Anthropic held back its most powerful model for months because it was too good at breaking into software. Then they gave it safety guardrails, cut the price by 60% from the Project Glasswing rate, and dropped it on June 9 with a two-week free window that ends on June 22. After that, every query costs usage credits.
Twelve days to use the best model on the market at no extra charge. That is the deal. The fine print is that after June 22, if you’re on a $200 Pro Max plan, you’re paying subsidized credits. If you’re on a standard plan, you’re paying full price. The window closes the same day.
What Fable 5 Actually Is
Claude Fable 5 is Anthropic’s first publicly available Mythos-class model, released June 9, 2026 — a tier above Claude Opus 4.8, priced at $10/$50 per million input/output tokens, which is exactly double what Opus 4.8 charges.
The two-model structure is the part that needs explaining. Claude Fable 5 and Claude Mythos 5 share identical weights. Mythos 5 is the restricted version handed to vetted cybersecurity researchers and infrastructure providers through Project Glasswing — the same classifiers lifted, the same underlying model. Fable 5 is what ships to the public, wrapped in safety classifiers that reroute high-risk queries to Opus 4.8. So when you’re using Fable 5 and it hands off your question to a noticeably less capable model, that’s the safety layer activating, not a bug.
Earlier in 2026, the Mythos Preview had already demonstrated the ability to find vulnerabilities in every major operating system and every major web browser. Mozilla used it to patch 271 Firefox vulnerabilities in two weeks. That’s what’s sitting underneath the public version, with classifiers on top.
The Benchmarks That Actually Matter
Most AI benchmarks are marketing dressed up in scientific formatting. The one worth paying attention to here is SWE-Bench Pro — a coding evaluation that tests whether a model can resolve real GitHub issues in large, production-grade codebases.
Fable 5 scores 80.3% on SWE-Bench Pro and 95% on SWE-Bench Verified. For context, Opus 4.8 sits at 69.2% on SWE-Bench Pro — Fable 5’s lead over the previous flagship is 11 points on the benchmark most developers actually trust.
The Cursor Bench numbers from independent developer testing confirm the same story: Fable 5 Max scores 72.9%, with GPT-5.5 Extra High at 64.3% as the closest competitor. The cost to reach those scores is what separates them — $18 per task for Fable 5 Max versus $4.37 for GPT-5.5 Extra High. Best-in-class, clearly. Cheapest, not even close.
What It Did in Real Codebases
The benchmark gap between models tends to collapse when you throw them at actual production bugs. Fable 5 doesn’t collapse.
One developer testing the model reported three separate bugs in a large codebase — a broken chat stream websocket, a non-functional computer use agent, and an iMessage connection that delivered messages without actually sending them. All three had defeated both GPT-5.5 and Opus 4.8. Fable 5 fixed each one in under 30 minutes from a lazy prompt — error message copy-pasted, minimal context, no detailed specification.
The code review quality confirmed the same pattern. Greptile, an AI code review tool that assigns confidence scores from one to five on pull requests, returned near-perfect scores on Fable 5’s output — multiple consecutive fives out of five, with almost no fours. Greptile’s confidence score is binary for most developers: under four means the PR isn’t ready to merge. Fable 5 consistently cleared that bar.
The UI Generation That Caught Everyone Off Guard
The second signal that this model is different came from front-end generation tests, which typically expose the gap between what AI models claim and what they actually render.
Five prompts, one shot each. A designer portfolio landing page with scroll-based animations. A Three.js gallery reconstruction from a screenshot and URL. A hero section redesign with a Figma MCP agent integration. A Craigslist modernization with hover interactions. A recreation of a GSAP and Three.js award-winning UI from a video reference.
The portfolio page came back with split-character text animations, timed easing on scroll reveals, and a working gallery section. The Three.js gallery rebuilt the concept from the reference URL — not a pixel copy, but a structurally accurate interpretation. The Craigslist refresh maintained the type-heavy density of the original while adding clean hover states. On five consecutive one-shot UI tests, Fable 5 went five for five.
That’s not the same as replacing a senior designer. A trained eye can tell the difference between original work and a model’s interpretation, and the award-winning site reconstruction exposed the ceiling clearly — good enough to be immediately useful as a base, not good enough to pass as the real thing without revision. But the gap between “immediately useful” and “requires days of custom code” is exactly where the tool earns its price.
The Pricing Problem Nobody Is Solving
At $10 per million input tokens and $50 per million output tokens, Fable 5 is double the price of Opus 4.8 and double GPT-5.5 on input. One developer who described himself as an AI skeptic — someone who had spent months arguing that current models were overhyped — signed up for the $200 Max plan, used Fable 5 to rebuild the entire visual metaphor of a productivity app, and blew past his monthly limit in two hours.
His complaint afterward was not that the model failed. It was that the model worked so well it removed the creative friction he’d built his entire developer identity around. That is a different category of complaint than “the model is bad.”
The compute constraint is real. Anthropic is spending close to a billion dollars a month on GPUs to run this infrastructure. The June 22 cutoff for free access on Pro plans exists because they physically cannot subsidize open-ended usage at this scale. The plan is to restore it as a standard feature when capacity allows — no committed date given.
Subscription plans — Pro, Max, Team, and seat-based Enterprise — include Fable 5 at no extra cost through June 22, 2026. Usage credits are required after June 23. For Cowork users specifically, the 5-hour usage limit in Cowork is doubled from June 5 through July 5.
The window is open. It closes in 12 days.
Sources
- Anthropic News: Claude Model Releases & Product Announcements
- SWE-bench: Software Engineering Evaluation Benchmarks for LLMs
- Greptile AI: AI-powered Code Review & Confidence Analytics
About the Author
Your 32-year-old tech brother who has burned through four AI subscriptions this quarter and refers to compute costs as “the vibe tax.”
