Anthropic Brings Mythos to the Masses with Claude Fable 5
Anthropic released Claude Fable 5, the most capable model it has ever made generally available and the first to carry its “Mythos-class” tier — the rung above the Opus class. It ships with always-on adaptive thinking, a 1-million-token context window, and 128K output tokens, and Anthropic says its lead over previous models widens on longer, harder tasks across software engineering, knowledge work, vision, and scientific research.
The timing is pointed. Fable 5 arrives just days after Anthropic publicly warned that frontier AI is becoming too dangerous and confidentially filed its IPO paperwork with the SEC. The company is threading both needles at once: ship the most powerful public model to date, and frame the release as proof that strong capability and hard safety limits can co-exist.
Same Model, Two Doors: Mythos 5 Stays Behind Trusted Controls
Fable 5 and Mythos 5 are the same underlying model — the difference is access control. Fable is the public version with safety classifiers switched on: in high-risk domains like cybersecurity, biology, chemistry, and distillation, it blocks responses and falls back to Claude Opus 4.8. Anthropic reports that at least 95% of Fable sessions ran entirely on the model’s own responses, with the fallback rarely tripping.
Mythos 5 is the unrestricted twin, and it is not for sale to the public. It’s offered only to approved customers — Glasswing cyber partners with the cyber safeguards lifted, and soon select biology researchers with bio and chemistry limits relaxed — through Anthropic’s Project Glasswing program. The bet is that a capability this strong can be broadly useful without being broadly dangerous, as long as the riskiest doors stay locked.
80.3% on SWE-Bench Pro — 21 Points Clear of GPT-5.5
Fable 5’s headline number is coding. It scored 80.3% on SWE-Bench Pro, which tests end-to-end resolution of real GitHub issues — about 21 points ahead of GPT-5.5 (58.6%), 11 ahead of Opus 4.8 (69.2%), and well clear of Gemini 3.1 Pro (54.2%). It also tops the field on knowledge work (GDPval-AA 1932 vs GPT-5.5’s 1769), tool use, legal, and spatial reasoning, landing at 96 to GPT-5.5’s 90 on the provisional aggregate.
The caveat worth keeping: SWE-Bench Pro measures bounded tasks. Several developers noted that the harder axis is sustained reliability across long sessions, where a model has to track architectural decisions made hundreds of turns earlier. That’s exactly where the 1M-token context and always-on thinking are supposed to pay off — but it’s also where benchmark numbers say the least.
$10/$50 Per Million Tokens, Free on Paid Plans Until June 22
Developers can call the model now via claude-fable-5 on the Claude API. Pricing is $10 per million input tokens and $50 per million output — double Opus 4.8, and notably above GPT-5.5’s $5/$30. From launch through June 22, Fable 5 is bundled into Pro, Max, Team, and seat-based Enterprise plans at no extra cost.
That free window has a hard edge. On June 23, Anthropic removes Fable 5 from those plans, and continued use requires usage credits until capacity expands. If you’re prototyping on a subscription this week, treat the included access as a trial, not a baseline — budget for the credit model before the cutover, especially for any agent or batch workload that burns output tokens at scale.
Day-One Everywhere: Bedrock, Cursor, Copilot, and More
The ecosystem rollout was immediate. Fable 5 landed on Amazon Bedrock at launch and appeared the same day in Cursor, Devin, Notion, Microsoft Foundry, GitHub Copilot, Cline, Replit, Base44, MagicPath, and MCP Atlas, among others. For most teams, adopting it isn’t a migration — it’s a dropdown change in the tools they already use.
That speed is its own signal. A coordinated day-one launch across the major clouds and coding tools means the integration work was lined up well in advance, and it short-circuits the usual lag between a frontier release and real-world availability. The friction now isn’t access; it’s cost discipline once the free window closes.
Early Hands-On: “First to 90%” and Better at One-Shotting Apps
Early adopters reported large real-world deltas. Analytics firm Hex said Fable was the first model to clear 90% on its core analytics benchmark for complex, long-running tasks, praising its “judgement and attention to nuance” on the hardest questions. Vibe-coding platform Base44 called it the best yet at one-shotting full apps with excellent tool-calling, and one developer reported Fable not only solving an architectural problem but also implementing four improvements to their LLM library along the way.
Simon Willison and others published first impressions within hours, and the AI-newsletter circuit flagged Fable’s release terms as “controversial” — a reference to the June 22 free-access expiry and credit model that left some users confused about what “included” actually meant until Anthropic staff clarified the rollout.
The Whole Release Is an Argument About Safety
Strip away the benchmarks and Fable 5 is really a thesis statement. Anthropic spent the prior week warning that AI is getting too dangerous, then turned around and shipped its most powerful public model. The reconciliation is the architecture itself: take the Mythos model, bolt on classifiers that block the genuinely dangerous domains, fall back to Opus 4.8 when they trip, and release the rest. The Mythos/Fable split — same brain, different doors — is the cleanest expression yet of “capability and restraint are separable.”
Whether that’s reassuring depends on how you read the 95% figure. Anthropic frames “95% of sessions ran entirely on Fable” as evidence the safeguards rarely get in the way. Read the other direction, it means roughly 1 in 20 sessions hit a guardrail — and the only thing standing between a public model and Mythos-class capability in a high-risk domain is a classifier the company is confident in. For regulated buyers, that’s the question that matters more than SWE-Bench: not how smart the model is, but how much you trust the gate.
The Real Decision Is the June 23 Cliff, Not the Benchmark
It’s easy to get lost in the 80.3%. The operational question for most teams is simpler: Fable 5 is free on your plan until June 22, then it isn’t. At $10/$50 per million tokens — double Opus 4.8 — an agent pipeline that runs hot on output can get expensive fast, and the included-access window is short enough to be a trap if you build around it without a budget.
The sensible play this week is to use the free window as a controlled evaluation: run Fable against the workloads where its coding and long-context edge actually moves your numbers, measure the token cost honestly, and decide where it earns the premium over Opus 4.8 versus where Opus is good enough. Treat “which model for which job” as a routing decision, not an all-in switch — the price gap makes selectivity the whole game.