On June 9, 2026, Anthropic simultaneously announced Claude Fable 5 — a Mythos-class model made safe for general use — and Claude Mythos 5, its full-capability twin reserved for trusted cyberdefenders in Project Glasswing. This is the first time a Mythos-class model has been publicly available to anyone.
The backdrop: Anthropic unveiled Claude Mythos Preview in April 2026 and immediately declined to release it publicly, citing the model's ability to autonomously chain zero-day exploits across every major OS and browser. Two months later, Anthropic says it has built robust safeguards that route high-risk queries to Claude Opus 4.8 — and the resulting Fable 5 now ships to all Pro, Max, Team, and Enterprise subscribers.
🔑 Key Distinction: Fable 5 and Mythos 5 share the same underlying model weights. The difference is purely at the guardrail layer: Fable 5 auto-routes sensitive cybersecurity, biology, chemistry, and model-distillation queries to Opus 4.8, while Mythos 5 lifts those restrictions for vetted partners.
Benchmark Breakdown
The numbers below are from Anthropic's official evaluation table. Every figure is verified against the published announcement.
Claude Fable 5 vs. the Competition — Full Benchmark Table
All scores sourced from Anthropic's June 9, 2026 announcement.
| Benchmark | Fable 5 | Opus 4.8 | GPT 5.5 | Gemini |
|---|---|---|---|---|
| Agentic CodingSOTA SWE-Bench Pro | 80.3% | 69.2% | 58.6% | 54.2% |
| Agentic CodingSOTA FrontierCode Diamond (xhigh) | 29.3% | 13.4% | 5.7% | — |
| Agentic CodingSOTA Terminal-Bench 2.1 | 88 | 82.7% | 83.4% | 70.7% |
| Knowledge Work GDPval-AA (score) | 1932 | 1890 | 1769 | 1314 |
| Knowledge Work Vision GDP.pdf (no tools) | 29.8% | 22.5% | 24.9% | 16.7% |
| Spatial ReasoningSOTA Blueprint-Bench 2 | 38.6% | 14.5% | 36.2% | 26.5% |
| Tool UseSOTA AutomationBench | 17.4% | 15.5% | 12.9% | 9.6% |
| Computer Use OSWorld-Verified | 85 | 83.4% | 78.7% | 76.2% |
| Computer Use (Mythos)MYTHOS OSWorld-Verified | 85.4% | — | — | — |
| LegalSOTA Legal Agent Benchmark | 13.3% | 10.4% | 2.1% | 0 |
| MultidisciplinarySOTA Humanity's Last Exam (no tools) | 59 | 49.8% | 41.4% | 44.4% |
| CybersecuritySOTA ExploitBench (Cap%) | 78 | 40 | 34 | — |
| HealthSOTA HealthBench Professional | 66 | 56.9% | 51.8% | — |
| BiologySOTA BioMysteryBench (human solved) | 83.9% | 80.4% | — | — |
* Cybersecurity, Biology, Health benchmarks show a larger gap because Fable 5 includes fallback routing to Opus 4.8 on sensitive inputs. Mythos 5 (no fallback) scores higher. Computer Use top score: Mythos 5 (85.4%) vs Fable 5 (85.0%).
What Makes This a Generational Leap
The numbers are impressive, but the more important shift is qualitative: Fable 5 can sustain focus across millions of tokens in long-running tasks. Previous Claude models degraded in quality mid-task. Fable 5 doesn't — it actively uses notes and memory to improve its outputs over time.
Stripe reported Fable 5 performed a codebase-wide migration across 50 million lines of Ruby in a single day — work that would take a whole engineering team over two months by hand.
Earlier Claude models needed complex helper harnesses with maps and navigation aids to play Pokémon FireRed. Fable 5 completed the entire game using raw vision screenshots alone — no maps, no game-state extras.
Anthropic's internal protein design experts found Mythos 5 accelerated aspects of drug design by ~10x. The model autonomously chose binding sites, ran protein design tools, and recovered from failures — matching skilled human researchers.
Mythos 5 is Anthropic's first model to consistently produce compelling novel scientific hypotheses. In blinded comparisons, scientists preferred Mythos 5's molecular biology hypotheses ~80% of the time over Opus-class models.
"Fable 5 changed how we work on the Claude Code team day to day. We used to verify that Claude did the work right. Now we verify that it's doing the right work."
— Felix Rieseberg, Claude Code & Cowork Lead, Anthropic
The Safety Architecture
The most technically interesting aspect of Fable 5 isn't the benchmarks — it's the safety architecture that makes public release possible. Anthropic built a classifier-based fallback system: when a request is flagged as touching cybersecurity, biology, chemistry, or model distillation, the response silently routes to Claude Opus 4.8 instead.
⚠️ Fallback Mechanics: The fallback triggers in less than 5% of sessions on average, per Anthropic. Anthropic acknowledges these safeguards are tuned conservatively and will sometimes catch harmless requests. Users see the fallback in the UI and are billed at Opus 4.8 prices for those responses.
For Mythos 5, safeguards are selectively lifted for Project Glasswing partners (US government, critical infrastructure providers, and selected biology researchers). Mythos 5 is priced identically to Fable 5 at $10/M input tokens and $50/M output tokens — less than half the price of Mythos Preview.
Fable 5 vs Mythos 5 — At a Glance
Claude Fable 5
- → Available to all Pro, Max, Team & Enterprise users
- → Auto-fallback to Opus 4.8 on high-risk queries
- → API model string:
claude-fable-5 - → Available on Claude.ai, API, Claude Code, AWS, GCP, Azure Foundry
- → Free for subscribers until June 22, 2026
- → $10/M input · $50/M output (90% discount with prompt caching)
Claude Mythos 5
- → Restricted to Project Glasswing partners
- → No cybersecurity / bio safeguard routing
- → Strongest cybersecurity model in the world
- → Available via Project Glasswing upgrade from Mythos Preview
- → Broader trusted-access program coming soon
- → Same price as Fable 5; mandatory 30-day data retention for safety
Industry Response
Early partners were given access weeks ahead of the public launch. Their assessments:
Cursor reported that Fable 5 is now state-of-the-art on CursorBench — scoring 72.9%, eight percentage points above the previous best — and described it as opening up "a class of long-horizon problems that were out of reach for earlier models."
GitHub said Fable 5 handled complex, long-horizon coding tasks "with a level of autonomy and reliability that exceeded previous benchmarks" and sees it pointing toward a future where developers can hand increasingly ambitious work to agents and trust the results across the software lifecycle.
Lovable integrated Fable 5 on launch day. Replit called it "a clear step forward on agentic coding and prototyping." Legal AI firm EvenUp noted in blind review its legal redlines matched or beat their current model every time.
What This Means for Developers
If you're building on the Claude API, the most important practical changes are:
Model string: Use claude-fable-5. The model is available in Claude Code via the /model command or by setting model: claude-fable-5 in your .claude/settings.json.
Agentic workflows: Feedback loops via /goal in Claude Code or Outcomes in Claude Managed Agents work identically to previous models. The key difference is Fable 5's dramatically improved ability to self-correct over extended task chains.
Multi-agent orchestration: In Claude Managed Agents, Fable 5 can be used as the orchestrator, delegating subtasks to smaller (cheaper) models, with the full Mythos-class reasoning reserved for planning and synthesis.
Cost: At $10/M input and $50/M output, Fable 5 is double the price of Claude Opus 4.8 — but prompt caching applies a 90% discount on cached tokens, which in long-context agentic tasks can bring effective costs below Opus pricing.
📌 Availability Note: Fable 5 is included at no extra charge for Pro, Max, Team, and Enterprise subscribers until June 22, 2026. After June 23, usage credits may be required until capacity expands. The Anthropic API supports Fable 5 immediately with standard consumption-based billing.
Bottom Line
Claude Fable 5 isn't an incremental update — it's a new capability tier. The model closes the gap between AI assistants and AI agents: it can take on tasks measured in days, not seconds, and stay reliable across them. The benchmark numbers are the most concrete evidence, but the qualitative shift (Stripe compressing months of engineering, Mythos 5 beating human scientists on drug design) is where the real story lives.
The safety architecture — fallback routing rather than capability reduction — is arguably Anthropic's most interesting engineering decision here. It lets them ship a Mythos-class model to the general public without waiting for perfect classifiers. It introduces new failure modes (false positives on legitimate queries), but the tradeoff buys access to the most capable publicly available AI model in the world, today.
