On this page (7 sections)
Quick answer
What is Claude Fable 5 and who is it for? Claude Fable 5, released 9 June 2026, is Anthropic's most capable widely released model. It carries a 1M-token context window, always-on reasoning, and top-of-chart benchmark scores, at $10 per million input tokens and $50 per million output — double Opus 5. It's built for the most demanding long-horizon agent work; for most everyday coding and building, the cheaper Opus 5 is the smarter choice.
Key takeaways
- Released 9 June 2026 as Anthropic's most capable widely released model — 'next-generation intelligence for long-running agents'.
- Priced at $10/$50 per million tokens, roughly double Opus 5, with a 1M context window and 128K output.
- Thinking is always on and can't be turned off; you control depth with the effort parameter.
- It ships with safety classifiers that can decline requests — a real integration difference from Opus 5.
- For most work, Opus 5 gets you ~95% of the capability at half the price. Fable 5 is the ceiling, not the default.
at a glance
| Released | 9 June 2026 |
| Model ID | claude-fable-5 |
| Pricing | $10 / $50 per million tokens (in / out) |
| Context / output | 1M tokens in · 128K out |
| Thinking | Adaptive, always on (can't disable) |
| Data retention | 30 days required (no zero-retention) |
Claude Fable 5 is the model at the very top of Anthropic’s lineup. When it landed on 9 June 2026, Anthropic described it as its most capable widely released model, built for “the most demanding reasoning and long-horizon agentic work.” A month later Opus 5 arrived and undercut it on price, which reframed the question around Fable 5: not is it the best, but when is the best worth paying for.
This breakdown answers that. What Fable 5 is, what genuinely sets it apart from Opus 5, and the honest case for when to spend double.
How we cover this. These are frontier models we analyze from Anthropic’s documentation and independent benchmark reporting, not tools we run through our own build brief. Every spec here is dated and sourced.
What makes Fable 5 the flagship
Fable 5 is what Anthropic calls “Mythos-class” — the tier above the standard Opus family. In practice that shows up in three ways.
It always thinks. Adaptive reasoning is permanently on; you can’t disable it. Instead of a thinking on/off switch, you tune depth with the effort parameter. That’s a design choice aimed at hard, multi-step problems where reasoning is the point. Anthropic also never returns the raw chain of thought — you get a summary or nothing, not the model’s internal working.
It’s built to run for a long time. The positioning is “next-generation intelligence for long-running agents.” With a 1M-token context window, the memory tool, code execution, compaction and task budgets all supported at launch, Fable 5 is aimed at agents that stay alive across many steps and hold a lot of state.
It has a twin you probably can’t use. Claude Mythos 5 shares Fable 5’s specs and price but is limited to approved customers through Project Glasswing for defensive security. The difference is the safety layer: Fable 5 carries classifiers that can decline requests; Mythos 5 doesn’t. For everyone outside that program, Fable 5 is the model, and it offers the same core capability.
The benchmarks
Fable 5 sits at the top of Anthropic’s own charts, and independent testing agrees it’s a step up from the Opus family. On SWE-bench Pro — the agentic coding benchmark most people watch — reporting from Vellum and Finout has put Fable 5 around 80%, comfortably ahead of Opus 4.8. Its lead is clearest on long, multi-step autonomous tasks, which is exactly the workload it’s designed for.
The nuance that matters in 2026: Opus 5 closed most of the gap. On Cursor’s CursorBench, Opus 5 lands within roughly half a percent of Fable 5 at half the price, and on the OSWorld computer-use benchmark Opus 5 actually comes out ahead. So Fable 5’s real edge is narrower than the price difference suggests — it shows up at the extreme end, not across the board.
The one thing developers must plan for
Fable 5 ships with safety classifiers that can decline a request. When that happens the API returns a refusal as a normal, successful response — not an error — and tells you which classifier fired. It’s aimed mainly at high-risk cyber and bio content, but any integration on Fable 5 needs to handle that path, and Anthropic provides server-side and client-side fallbacks to retry on another model.
Two more operational notes. Fable 5 requires 30-day data retention and can’t be used under zero-retention policies, which rules it out for some teams. And it was briefly withdrawn and then redeployed shortly after launch, so availability has settled but the rollout wasn’t perfectly smooth. Opus 5 carries none of these constraints, which is part of why it’s the easier default.
Fable 5 vs Opus 5, side by side
| Fable 5 | Opus 5 | ||
|---|---|---|---|
| Tier | Flagship (Mythos-class) | Opus | |
| Price (in / out) | $10 / $50 | $5 / $25 | |
| Released | 9 Jun 2026 | 24 Jul 2026 | |
| CursorBench | Top | Within ~0.5% | |
| OSWorld | Strong | Ahead, at ⅓ the cost | |
| Safety refusals | Yes — plan for them | Fewer restrictions | |
| Best for | Hardest long-horizon agents | Everyday coding & agents |
When Fable 5 is actually worth it
Frontier autonomous agents
Long-running, multi-step agents where the last few percent of reasoning quality changes the outcome, and the extra cost is a rounding error next to the value. This is the job it was built for.
See where it fits →Everyday coding & building
For normal coding, chat and app-building, Opus 5 gets you within a whisker of Fable 5 at half the price, with no refusal handling to build around. Use it.
Read the Opus 5 breakdown →Zero-retention / restricted data
Fable 5 requires 30-day data retention and can’t run under zero-retention policies. If that’s a hard constraint, Opus 5 or Sonnet 5 is your ceiling.
Compare the tiers →How this ties back to the tools you build with
You’ll rarely pick Fable 5 directly inside an app builder — the builders we review optimize for the cost and speed of everyday models, which in 2026 usually means Opus 5 or Sonnet 5 rather than the flagship. Where Fable 5 earns its place is bespoke, high-stakes agent systems you’re wiring up yourself. If you’re building normal software, the practical move is a current mid-tier Claude model inside a tested app builder.
The bottom line
Fable 5 is the most capable model Anthropic ships, and if you’re running the kind of long-horizon autonomous agent it was designed for, it earns the premium. For everyone else, the arrival of Opus 5 made the choice easy: you get the vast majority of Fable 5’s capability at half the price and without the refusal handling. Treat Fable 5 as the ceiling you reach for deliberately, not the model you default to.
Frequently asked questions
How much does Claude Fable 5 cost?
Claude Fable 5 vs Opus 5 — which should I use?
What does 'Mythos-class' mean?
Why does Claude Fable 5 refuse some requests?
Does Claude Fable 5 support a large context window?
Sources & data
- Anthropic — Introducing Claude Fable 5 and Claude Mythos 5 (9 Jun 2026)
- Anthropic — Models overview (specs & pricing)
- Benchmark reporting: Vellum and Finout (2026)
Review changelog
Jul 2026 — Published alongside the Opus 5 launch coverage; specs from Anthropic docs