Cold open: You’re hunched over your laptop at 10 a.m., watching a usage meter climb so fast it looks like it’s on fast-forward. The chatbot politely refuses a child’s biology question and hands you an earlier model instead. I remember the exact Reddit thread where a developer posted: “Completely unusable right now.”
I’ve followed Anthropic since Claude first stirred the conversation, and I tell you this as someone who tests models until they break. You should expect tension when a company ships powerful AI with strict limits. You should also expect questions that go far beyond whether a bot will answer schoolwork.
Teachers and parents report basic biology homework being blocked
Classroom prompts that used to sail through now get flagged as dangerous and routed to Opus 4.8. Anthropic released Fable 5 on Tuesday as a “Mythos-class” model, positioned as the most capable tool it had made generally available. But the company also stressed that when prompted about sensitive topics such as cybersecurity, biology, chemistry, or other high-risk areas, Fable 5 will politely decline and fall back to Opus 4.8.
Why is Fable 5 blocking basic biology questions?
You’ll hear that the model’s hypersensitivity was deliberate. Anthropic wrote that it tuned safeguards conservatively “to release the model both safely and quickly.” That conservative tuning means false positives: harmless school questions flagged as potential threats. I’ve seen a third-grade cell question treated like a red team probe.
On X and Reddit the complaint sounded the same: a helpful assistant turned risk-averse sentinel. That trade-off—safety versus everyday usefulness—is at the center of this standoff.
A developer watched usage climb 2% per minute on their $200/month plan (€185)
One Max 20x subscriber posted that Fable 5 was burning through their subscription at roughly 2% per minute. That’s the kind of discovery that stops you. Fable charges twice the token cost of Opus 4.8 for both inputs and outputs, so longer “agentic” sessions—where the model runs multi-step tasks—eat through allowances fast.
How much does Fable 5 cost?
The model is included with Pro, Max, Team, and seat-based Enterprise plans for now, but Anthropic said it will move to a pay-as-you-go credits model starting June 23. For developers used to fixed monthly windows, that switch feels like a tax on ambitious projects: the most powerful assistance becomes metered. Think of it like a toll booth on a highway of ideas.
Early testers noticed Mythos access was limited under Project Glasswing
Anthropic rolled out Mythos to small groups under Project Glasswing to map cybersecurity risks before a broader release. The observation was simple: access was gated, and some organizations got in early while others did not. That rollout pattern stoked a familiar worry—power concentrated among those with the most security and money.
People on X warned that pay-as-you-go will create winners and losers. One user wrote, “The permanent underclass everyone keeps tweeting about has a start date now: June 23.” The fear isn’t abstract: it’s that an arms race of compute and model access creates an advantage that compounds as models improve.
Security concerns and compute limits drove the rollout
Anthropic says safety is the main reason Mythos is limited; observers point out that compute economics matter too. Running and training models like Mythos and Fable isn’t free for the company. A sudden flood of long-running agent tasks could strain capacity and make latency worse for everyone.
So Anthropic did two things: ship a heavily guarded, subsidized Fable 5 to subscribers while keeping Mythos 5 for controlled testers, then move to metered pricing. From a platform operator’s view, it’s a pragmatic throttle. From a developer’s view, it can feel like exclusion.
Developers and companies are already adjusting workflows
Teams reported changing how they allocate tasks between Opus 4.8 and Fable 5 to avoid token shocks. That small operational pivot is a real-world adaptation—shifting heavy lifting to cheaper models and reserving Fable for hard problems. It’s an early glimpse of how people will ration access when compute becomes expensive.
If you’re building a startup or running a dev team, plan for two levers: budget and model hygiene. Budget for pay-as-you-go credits after June 23; establish prompts and checkpoints so your agentic sessions are lean. I’d also watch how OpenAI and other vendors respond—competition will shape pricing and capabilities.
Will pay-as-you-go create an AI elite?
The worry is legitimate: when superior models are costly, organizations with capital can iterate faster, ship smarter products, and hire talent that leverages powerful AI. That creates a feedback loop where winners widen their lead. But markets and regulation can curb extreme concentration—if policymakers and platforms choose to act.
Anthropic’s conservative approach to safety and staged access is both a firewall and a gate. It protects against misuse, but it also creates scarcity. I don’t want to moralize; I want you to see the mechanics so you can plan.
Community friction is already shaping the conversation
Reddit threads filled within hours of Fable’s release, with developers calling the model “completely unusable” and pleading for relaxed guardrails. Anthropic acknowledged the problem and promised to reduce false positives as models improve. That is a public relations and product problem wrapped together.
You should expect more of these moments—companies balancing customer satisfaction, security, and economics under public pressure. If Anthropic loosens constraints too fast, it risks new vectors of harm. If it stays strict, it risks alienating the developers who make its models useful.
I’ve watched similar inflection points at OpenAI when capability jumps forced reassessments of content policies and pricing. Your choice as a user or leader is to test carefully or budget for the future, because cheap access to powerful models may be increasingly rare.
Two metaphors have been left on the table to describe what follows: Fable’s filters are like a museum guard who locks the door at the faintest whisper, and the pricing change is like a toll booth on a highway of ideas. These images aren’t meant to frighten you; they’re meant to clarify the trade-offs.
I’ll keep testing, and you should, too—pay attention to token burn, policy updates, and how companies such as Anthropic, OpenAI, and their peers frame safety versus accessibility. Who will control access to the fastest AI tooling: a cooperative ecosystem or a gated few?
You’re not even allowed to ask Fable about basic biology questions, let alone anything that could potentially be dangerous. pic.twitter.com/FOlGpPJqsB
— Crémieux (@cremieuxrecueil) June 9, 2026