The Short Version

Claude Fable 5 is the consumer-safe release of Anthropic's Mythos architecture — the same underlying model the company previously reserved for restricted government and security partners. "Fable" is Mythos with cybersecurity safeguards left on; Mythos 5 is the identical model with those guardrails removed for authorized users only (Project Glasswing — US government cyber defenders and infrastructure providers).

The headline: on nearly every published benchmark of raw capability, Fable 5 is the strongest model anyone outside a lab had ever been able to call via an API. It didn't edge out the competition — it cleared it by double digits on coding.

80.3%
SWE-bench Pro — vs 58.6% for GPT-5.5
#1
Artificial Analysis Intelligence Index at launch
90%+
first model ever to break 90% on Anthropic's core analytics benchmark
1 day
to migrate a 50M-line Ruby codebase (≈2 months by hand)

Benchmarks: How Fable 5 Compares

The numbers below are the published launch benchmarks. The coding gap is the story — an 80.3% on SWE-bench Pro is roughly 11 points above the next-best generally available model (Opus 4.8 at 69.2%) and more than 20 points above GPT-5.5 and Gemini 3.1 Pro.

BenchmarkFable 5Opus 4.8GPT-5.5Gemini 3.1 Pro
SWE-bench Pro (coding)80.3%69.2%58.6%54.2%
Core analytics (long-running)90%+~80%
GDP.pdf (vision/document)29.8%22.5%24.9%16.7%
Hebbia Finance (senior reasoning)Highest of any model

A few things worth flagging about these numbers. The vision/document scores look low across the board because GDP.pdf is a deliberately brutal benchmark — dense, real-world financial PDFs. Fable still leads it. And the finance and analytics gains aren't toy-benchmark wins: they map directly onto the kind of multi-step, document-grounded reasoning that real knowledge work depends on.

Reality check: Benchmarks measure ceiling, not daily experience. For most chat and writing tasks you would struggle to feel the difference between Fable 5, Opus 4.8 and GPT-5.5. The gap shows up on hard, long, multi-step agentic work — large refactors, codebase migrations, deep research, financial modeling.

What's Actually New

🧠
Millions-of-tokens context
Built for very large context windows with persistent file-based memory, so it holds state across long agentic sessions instead of "forgetting" mid-task.
🛡️
Safety routing
Automatically reroutes to Opus 4.8 when classifiers detect cyber-exploit, dual-use bio/chem, or model-distillation queries. Reportedly triggers in <5% of sessions.
📊
Tool-free document reasoning
Interprets charts, tables and dense PDFs natively, without external parsing tools — a big deal for finance and research workflows.
⚙️
Mythos-class engine
The same model the security version (Mythos 5) uses — described by Anthropic as the strongest cybersecurity capabilities of any model in the world.

Pricing: The Catch

Fable 5 is not cheap. At launch it was priced at $10 per million input tokens and $50 per million output tokens — double Opus 4.8's rates ($5 / $25).

ModelInput / 1MOutput / 1M
Claude Fable 5$10$50
Claude Opus 4.8$5$25
GPT-5.5$5$30

The economic argument for Fable was never per-token price — it was tasks completed per dollar. If one model finishes a two-month migration in a day on the first try, the headline rate stops mattering. But for high-volume, simple workloads, the routing fallback to Opus and the 2× price made Opus 4.8 the saner default for most teams anyway.

Availability today: Fable 5 and Mythos 5 were suspended on June 12, 2026 — three days after launch — following a US government export-control directive restricting access to US nationals. If you're evaluating it, the practical answer right now is: you probably can't. We break down exactly what happened and why in the article linked below.

Should You Build on Fable 5?

Assuming access returns, the decision is straightforward:

For a full side-by-side of the entire June 2026 lineup — not just Fable — see our best AI models right now breakdown, or jump straight to the interactive AI Model Selector and benchmark explorer.

One prompt, many models — keep them organized

Models change every few weeks. PromptChief lets you save prompt versions per model, so when the next frontier release drops you're refining, not rewriting. Store, test and reuse your best prompts across Claude, GPT and Gemini.

Try PromptChief Free →