Claude Fable 5 Explained: Anthropic's First Claude 5 Model

What Is Claude Fable 5? The Claude 5 Family Begins

For the past two years, Anthropic's lineup followed a familiar ladder: Haiku for speed, Sonnet for balance, Opus at the top. Claude Fable 5 breaks that pattern. According to Anthropic's announcement, Fable 5 is a Mythos-class model — a tier above Opus-class — and the first Claude 5 model to ship.

The Mythos-class story started before the public launch. In April 2026, Anthropic began Project Glasswing, releasing the first Mythos-class model (Claude Mythos Preview) only to a limited group of cyber defenders and critical software infrastructure providers. Fable 5 is the moment that capability tier went public: as TechCrunch and CNBC covered it, Anthropic released a version of Mythos that the public can access — made "safe for general use" through added safeguards.

Anthropic's headline claims are specific: Fable 5 is state-of-the-art on nearly all tested benchmarks of AI capability, with exceptional performance in software engineering, knowledge work, vision, and scientific research. Two claims matter most for agent builders: the longer and more complex the task, the larger Fable 5's lead over other models, and it can work autonomously for longer than any previous Claude model.

Fable 5 vs Mythos 5: Same Model, Different Access

This is the part most coverage gets muddled, so let us be precise. Fable 5 and Mythos 5 are the same underlying model. The names even hint at it — Anthropic notes that "fable" comes from the Latin fabula, akin to the Greek mythos. What separates them is not capability but safeguards and access.

The dual-use safeguards on Fable 5 cover three areas:

• Cybersecurity — detecting queries about discovering and exploiting software vulnerabilities, plus agentic hacking patterns like reconnaissance and lateral movement
• Biology and chemistry — broad safeguards on most bio/chem requests, since Mythos-class models outperformed even protein-specialized models on unpublished tasks
• Distillation — blocking attempts to extract Fable 5's capabilities to train competing models

For everyday development work, the practical impact is small: Anthropic reports more than 95% of Fable sessions involve no fallback at all. The safeguards also held up under scrutiny — an external bug bounty found no universal jailbreaks in over 1,000 hours of testing, and Fable 5 complied with zero harmful single-turn cyberattack requests across 30 public jailbreak techniques.

What's New vs Claude Opus 4.x

Here is how the two tiers compare on what has been publicly verified so far:

Notice the relationship in that last row: Opus 4.8 is not being replaced — it literally serves as the fallback model inside Fable 5's safeguard system, and Anthropic's alignment assessment found Mythos 5's rate of misaligned behavior was low and similar to Opus 4.8's. The two tiers are designed to coexist, which also means model routing (Fable for the hard problems, Opus or Sonnet for routine work) is now a three-tier decision rather than two.

Benchmarks and Pricing: What's Verified So Far

It is early days, and we will only repeat figures with a source behind them. From Anthropic's announcement and launch-partner evaluations:

• State-of-the-art on nearly all tested benchmarks, per Anthropic, spanning software engineering, knowledge work, vision, and scientific research
• Top performer on Cognition's FrontierCode evaluation and the highest score on Hebbia's Finance Benchmark for senior-level reasoning
• First model to break 90% on Hex's core analytics benchmark for complex, long-running analytical tasks
• Nearly saturating ViBench, an end-to-end vibe-coding benchmark — relevant if you follow the vibe coding workflow
• Stripe reported Fable 5 "compressed months of engineering into days," completing a 50-million-line Ruby codebase migration in a day that would otherwise have taken a team over two months

One detail we are deliberately not quoting: a context window figure. Anthropic's announcement does not headline one, and we have not found a verified number — so treat any specific claim you see elsewhere with caution until it appears in the official docs.

Claude Fable 5 in Claude Code

Fable 5 now powers Claude Code, and this is where the "lead grows with task length" claim becomes tangible. Long-horizon coding tasks — the multi-hour refactors and multi-agent team workflows we have covered before — are exactly where Anthropic says the gap over previous models is largest. One launch-partner CTO put it plainly: Fable 5 "delivers more capable engineering in fewer turns than prior models — handling the complex multi-agent workflows our employees run daily in Claude Code."

"Fewer turns" matters more than it sounds. Every turn an agent wastes on a wrong approach burns context and tokens. A model that gets to the correct fix in 3 turns instead of 7 compounds with everything in our 1M context window guide — large context plus stronger long-horizon reasoning is what makes the single-session workflow (research, plan, implement, verify) actually hold together on big codebases. And because Fable 5 costs 2x Opus per token, the habits in our token optimization guide — prompt caching, focused context, routing simple tasks to cheaper models — matter twice as much now.

What It Means for Businesses Building AI Agents

If you are evaluating AI agents for your business, Fable 5 changes two variables in the equation: reliability and economics.

Reliability. The main reason autonomous agents fail in production is not that models cannot write code or draft emails — it is that they drift, lose the plot, or make a bad judgment call somewhere in step 14 of a 20-step workflow. A model whose advantage specifically grows with task length and that can work autonomously longer than any previous Claude directly attacks that failure mode. Workflows that needed a human checkpoint every few steps can now plausibly run end-to-end.

Economics. At $10/$50 per million tokens, Fable 5 is the most expensive Claude per token. But per-token price is the wrong metric for agents — cost per completed task is what hits your budget. A model that finishes in fewer turns, makes fewer errors that need re-work, and escalates to humans less often can be cheaper per outcome despite the higher rate. We break down that math in our AI agent development cost guide; the short version is that model fees are usually a minority of total project cost, so a 2x token rate on the hardest 10% of tasks barely moves the total — while failed automations move it a lot.

The sensible architecture is tiered: Fable 5 for planning, judgment, and the long-horizon steps; Opus or Sonnet for routine execution. If you are outsourcing the build, this is a good litmus test — when choosing an AI agent development company, ask whether they are already building on the Claude 5 family and how they decide which tier handles which step. At Codeloop, model routing is a standard part of how we build custom AI agents.

How to Try Claude Fable 5

1. 1. Claude API — use model ID claude-fable-5. Available everywhere since June 9, 2026.
2. 2. Claude Code — switch with /model and give it a genuinely hard, long-horizon task. Short tasks will not show you the difference; a cross-cutting refactor will.
3. 3. claude.ai — available on Pro, Max, Team, and seat-based Enterprise plans (usage credits apply after the June 22 launch window).
4. 4. Mythos 5 — only via Anthropic's trusted-access programs (Project Glasswing partners today, a biology program planned). For everyone else, Fable 5 is Mythos-class access.

Planning Your Next AI Agent Project?

At Codeloop, we design and build production AI agents on the Claude family — including tiered architectures that put Fable 5's reasoning where it pays off and cheaper models everywhere else. If you want to know what the Claude 5 family means for your roadmap, we can map it out with you.

Talk to Us About Your Project