Anthropic ships Fable 5 — the first Mythos-class model and, fairly undisputedly, the best AI we've ever used. But at this level, getting the most out of state-of-the-art isn't your old prompts with a new model: it's controversial guardrails, usage-based reality, and a new skill called task imagination.
Fable 5 doesn't just beat the benchmarks — it raises the bar for ambition.
Anthropic's first Mythos-class model is a genuine leap, especially on agentic coding, where it one-shots things that used to take teams months. But the real story is what it demands of us: in a usage-based world, we have to become token-efficiency optimizers who match models to use cases, and we have to develop "task imagination" — the ability to hand an agent days of responsibility, not minutes of tasks. The frontier has moved from answers to tasks to responsibilities, and most of us aren't dreaming big enough to use it.
On Tuesday June 9th, Anthropic released Claude Fable 5 — NLW calls it fairly undisputedly the best AI model we've ever been able to use. It arrives just weeks after Opus 4.8, which still plays a big role in the Fable-led ecosystem.
Mythos 5 is effectively the same model as the public Fable 5, minus the controversial safeguards. It's available only through Project Glasswing, deployed in collaboration with the US government, with broader "trusted access" promised later.
Anthropic added "Fable" as a class above haiku, sonnet, and opus. It's also the first time since GPT-5's disastrous August rollout that any lab put a full new base number on a model — the turn-of-2026 jumps were all 4.5/4.6. The naming alone signals Anthropic isn't playing.
NLW usually distrusts saturated benchmarks, but the leaps here are real: 80.3% on SWE-bench Pro, 88% on Terminal Bench, and a top blended ranking from Artificial Analysis. The model's clear purpose is agentic coding.
The new benchmark combines unit tests with assessments of scope, discipline, style, and codebase standards — measuring not just whether code works, but whether it's good enough to merge into production. Fable 5 scored 29.3%, more than doubling Opus 4.8's 13.4%.
— Sean Wang, Cognition / Latent Space. Cognition's Sean Wang's framing for why Frontier Code is needed: even code that nominally solves a problem is often unusable by the organization running it.
API costs are double Opus, and Anthropic is positioning subscription access as an introductory offer: Fable will be removed from Pro-tier plans on June 23rd, after which access is pay-per-usage. More evidence we're in a firmly usage-based pricing paradigm.
Users report Fable refusing or rerouting on basic biology — the word "cancer" flagged as a biosecurity risk, "tell me about mitochondria" pausing the chat. Anthropic admits it ratcheted up bio/chem filtering precisely because the model is more capable.
When Fable's classifiers detect cybersecurity, biology/chemistry, or distillation requests, the response is handled by Opus 4.8 instead, with the user informed. Anthropic says 95% of sessions never hit a fallback, arguing a downgrade beats an outright refusal.
Anthropic added interventions limiting Claude's effectiveness on pre-training pipelines, distributed training, and accelerator design — aimed at actors who'd violate its terms (read: Chinese models distilling its work). NLW sees a dragnet catching legitimate researchers.
— Will Brown, Prime Intellect. Critics including Nathan Lambert and Dean Ball blasted not just the research limits but that they're invisible and undisclosed — "shockingly hostile," said Ball. SemiAnalysis claims models will secretly degrade output quality on interesting ML work.
— Tenebris, on X. The counter-camp says the pearl-clutching is naive — locking down capability and capturing the upside was always the plan. OpenAI staffer Adam GPT quipped, "Well, look at that. OpenAI ends up being the OpenAI lab."
Mythos-class prompts and outputs are retained for 30 days with human review on every platform. Critics warn it could violate NDAs — especially with memory on, which pulls sensitive past chats into context. NLW expects this constraint won't last long.
— John vs Malik, on X. Amid "I'll be out of usage in an hour" panic, others found Fable cheaper than Opus in practice: it costs more per token but one-shots far more often, so users burn less time re-prompting.
— Ali K. Miller. Ali K. Miller called Fable 5 an actual leap with high-performing models that can run 100-plus hours — reframing work away from a 9-to-5 of constant babysitting toward giving complex, goal-oriented prompts and aligning your org on what to kick off.
Riley Brown reported Fable one-shotting "Replit Mobile," a Swift app that builds web apps, and rebuilding a working Lovable clone in two prompts. Skeptics noted a real company is more than an interface — but the speed was a genuine moment.
Matt Schumer said Fable "solved 3D world building" with custom Three.js in-browser. Jake Fitzgerald got a humanoid robot design after 2 hours and 1.4M tokens. Hugging Face's Victor said Fable did an "AGI-level job" on his Boeing 747 Three.js benchmark.
Per Anthropic's launch post, Stripe reported Fable 5 compressing months of engineering into days — performing a codebase-wide migration on a 50-million-line Ruby base in a day that would have taken a team over two months by hand.
— Todd Saunders. Todd Saunders had Claude transcribing a customer call in the background — and building the requested features in real time as the customer described them. Autonomous looped building triggered straight from a sales conversation.
— Citrini Research. Citrini Research captures the new dynamic: state-of-the-art no longer reveals itself across all tasks. The gains show up specifically in things that simply weren't possible before — visible to the people pushing hardest.
NLW's standout everyday win: in a strategic debate, Fable disagreed clearly, then updated its position on new information without collapsing into telling him what he wanted to hear. That alone is a massive upgrade for strategic ideation, a top real-world use case.
In hours of unattended work, Fable rebuilt Superintelligent's audit input system on Whisper, the Agent Transformation Intensive site and platform, and turned AIDB's shareable-nuggets mockups into a real production pipeline. The takeaway: far less management, much bigger ambition.
— Felix Reisberg, Anthropic. Felix Reisberg, who leads Claude Code and CoWork, says a third era quietly began: from answers, to tasks, to loops. He no longer tells Claude to investigate a crash report — it watches every crash and its job is to keep the apps from crashing. He expects 2027 apps to look nothing like today's.
With models that can run for days, NLW argues we'll all have to up-level ambition — and become token-efficiency optimizers who match models to use cases. Nate B. Jones's framing: most of us have nothing that's ever taken even an hour on AI, so the scarce skill is imagining tasks worth handing to a model that works for days.
— Thibault, OpenAI / Codex. Asked whether Anthropic had blown past OpenAI with three models in two months — and Fable isn't even its best — Codex product lead Thibault offered a confident, cryptic reply. NLW: we could be in for quite a week.