Claude Fable 5 Explained: Features, Limits & Setup (2026)

Claude Fable 5 is Anthropic’s most powerful AI model ever released to the public – and it runs on the same foundation as Claude Mythos, the restricted model previously available only to critical infrastructure defenders and government partners. Launched on June 9, 2026, Fable 5 brings Mythos-class intelligence to anyone with a Claude account, API key, or AWS Bedrock integration. The performance jump over Claude Opus 4.8 is not incremental. It’s the kind of leap that makes you rethink what you should be delegating to AI.

This article cuts through the naming confusion, explains what Claude Mythos and Claude Fable actually mean, benchmarks Fable 5 against Opus 4.8, and gives you a practical setup guide so you can use it before the free subscription window closes on June 22, 2026. I’ve been testing Fable 5 since launch and will share what actually surprised me – the good and the frustrating.

TL;DR: Claude Fable 5 at a Glance

• Claude Fable 5 = Claude Mythos with safety classifiers – same underlying model, public access, blocked on a narrow set of high-risk topics
• Launched June 9, 2026 – state-of-the-art across coding, vision, knowledge work, and scientific research benchmarks
• API model ID: claude-fable-5 – available on Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry
• Pricing: $10 per million input tokens, $50 per million output tokens – less than half the price of Mythos Preview
• Free on Pro, Max, Team, and Enterprise subscription plans through June 22, 2026; usage credits required after June 23
• Safety classifiers fire in fewer than 5% of sessions and fall back to Claude Opus 4.8 – not a hard error

Claude Fable 5 vs Claude Opus 4.8 vs Claude Mythos 5: Quick Comparison

Model	Best For	Biggest Strength	Main Weakness	Pricing (per 1M tokens)	Public Access
Claude Fable 5	Long-horizon agentic work, coding, finance, vision	Autonomous complex task execution, FrontierCode #1	Safety classifiers block ~5% of sessions; drops off subscriptions June 23	$10 in / $50 out	Yes – API + claude.ai
Claude Opus 4.8	General reasoning, safe production workloads	Broad capability, zero refusals from classifiers	Outclassed by Fable 5 on nearly all benchmarks	Lower than Fable 5	Yes – all plans
Claude Mythos 5	Cyberdefense, drug design, critical infrastructure	No safety classifiers – full model power	Extremely restricted; Project Glasswing only	$10 in / $50 out	No – invite only

Key Takeaways: What You Need to Know About Claude Fable 5

• Claude Fable 5 and Claude Mythos 5 share the same underlying model – Fable has safety classifiers on top; Mythos does not
• Context window: 1 million tokens. Max output: 128,000 tokens per request – both are class-leading for production use
• Fable 5 scored highest on Cognition’s FrontierCode benchmark among all frontier models, even at medium effort levels
• Refusals on sensitive topics return HTTP 200 with stop_reason: "refusal" – integrations must handle this explicitly, not as an error
• Stripe used Fable 5 to complete a 50-million-line Ruby codebase migration in one day that would have taken a full team over two months
• Knowledge cutoff: January 2026
• Fable 5 is available on Claude.ai, Claude Code CLI, Claude Code for VS Code, Claude Cowork, Amazon Bedrock, Vertex AI, and Microsoft Foundry

What Is Claude Fable 5? The Naming Explained

Claude Fable 5 is Anthropic’s public release of its Mythos-class model – the most capable AI system Anthropic has built to date. The “Fable” name signals exactly what this is: a version of Mythos made safe enough to tell to the world. Fable (from the Latin fabula, meaning “that which is told”) and Mythos share the same model weights. The only architectural difference is that Fable 5 runs a layer of safety classifiers that intercept requests touching cybersecurity exploits, dangerous biology, chemistry synthesis, and model distillation.

Anthropic first introduced Claude Mythos Preview in April 2026 through Project Glasswing – a restricted program involving AWS, Microsoft, Apple, and CrowdStrike. That model demonstrated an ability to autonomously discover and chain zero-day exploits across major operating systems, which is exactly why it stayed locked down. Fable 5 is Anthropic’s answer to the obvious question: how do you give the world access to a model this capable without handing over a cyberweapon?

Claude Fable 5 is not a watered-down Mythos. It’s the same engine with a kill switch on a narrow set of dangerous outputs. For the 95%+ of professional use cases, you’re running full Mythos-class intelligence.

Key takeaway: When people ask “Is Claude Mythos and Claude Fable the same?” – the answer is yes, architecturally. Fable 5 = Mythos 5 + safety classifiers. If you need the unconstrained version and you’re a vetted infrastructure operator, Mythos 5 is accessible through Project Glasswing. Everyone else gets Fable 5, which is still the most powerful publicly available Claude model ever shipped.

Claude Fable 5 Features: What Makes It Different

Software Engineering: The Longest Reach of Any Claude Model

Fable 5’s coding performance is where the model separation from Opus 4.8 becomes impossible to ignore. On Cognition’s FrontierCode evaluation – which tests both task completion and production-code quality standards – Fable 5 scores highest among all frontier models, even at medium effort settings. Cursor’s CEO Michael Truell confirmed it directly: Fable 5 is the new state-of-the-art on CursorBench and opens a class of long-horizon problems that were previously out of reach.

I ran a long refactoring task on Fable 5 that Opus 4.8 consistently botched on context coherence after ~30 turns – Fable 5 held the full codebase context and delivered clean output on the first pass. The token efficiency gains are real. Anthropic’s own data shows Fable 5 proactively self-updates skills based on learnings and verifies its work before delivery, behaviors that Opus 4.8 required explicit prompting to achieve. If you use Claude Code for vibe coding, Fable 5 is the model you want underneath it.

Best for: Production-grade codebase migration, long agentic coding sessions, CI/CD pipeline integration, and complex refactors across millions of lines.

Vision: Rebuilding Apps from Screenshots

Fable 5 is the new state-of-the-art for vision tasks. It can extract precise numbers from dense scientific figures and – this one surprised me – rebuild a complete web app’s source code from screenshots alone. The Pokémon FireRed benchmark is the cleanest illustration: earlier Claude models needed helper harnesses with maps and navigation aids to complete the game. Fable 5 beat FireRed using only raw game screenshots with no additional scaffolding.

For teams building document-processing pipelines, financial analysis tools, or any workflow involving chart and table interpretation, this is a material upgrade. On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 leads all models with the biggest gains in document-based reasoning and chart interpretation – exactly the tasks that break most vision models in production.

Main limitation: Vision performance shines on structured documents and screenshots. It doesn’t change the fact that image generation isn’t on-model – you still need separate tools for creative visual output.

Long-Context Memory: 1M Tokens and Actually Using Them

The 1M token context window isn’t new – Opus 4.8 already had it. What Fable 5 does differently is stay coherent and strategic across that full context. Anthropic’s Slay the Spire test is the clearest data point: giving Fable 5 access to persistent file-based memory improved its performance three times more than it did for Opus 4.8. Fable 5 also reached the game’s final act three times more often. That’s not a benchmark quirk – it’s a direct indicator of how well the model maintains long-running strategic context versus just window-stuffing tokens.

For knowledge work – legal document analysis, financial modeling, deep research – the combination of 1M context and better coherence makes Fable 5 the first Claude model I’d trust on truly large-document workflows without extensive chunking scaffolding. IMC reported that Fable 5 aced their trading analysis evaluations nearly across the board, covering factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis.

I threw a 400-page contract at Fable 5 and asked it to cross-reference conflicting clauses across exhibits. Opus 4.8 would occasionally lose the thread around clause 80. Fable 5 didn’t miss a single reference point.

How to Use Claude Fable 5: Setup Guide

On Claude.ai (Subscription Plans)

If you’re on a Pro, Max, Team, or seat-based Enterprise plan, Fable 5 is already available in your model selector at no extra cost – but only through June 22, 2026. After June 23, Anthropic pulls it from subscription plans and shifts it to usage credits. Log into claude.ai, open the model selector in any new conversation, and choose Claude Fable 5. It’s also the model now powering Claude Code’s strongest agentic results, Claude Cowork, and the Claude Code CLI. Free plan users do not have access to Fable 5.

Via the API

For API integrations, the model ID is claude-fable-5. Fable 5 is fully available on the Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry from today with no phased rollout. The critical integration change: Fable 5 can return stop_reason: "refusal" as a successful HTTP 200 response – this is new behavior compared to Opus 4.8. Your error handling must check for this explicitly. Anthropic recommends building a fallback chain to Opus 4.8 for any refused requests. Server-side fallback via the fallbacks parameter is in beta on the Claude API and AWS.

{

  "model": "claude-fable-5",

  "max_tokens": 4096,

  "fallbacks": ["claude-opus-4-8"],

  "messages": [{"role": "user", "content": "Your prompt here"}]

}

Both Fable 5 and Mythos 5 are designated Covered Models with a 30-day data retention requirement – zero data retention is not available for either. If your compliance setup requires ZDR, stick with Opus 4.8 for now. You can read more about building robust Claude pipelines in our guide on how to build Claude Managed Agents.

Why it matters: Developers who don’t handle refusals properly will see unexpected silent failures in production. This is the single biggest integration gotcha with Fable 5 – a refused request looks like a success to basic error checking.

Claude Fable 5 Pricing and Availability

Fable 5 and Mythos 5 share the same pricing: $10 per million input tokens and $50 per million output tokens. That price point is less than half what Claude Mythos Preview cost during its restricted release – a deliberate move by Anthropic to drive adoption. For context, Opus 4.8 is cheaper per token, but if your tasks benefit from Fable 5’s stronger reasoning and agentic capabilities, the ROI difference is significant. Our detailed breakdown of Opus 4.8’s pricing model and capabilities is in the Claude Opus 4.8 guide.

The subscription access timeline matters for planning: free included access on Pro/Max/Team/Enterprise runs through June 22. On June 23, it shifts to usage credits with an unclear timeline for restoration as standard. Anthropic says it will restore Fable 5 to subscription plans as capacity allows, but there’s no firm date. API and consumption-based Enterprise plans have no interruption – Fable 5 is fully available there from day one.

What Most People Misunderstand About Claude Fable 5

The biggest misconception doing the rounds is that Fable 5’s safety classifiers make it a “censored Mythos.” That framing is wrong in both directions. The classifiers only target a narrow band of genuinely dangerous outputs – cybersecurity exploit generation, bio/chem synthesis guidance, and model distillation. For the 95%+ of sessions that never touch these topics, Fable 5 runs identical to Mythos 5 with zero interference. Treating Fable as a lesser model because it has guardrails is like refusing to fly commercial because TSA exists.

The second misunderstanding is about Mythos being fundamentally different and more powerful. Mythos 5 is not a separate model with different weights or training – it’s Fable 5 with the safety classifiers removed in specific areas. The intelligence ceiling is identical. If you’ve been holding out for Mythos access thinking it’s a different beast, you’re already running essentially the same model on Fable 5. The distinction matters if and only if your work involves the specific domains those classifiers cover.

A third confusion: people think the June 22 cutoff ends access entirely. It doesn’t. It ends free-included access on subscription plans. API users and consumption Enterprise users keep full uninterrupted access. Even subscription users can still access Fable 5 after June 23 – they just need usage credits to do it.

What Actually Matters with Claude Fable 5

The real story isn’t the benchmarks. Every frontier model ships with benchmark tables. What actually matters with Fable 5 is the autonomous task duration. Previous Claude models needed frequent check-ins on complex multi-step tasks – you’d get 15-20 solid turns before coherence started degrading. Fable 5 runs for genuinely extended periods on complex knowledge and coding tasks without intervention. Stripe didn’t get months of work done in a day because the model was faster token-for-token. They got it done because Fable 5 could hold a coherent plan across an entire migration without a human steering it at every fork.

That autonomy shift is what changes the calculus for AI integration in 2026. You’re not just getting a smarter assistant – you’re getting a model that can own a deliverable end-to-end. That matters differently for different teams. For lean engineering teams, Fable 5 at $10/M input tokens is genuinely competitive with hiring contract developers for well-scoped tasks. For researchers, the combination of long-horizon autonomy and the life sciences capabilities (Mythos 5 matched or beat skilled human operators on protein design with no human assistance) signals that AI-native workflows in technical fields are no longer theoretical.

Check out how people are already rethinking AI coding workflows: our article on replacing Claude Code with Gemini CLI for 7 days shows the competitive landscape Fable 5 now has to beat. Spoiler: Fable 5 makes the comparison harder for competitors, not easier. If you want to get maximum output from Fable 5’s reasoning depth, our advanced prompt engineering guide covers the techniques that transfer directly.

Who Should NOT Use Claude Fable 5

Teams with strict ZDR compliance requirements. Fable 5 carries mandatory 30-day data retention. If your legal, healthcare, or financial workflows require zero data retention contracts with Anthropic, Fable 5 is off the table entirely. Use Opus 4.8 until Anthropic adds ZDR support for Fable-class models.

Budget-constrained teams running high-volume, low-complexity tasks. Fable 5 at $50/M output tokens is expensive for simple content generation, classification, or summarization tasks that Opus 4.8 handles competently at lower cost. If your median task is “summarize this support ticket,” you’re overpaying for Fable 5. Route those to cheaper models and reserve Fable 5 for tasks that genuinely need its depth.

Developers who haven’t updated their refusal handling. If your integration doesn’t check for stop_reason: "refusal" in successful 200 responses, Fable 5 will silently fail on a small percentage of requests with no visible error. Don’t put Fable 5 in production on existing integrations without the refusal + fallback logic in place.

Anyone whose workflow primarily touches the blocked domains. Security researchers working on offensive techniques, chemistry educators covering synthesis routes, and similar use cases will hit the classifiers regularly. For those users, Fable 5 is more frustrating than useful – the fallback to Opus 4.8 on blocked requests is functional but creates unpredictable response quality within a single session.

Real-World Recommendation: Who Gets the Most From Claude Fable 5

Use Fable 5 if you run long autonomous coding or knowledge work tasks and you’re already on the Claude API or a paid subscription. The free window through June 22 is the best reason to experiment now – run your most demanding tasks during this period to validate whether Fable 5 outperforms Opus 4.8 enough on your specific workload to justify the higher token cost after the cutoff.

Avoid Fable 5 as your default model if your task mix is broad and varied. The smart play is a routing architecture: Fable 5 for long-horizon, complex, high-value tasks; Opus 4.8 for the long tail of simpler requests. Build the fallback chain from day one – you’ll need it anyway for the refusal handling, and it will save you money at scale.

For Claude Code users specifically: Fable 5 is now the best model to run underneath Claude Code for anything beyond quick edits. The agentic coding gains are measurable and the FrontierCode benchmark lead at medium effort is meaningful – it means you burn fewer tokens reaching good output. If you haven’t already, check our full breakdown of Claude Code for vibe coding to see where Fable 5 fits in an end-to-end workflow.

My real-world call: if you’re a solo developer or small team and your monthly API spend is under $200, Fable 5 is worth routing your most complex tasks to immediately. Above that volume, instrument your task complexity distribution first – you may find Opus 4.8 handles 60% of your volume just as well at lower cost.

FAQ’s (Frequently Asked Questions)

Conclusion: Should You Use Claude Fable 5?

Claude Fable 5 is the best publicly available Anthropic model – and right now, it’s free on subscription plans for the next ten days. The performance over Opus 4.8 is real, the autonomous task capabilities are a genuine step change, and the pricing at $10/M input tokens is sharper than what Mythos Preview cost. The question isn’t whether Fable 5 is worth using – it clearly is. The question is whether your specific workflow extracts enough value to justify $50/M output tokens after June 23.

The biggest tradeoff is compliance: mandatory 30-day data retention blocks the teams that most want frontier AI capabilities (legal, finance, healthcare) from adopting Fable 5 until Anthropic adds ZDR support. That’s the gap to watch for in the next model cycle. For everyone else – use the free window to run your most demanding tasks, validate the uplift, build your refusal handling, and make an informed cost decision before the credits clock starts.

Fable 5 is where Anthropic’s frontier lands in mid-2026. If you’re following the broader AI model landscape – including what Google’s Gemini Spark agents and Gemini 3.5 Flash are doing on the other side of the aisle – Fable 5 sets a high bar. Start with the official Anthropic announcement for the full benchmark tables, then open claude.ai and run it yourself. That’s the only test that actually matters for your workflow.