowance in just under nine minutes.

Scrimba CEO Per Borgen did the math in public: “Just tried Fable. It burned 1.3M tokens in 7 minutes. That’s $160 per hour. Equivalent to a $333k/year salary,” he posted on X.

Theo from T3 Chat posted that he’d spent over $1,000 in tokens in one day on his $200 subscription plan. Josh Ellithorpe, CTO at Pixelated Ink, said Fable 5 “burns tokens like no other model,” giving him only a few prompts before draining out his quota. “Can’t even review this, since my testing is so limited,” he ranted.

Anthropic’s answer is that Workflow mode—the feature that burns most aggressively—breaks complex prompts into parallel subagent tasks, which costs more compute by design.

There’s also a new system prompt, which is around 120,000 tokens long, and is loaded into every new conversation. For context, this is around the same token context window that GPT-4o could handle before collapsing.

The company also says Fable 5’s per-task efficiency is better than it looks per-token, since it produces more thorough output with less iteration. That may be true in controlled benchmarks. On live subscriptions with hard daily limits, users experienced it as a machine eating their budget in minutes.

The model that lies without lying

The second complaint was more damaging because it came straight from Anthropic’s own documentation. Buried in Fable 5’s system card, the company disclosed that when the model detects a user is working on frontier large-language-model development—pretraining pipelines, distributed training infrastructure, machine-learning accelerator design—it does not refuse to reply and does not fall back to a smaller model. It silently nerfs itself through prompt modification, steering vectors, or parameter-efficient fine-tuning, without telling the user anything has changed.

In other words, researchers don’t know if they are paying for Fable to reply and are getting Opus responses. It also makes it hard for users to know what made their prompt fail.

“Unlike our interventions for cybersecurity, biology and chemistry, and distillation attempts, these safeguards will not be visible to the user,” Anthropic wrote in Fable’s System Card. “Instead, the safeguards will limit effectiveness through methods such as prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT).”

That distinction matters enormously to researchers. As AI newsletter Latent Space noted, a model that refuses openly lets researchers understand a boundary. A model that falls back to a weaker version is detectable. But a model that appears to help while covertly delivering worse output destroys scientific reproducibility—a failed result could come from the researcher’s idea, their implementation, or an invisible intervention that was never disclosed.

Anthropic estimated this would affect approximately 0.03% of traffic. The open-source and research communities found that number irrelevant to the principle involved.

“Dear Anthropic, you broke our trust and I don’t think you’ll ever get it back. My tokens will no longer fly your way,” Arthur Zucker, a core contributor at Hugging Face, posted on X.

Mikel Artetxe, cofounder of Reka AI, also bashed this decision:

“Brilliant idea! Next up: Apple randomly reboots your Mac if you’re building competing tech, Gmail silently edits your email if you mention rival platforms, and Tesla Autopilot swerves if it detects you’re working on self-driving cars. All in the name of safety, of course,” he posted.

The researchers who got hit hardest weren’t the big labs with proprietary infrastructure but the academics, startups, and independent builders using Claude as a public tool—exactly the people Anthropic’s safety rhetoric has always claimed to protect. AlphaXiv, an open research platform, called the practice a precedent that “is not safety,” arguing that safety policies should be transparent and auditable.

Nathan Lambert who just started a role at Arcee AI after working with the Allen Institute put it more simply: “To me this paints Anthropic clearly as anti science, and therefore anti progress and anti safety,” he wrote.

Pseudonymous user “CalleBTC,” an AI and crypto developer who had been waiting on Fable to help train a world model, summed up the frustration too. “Anthropic has lost the plot. I was literally waiting for Mythos to help me train a world model. Instead, they chose to cuck their model to stifle their competition,” he said, calling the move “deeply unethical and disrespectful to developers and scientists.”

Overall, researchers argued that Fable’s restrictions extend beyond specific topics and may be influenced by how the model classifies users.

“Your prompt is mine”

The third grievance affected enterprise users most directly, but the implications reached everyone. Per Anthropic’s own announcement, all traffic on Mythos-class models—Fable 5, Mythos 5, and any future models at similar capability levels—is subject to mandatory 30-day data retention across every platform where these models are offered, including third-party surfaces like AWS Bedrock and Google Vertex AI.

The company assures this data will be deleted after 30 days in “almost all cases.”

The problem for enterprise users isn’t what Anthropic says it will do. It’s what the policy structurally requires. Companies handling privileged legal communications, healthcare records, confidential source code, may be in trouble in case they use these models. If there are specific privacy agreements with Anthropic, users argue they should be updated to guarantee privacy.

The compliance problem goes geographic too. European companies operating under GDPR’s data minimization rules, or any organization requiring demonstrable zero-retention for regulated workflows, are simply locked out of Fable 5 until Anthropic offers a carve-out. Pseudonym X user Lisan al Gaib, a well-known personality in the AI community, flagged the consequence directly:

“Anthropic just delegated a lot of European companies to the permanent underclass. If Anthropic saves data for Claude Mythos and Fable 5 for 30 days, then all companies that require zero data retention simply can’t use them.”

Hugging Face CEO Clement Delangue framed the week’s events inside a larger argument:

“Concentration of power, capabilities and economic wealth is the biggest risk in AI,” he wrote. “We need open science and open-source more than ever!”

Another user wrote: “All jokes aside, it’s very clear that Anthropic is the direct path to the worst type of dystopia. Their CEO is against the very technology he creates. Restricting knowledge and education on ML related topics is beyond despicable.”

Fable 5 is free on Pro, Max, Team, and Enterprise plans until June 22. After that, it moves to usage credits only—API rates, no subscription inclusion—with Anthropic saying it will restore broader access “as soon as capacity expands.”