Anthropic released Claude Fable 5 on June 9, 2026. It was not framed as "the next Opus." It was framed as something new: the first publicly available model in a capability tier Anthropic calls Mythos-class, sitting above Opus.
Three weeks later, on June 30, Anthropic shipped Claude Sonnet 5 and Claude Opus 5 — with Opus 5 described as coming close to Fable 5's frontier intelligence at half the price. So the obvious question for anyone deciding what to build on: does Fable 5 still earn its premium?
The Two-Tier Split
Fable 5 and Mythos 5 are the same underlying model with different safeguards:
- Claude Fable 5 — safety classifiers active, available to everyone (API, Bedrock, Claude Platform on AWS/GCP/Azure Foundry, and Pro/Max/Team/Enterprise plans).
- Claude Mythos 5 — safeguards lifted in specific domains, restricted to Project Glasswing partners (US government-aligned cyberdefense infrastructure providers) and vetted biomedical researchers through a trusted-access program.
You will almost certainly only ever touch Fable 5. It's the one available via the claude-fable-5 model identifier.
What It's Actually Good At
Independent benchmarking puts Fable 5 around 80% on SWE-Bench Pro — a meaningful jump over Opus 4.8, and the widest gap on long, complex, multi-step tasks rather than short ones. Anthropic's own examples lean into that pattern rather than headline numbers:
| Example | Result |
|---|---|
| Stripe, 50-million-line Ruby migration | Compressed months of engineering into days |
| Pokémon FireRed | Completed using vision alone — prior models needed helper tools |
| Complex analytics benchmark | 90%+, a 10-point jump over Opus |
| Protein/AAV design (Mythos 5) | Outperformed specialized protein-only models |
The theme across all of these: Fable 5's edge shows up on tasks that run long and require holding a lot of context coherently, not on short single-turn prompts where the gap to Opus is much smaller.
Pricing
| Input | Output | |
|---|---|---|
| Claude Fable 5 | $10/M tokens | $50/M tokens |
| Claude Mythos Preview (prior) | ~2x this | ~2x this |
Fable 5 is priced at less than half of what the earlier Mythos Preview cost — but it's still 2x Opus 4.8's input price and 2x its output price. Prompt caching still gets the usual 90% discount on cached input tokens, which matters more here than usual given the token cost per request.
The Safety Layer
This is the part that's easy to miss if you're just comparing benchmark tables. Fable 5 ships with three classifier systems that silently fall back to Opus 4.8 when triggered:
- Cybersecurity — blocks offensive cyber tasks and agentic exploitation attempts. Anthropic reports zero universal jailbreaks across 1,000+ hours of external red-teaming.
- Biology & chemistry — deliberately overly broad at launch, prioritizing caution over false negatives.
- Distillation — prevents the model from being used to extract capabilities into a competing model.
Anthropic states over 95% of Fable 5 sessions never hit a fallback. If you're building something that brushes up against security tooling, bio/chem content, or anything that looks like model distillation, expect occasional silent downgrades to Opus 4.8 — worth knowing before you debug a mysterious quality dip that isn't a bug at all.
One more operational detail: Mythos-class traffic gets 30-day data retention, logged human access, and is excluded from training — a tighter policy than Anthropic's standard API terms.
Does It Still Matter After Opus 5?
Given Opus 5 landed three weeks later at roughly half Fable 5's price and "close to" its frontier intelligence, the honest answer is: for most day-to-day coding and writing work, Opus 5 is now the better default. Fable 5's premium is worth paying specifically when a task is long-horizon, high-context, and where the last few points of reasoning quality compound — large-scale migrations, research synthesis across huge documents, agentic workflows that run for hours unattended. For everything else, you're paying 2x for headroom you won't use.
Sources: