Model Launch

Claude Fable 5 Is the Best Model Anthropic Will Let You Use. For Now.

June 9, 2026 · 6 min read

Anthropic just launched Claude Fable 5 and Claude Mythos 5 — the first Mythos-class models to reach the public API. They're the same model underneath. The difference isn't intelligence. It's guardrails. Fable 5 is the one you can actually use today, with classifiers that watch for dangerous territory and fall back to Opus 4.8 when they trip. For 95% of sessions, you get Mythos-level capability at $10/$50 per million tokens. Here's what that means for the code you're building.

One model, two names, one big tradeoff

Anthropic's model hierarchy now looks like this: Haiku → Sonnet → Opus → Mythos. Fable 5 and Mythos 5 sit at the top as the same underlying Mythos-class model. Fable 5 is the version with safeguards turned on — classifiers that detect queries related to cybersecurity, biology/chemistry, and AI distillation, and automatically switch the response over to Claude Opus 4.8. Mythos 5 is the same brain with those classifiers lifted, available only to Glasswing consortium partners and select biology researchers.

I've been using Fable 5 since it dropped, and the honest reality is this: for the vast majority of developer work — coding, debugging, architecture, code review — you will never notice the classifiers. Anthropic says they trigger in less than 5% of sessions. When they do, you get Opus 4.8 instead of a refusal, which is a genuinely thoughtful design choice. A fallback to Opus 4.8 on a flagged cyber query is infinitely better than the model going silent.

The safeguard design deserves attention because it's the reason this model exists publicly at all. Mythos Preview, released in April, triggered Anthropic's ASL-4 safety protocol during evaluation. It was restricted to a handpicked consortium of cyber defenders and infrastructure providers. The version now in your API key — Fable 5 — exists because Anthropic built classifiers robust enough to satisfy their internal safety bar. Over 1,000 hours of external red-teaming produced zero universal jailbreaks. The UK AISI got close within a short testing window but didn't crack it. That's the engineering that bought public access.

What Fable 5 actually delivers

The headline numbers are strong, but the developer-relevant ones are stronger:

Long-horizon coding. Stripe reported that Fable 5 compressed months of engineering into days. In a 50-million-line Ruby codebase, it performed a codebase-wide migration in a day that would have taken a team over two months manually. On Cognition's FrontierCode evaluation — which tests whether models can meet production codebase standards, not just pass toy problems — Fable 5 scores highest among all frontier models, even at medium effort.

Vision that actually works. Fable 5 is the new state-of-the-art for vision tasks. It can extract precise numbers from scientific figures and rebuild a web app's source code from screenshots alone. The most telling demo: earlier Claude models needed a complex helper harness to play Pokémon FireRed. Fable 5 completed the game with vision alone — raw screenshots, no maps, no game-state information. That's the difference between "can see images" and "can understand what it's looking at well enough to act on it."

Memory that compounds. Fable 5 stays focused across millions of tokens and uses its own notes to improve over time. When given access to persistent file-based memory while playing Slay the Spire, it improved three times more than Opus 4.8 did with the same setup, and reached the game's final act three times more often. For developers shipping long-running agent workflows, this is the metric that matters.

Scientific reasoning at a new tier. Mythos 5 (the unrestricted version) accelerated drug design by roughly 10x — matching or beating skilled human operators at choosing binding sites, running protein design tools, and recovering from failures. In blinded comparisons, scientists preferred Mythos 5's molecular biology hypotheses over Opus-class models ~80% of the time. One hypothesis was independently corroborated by an external lab working on the same problem. It also conducted a week of autonomous genomics research, training a custom ML model that outperformed a recent Science paper — despite being 100 times smaller.

Developer quickstart

Pricing context: Opus 4.8 runs $5/$25 per million tokens. Fable 5 at $10/$50 is double the cost — but less than half of what Mythos Preview cost. Whether that premium is worth it depends entirely on your workload. For routine chat and straightforward code generation, stick with Opus 4.8. For long-horizon autonomous coding, complex debugging across a large codebase, or vision-heavy tasks, Fable 5's premium pays for itself in fewer turns and fewer failures.

The June 23 credit cutoff

Anthropic is handling the launch with an unusual pricing transition:

• June 9–22: Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost.
• June 23: Fable 5 will be removed from those plans and require usage credits. If capacity allows, this window will be extended.
• Eventually: Anthropic aims to restore Fable 5 as a standard part of subscription plans once capacity permits.

API users are unaffected — Fable 5 is fully available on the API and consumption-based Enterprise plans from day one. The credit game is for subscription-plan users only. If you're building on the API, you can ignore the calendar and just start using claude-fable-5 now.

When to reach for Fable 5 vs. Opus 4.8

This is the practical question, and after a few days of using both, here's my framework:

Use Fable 5 when:

• You're running autonomous agent loops that span hundreds of tool calls
• Your task involves interpreting images or screenshots as part of a coding workflow
• You need the model to maintain state and reasoning across a long session without degrading
• You're doing complex debugging where the model needs to form and test hypotheses
• You're working across a massive codebase and need reliable long-context retrieval

Use Opus 4.8 when:

• You're doing single-turn code generation or quick Q&A
• Cost is the primary constraint and your tasks are well-scoped
• Your prompt involves cybersecurity, biology, or chemistry topics (Fable will just fall back to Opus anyway)
• You're running high-volume batch processing where the 2x price difference compounds

The honest answer for most developers: start with Opus 4.8 for everything, then upgrade specific high-value or high-failure-rate workloads to Fable 5. The price gap means indiscriminate use is wasteful. But for the tasks where Opus 4.8 falls short, Fable 5's step change in reliability is real — and it's cheaper and more useful than Opus 4.8 failing three times before succeeding once.

What Mythos 5 unlocks (that you can't have yet)

Mythos 5 is the same model with the cyber safeguards lifted, deployed through Project Glasswing in collaboration with the US government. It has the strongest cybersecurity capabilities of any model in the world. Anthropic is planning a broader trusted access program — cybersecurity organizations will be able to apply systematically, and a separate biology track will give researchers access to Fable 5 with the biology/chemistry classifiers removed.

The long-term trajectory is clear: Anthropic wants Mythos-5-level access to expand. The classifiers on Fable 5 are deliberately conservative to get the model out the door, and they've said explicitly they plan to narrow them as they gather data and refine the safeguards. If you're an infrastructure provider or cyber defender, reach out to your Anthropic account team about Glasswing access. For everyone else, Fable 5 is the best you can get — and for now, it's genuinely very good.

The bet Anthropic is making

This launch is an experiment in a question the industry hasn't answered yet: can you ship a model powerful enough to cause real harm if you wrap it in strong enough safeguards? Anthropic's answer is Fable 5 — same brain as the model that triggered their highest safety alert, but with classifiers that have survived 1,000+ hours of professional attacks.

The classifiers will frustrate you. They're tuned conservatively and will catch harmless queries. Some developers will hit the fallback to Opus 4.8 on something innocent and wonder why they're paying $10/$50 for a model that won't answer. But the alternative — the one every other lab has chosen so far — is not shipping the model at all. Anthropic shipped it.

The 30-day data retention requirement is the other cost of this approach. All Fable 5 and Mythos 5 traffic is retained for 30 days for safety monitoring (not training). Anthropic has added privacy protections including logging all human access to the data and ensuring deletion after 30 days. If your use case can't tolerate that, Fable 5 isn't for you. For most developer workloads, it's a reasonable price for access to a model that would otherwise be locked behind government partnerships.

Fable 5 won't answer every question you throw at it. But it'll answer 95% of them at a level no other publicly available model can match. That's the trade. I'll take it.