Picture this: a software team at Stripe facing a migration across 50 million lines of Ruby code. Normally, that's a multi-month project requiring a whole engineering team working by hand. With Claude Fable 5, the entire migration was done in a single day. Not days. One day.
That's the kind of thing Anthropic is promising with its newest model and from what we've seen in the benchmarks and early partner reports, they're not exaggerating.
Claude Fable 5 launched on June 9, 2026, alongside its restricted sibling Claude Mythos 5. This article covers everything: what the model actually does differently, how the pricing works, the new safety classifiers (and why they'll occasionally frustrate you), who gets access to what, and whether it's worth switching from whatever you're using today.
Let's get into it.
What Is Claude Fable 5?
Claude Fable 5 is Anthropic's most capable widely released model. It belongs to what Anthropic calls the "Mythos class" — a tier of model capability they've previously kept behind locked doors due to cybersecurity concerns. Fable 5 is the first Mythos-class model made available to the general public, albeit with a set of safety classifiers that gate the most sensitive capabilities.
The model Claude Fable 5 was built for is best described as long-horizon agentic work — tasks that aren't quick questions but multi-step, multi-hour (or multi-day) autonomous operations. Think less "summarize this email" and more "analyze this codebase, identify the bottlenecks, and refactor them."
There's also Claude Mythos 5, which is the same underlying model but with fewer safety restrictions. Mythos 5 is currently restricted to a small group of approved cybersecurity organizations and infrastructure providers through a program called Project Glasswing. For most of us, Fable 5 is the one that matters.
What's Actually New
Long-Horizon Autonomy
This is the headline feature. Previous Claude models could do solid work on contained tasks but would drift, lose focus, or need hand-holding on anything that stretched across a long context window. Fable 5 sustains productive output over extended periods — we're talking hours, potentially days, of continuous goal-directed work.
When Anthropic tested Fable 5 on the deck-building game Slay the Spire, giving it access to persistent file-based memory improved its performance three times more than for Opus 4.8. Fable 5 also reached the game's final act three times more often. That sounds like a novelty, but the underlying skill using persistent memory to build on prior experience rather than starting fresh each session — is exactly what matters for real autonomous workflows.
Vision Capabilities
Fable 5 is Anthropic's new state-of-the-art vision model. It can extract precise numbers from dense scientific figures, rebuild web app source code from screenshots alone, and handle noisy, flipped, or low-resolution images using bash and crop tools it calls on its own.
The most striking demonstration: earlier Claude models couldn't play Pokémon FireRed without an elaborate helper harness giving them extra tools and game-state information. Fable 5 beat the game using only raw game screenshots no maps, no navigation aids, no extra context. One continuous timelapse from start to finish, vision only.
For enterprise use cases, this translates into things like processing scanned documents, interpreting complex charts without special tooling, and working with visual data that previous models would bounce back as "unclear."
Knowledge Work and Reasoning
IMC, the trading and market-making firm, ran their trading analysis evaluations on Fable 5 and reported it "aced" them nearly across the board factual lookup, conceptual reasoning, root-cause analysis, expected-value analysis. That's not a generic benchmark; that's a firm stress-testing the model against their actual workflows.
On Hebbia's Finance Benchmark for senior-level reasoning, Fable 5 scored highest of any current model, with notable gains in document-based reasoning, chart interpretation, and complex problem solving. If your work involves dense documents, financial models, or multi-step analytical chains, this is where Fable 5 earns its price tag.
Software Engineering
Beyond the Stripe example up top: on Cognition's FrontierCode evaluation which tests whether models can pass difficult coding tasks while meeting production-quality standards Fable 5 scores highest among frontier models. The emphasis on production quality matters. It's easy to pass a coding test with working code that no real engineering team would accept. FrontierCode specifically penalizes that.
Code review, debugging, and codebase-wide search also improved meaningfully. Fable 5 can scan repository history, trace root causes across large codebases, and catch bugs that Opus 4.8 would miss. The one significant caveat: the safety classifiers block offensive cybersecurity work, so if your use case involves penetration testing, exploit development, or security tooling, expect fallbacks to Opus 4.8.
Drug Design and Life Sciences
This is where Mythos 5 diverges sharply from Fable 5, so worth noting separately. Anthropic's internal protein design team reported a roughly ten-times acceleration on aspects of the drug design process. In one test, Mythos 5 — given protein design tools but no human assistance matched or beat skilled human operators at selecting binding sites, running protein design tools, and recovering from failures.
Nine of 14 protein targets yielded strong drug design candidates currently under investigation. This is not a demo. It's active research with real outputs.
For Fable 5 users, biology and chemistry queries will fall back to Opus 4.8 by default. The full life sciences capability is Mythos territory for now, though Anthropic is opening a trusted access program for biology researchers in the coming weeks.
The Safety Classifiers: What They Are and How They'll Affect You
This is genuinely new territory for a public model release, and it's worth understanding properly rather than dismissing as boilerplate safety theater.
Why They Exist
Mythos-class models have real, quantifiable capability in offensive cybersecurity and biological research that prior models simply didn't have. Anthropic ran tests showing Fable 5's performance on predicting unpublished viral properties outperformed specialized protein language models — using only biological reasoning, without being explicitly trained on the task.
That's useful for gene therapy research. It's dangerous if the same capability is applied to designing harmful viruses. Same queries, very different intent.
Similarly, Fable 5's ability to perform agentic hacking — not just finding exploits but handling reconnaissance, lateral movement, and the broader attack chain is something that could realistically accelerate cyberattacks in the wrong hands. The classifiers exist because the uplift is real.
What Gets Blocked (and What Happens Instead)
Three categories trigger classifier fallback to Opus 4.8:
Cybersecurity — exploitation techniques, offensive cyber tasks, malware, attack tooling. One partner tested Fable 5 across 30 different public jailbreak techniques; zero harmful requests got through. Benign cybersecurity work can sometimes trigger this too. If you're a developer doing legitimate security work, you may see unexpected fallbacks.
Biology and chemistry — Anthropic has deliberately set this to broad and conservative, covering most requests in the domain until they can narrow it with more confidence. They've been explicit that this will produce false positives. Biomedical researchers are being offered a trusted access program to Mythos 5 with these restrictions lifted.
Distillation — attempts to systematically extract Claude's capabilities to train competing models. Flagged requests fall back to Opus 4.8.
When fallback happens, you're informed. You're not left hanging with an opaque refusal — you get an Opus 4.8 response. More than 95% of Fable 5 sessions involve no fallback at all.
Robustness Against Jailbreaks
Anthropic ran an external bug bounty — over 1,000 hours of testing — and found no universal jailbreaks. External red teams also failed on long-form agentic tasks, though the UK AISI made some progress toward one in an initial testing window, which Anthropic disclosed. They're not claiming perfection; they're claiming the safeguards are slow and costly enough to prevent scaled misuse, which is the realistic goal.
An internal graph shows Fable 5 with blocking cyber safeguards stopping all progress on offensive tasks, compared to prior models that showed incremental progress. That's a meaningful difference, even if it's not absolute.
Pricing
Here's the full pricing breakdown for both models:
Both Claude Fable 5 and Claude Mythos 5 are priced identically. Base input tokens cost $10 per million tokens, and output tokens cost $50 per million tokens those are the two numbers that matter most for most workloads.
Prompt caching adds a few more tiers worth knowing. Writing to cache with a short 5-minute window costs $12.50 per million tokens. The longer 1-hour cache write costs $20 per million tokens more expensive upfront, but it pays off if your context stays stable over many calls. Cache hits and refreshes drop all the way to $1 per million tokens, which is where long-running agentic workflows start to get genuinely economical. If you're building a pipeline where the same large system prompt or context gets reused across hundreds of calls, that $1 rate makes a real difference to your monthly bill.
One important note Anthropic is leading with: this is less than half the price of Claude Mythos Preview, which Fable 5's performance matches or exceeds on most tasks. If you were using Mythos Preview, you're getting a better model for a lower price.
The output token cost of $50/MTok is the number to watch on long-horizon agentic tasks. Fable 5 at xhigh or max effort settings can generate substantial output Anthropic recommends starting with a max_tokens setting of at least 64k on agentic runs, adjusting from there. Budget accordingly.
For comparison, cache hits at $1/MTok mean that long-running tasks with stable system prompts or large repeated contexts become significantly more economical. If you're building an agentic workflow where the same 100k-token context gets reused across many calls, prompt caching will materially reduce your costs.
The model string for API access is claude-fable-5.
API and Technical Details
A few things developers need to know before migrating:
Adaptive thinking is always on. Fable 5 doesn't let you disable thinking. The thinking: {"type": "disabled"} parameter is not supported and returns an error. You control thinking depth through the effort parameter instead.
Raw chain-of-thought is never returned. You can get a summarized version of the reasoning by setting thinking.display to "summarized", or omit it entirely (the default). This is a deliberate change — attempts to extract raw reasoning are flagged as potential distillation.
The refusal stop_reason. When classifiers trigger, the API returns stop_reason: "refusal" as a normal HTTP 200, not an error. Your error handling needs to account for this explicitly. Anthropic provides three fallback approaches: server-side (pass a fallbacks parameter), client-side (SDK middleware), or manual.
Context window and output: 1M token context window by default, up to 128k output tokens per request.
Effort levels for Fable 5, from low to high capability: low, medium, high (default), xhigh, max. Lower effort settings on Fable 5 still often exceed xhigh performance on prior models — the floor is higher. Use high as your starting point and adjust from there.
Data retention: Fable 5 and Mythos 5 require 30-day data retention and are not available under zero data retention agreements. This covers all traffic, first- and third-party. Anthropic is explicit that this data won't be used for model training — it's for detecting novel jailbreaks and identifying false positives in the classifiers.
Prompting for Best Results
Fable 5 behaves differently enough from Opus 4.8 that old prompts sometimes degrade output quality. The model is more capable and follows instructions more precisely, which means over-specified instructions can get in the way.
A few things that actually matter:
Give it harder tasks. Teams that test Fable 5 on simpler workloads consistently undersell what it can do. Pick a task that was previously too complex or too long-running, and start there.
Ground progress claims during long runs. On autonomous pipelines, instruct the model to audit each progress claim against an actual tool result before reporting. In Anthropic's testing, this instruction nearly eliminated fabricated status updates on tasks designed to elicit them. Something like: "Before reporting progress, audit each claim against a tool result from this session."
Avoid instructions to reproduce reasoning. If your existing prompts tell the model to "show your thinking" or "explain your reasoning step by step," those instructions can trigger the distillation classifier and cause unexpected fallbacks to Opus 4.8. Audit your system prompts for this.
Use explicit boundaries. Fable 5 can take unrequested actions — drafting documents you didn't ask for, creating defensive git branches, going beyond scope. Define what it should and shouldn't do explicitly, especially for long autonomous runs.
Memory files matter. For long sessions, give the model a place to write notes (even just a Markdown file). Fable 5 performs noticeably better when it can record and reference lessons from earlier in a task.
Availability
Claude Fable 5 is generally available as of June 9, 2026 on:
- Claude API
- Claude Platform on AWS
- Amazon Bedrock
- Vertex AI
- Microsoft Foundry
Claude Mythos 5 is available in limited release only, to approved Project Glasswing partners.
For subscription plans (Pro, Max, Team, seat-based Enterprise): Fable 5 is included at no extra cost through June 22, 2026. On June 23, usage will require usage credits. Anthropic's stated plan is to restore Fable 5 as a standard part of subscription plans once they have sufficient capacity, with advance notice before any changes.
Who Should Actually Use This
Here's the honest breakdown:
Use Fable 5 if: You're running long autonomous workflows. You're doing complex code migrations, refactoring, or multi-step engineering work. You need serious vision capabilities dense charts, screenshots, technical figures. You're doing senior-level financial analysis or research. The longer and more complex the task, the more the performance gap over Opus 4.8 matters.
Stick with Opus 4.8 if: Your tasks are well-contained and conversational. You're doing legitimate cybersecurity work that keeps hitting the classifiers. You need zero data retention. You're working in biology/chemistry and don't want constant fallbacks. For these cases, Opus 4.8 is genuinely excellent and doesn't carry the data retention requirement.
Wait for Mythos 5 access if: You're in life sciences research and need the full capabilities without the biology classifier. Anthropic is opening a trusted access program specifically for biomedical researchers — you'll get Fable 5 capabilities with the biology/chemistry restrictions lifted.
The Bottom Line
Claude Fable 5 is a legitimate step change, not a marketing bump. The Stripe codebase migration, the drug design results, the Pokémon finish with vision-only — these are concrete demonstrations from real partners with real stakes in whether the model actually works.
The safety classifiers will occasionally get in your way if your work touches cybersecurity or biology. That's a real tradeoff, and Anthropic isn't hiding it. The 5% fallback rate they're citing is for average users; if your workflow regularly involves those domains, your rate will be higher.
The pricing at less than half of Mythos Preview — makes the calculus pretty clear for anyone who was already using frontier Claude models. And for developers new to this capability tier, the 1M context window, the 128k output per request, and the serious autonomous task performance make this worth testing on your hardest problems first, not your easy ones.
Sources
- Anthropic — Claude Fable 5 and Claude Mythos 5 Launch Announcement (June 9, 2026) Official release post covering capabilities, safety classifiers, Project Glasswing expansion, and benchmark results. https://www.anthropic.com/news/claude-fable-5-and-claude-mythos-5
- Anthropic Docs — Introducing Claude Fable 5 and Claude Mythos 5 Full API documentation: model IDs, context window specs, pricing, availability, refusal handling, fallback options, and billing rules. https://docs.anthropic.com/en/docs/about-claude/models/claude-fable-5
- Anthropic Docs — Prompting Claude Fable 5 Developer guide covering behavioral differences from Opus 4.8, effort levels, long-run scaffolding, memory systems, subagent delegation, and classifier-aware prompt patterns. https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/claude-fable-5
- Anthropic Docs — Effort Parameter Full reference for the effort parameter across all supported models (Fable 5, Mythos 5, Opus 4.8, Sonnet 4.6, and others), including recommended levels by model and use case. https://docs.anthropic.com/en/docs/build-with-claude/extended-thinking/effort
- Anthropic — Project Glasswing Overview Background on the trusted access program for Mythos-class cybersecurity deployments and the US government collaboration underpinning Mythos 5's restricted rollout. https://www.anthropic.com/glasswing
