Claude Fable 5: Anthropic makes its Mythos model accessible to the public

LLM & Modèles 🟢 Beginner ⏱️ 13 min read 📅 2026-06-10

Claude Fable 5: Anthropic makes its Mythos model available to the public

🔎 A model too dangerous becomes paid

On June 9, 2026, Anthropic released Claude Fable 5. This is not a simple update. It is the first public version of Mythos, the family of models that Anthropic had kept under lock and key since April 2026 because it deemed them too powerful for general public release.

Fable 5 is Mythos with a safety brake. Scores explode on benchmarks — up to 10% above Claude Opus 4.8. But the pricing does too: $10 per million input tokens, $50 for output. Double that of Opus 4.8.

The timing is anything but coincidental. Anthropic is preparing its IPO. The Pentagon just cut its military contracts. And Claude Mythos just discovered over 10,000 critical vulnerabilities in a month via the Glasswing project — which raises the question of what happens when this power is in everyone's hands.

The key points

Fable 5 = secured Mythos: this is the first model in the Mythos lineage made public, with safeguards that the internal model did not have.
Performance +10% over Opus 4.8 on certain reasoning and code benchmarks, according to Anthropic's data.
Premium pricing: $10/$50 per million tokens (June 2026, check on anthropic.com), which is double that of Opus 4.8.
Automatic routing: requests deemed "high-risk" are switched to Opus 4.8 without user intervention.
Strategic context: released right in the middle of IPO preparation, after the termination of contracts with the Pentagon.

Recommended tools

What Fable 5 really is — and what it is not

Fable 5 is not a new model from scratch. It is a throttled version of Mythos, the model family that Anthropic classified as "high danger" in April 2026.

The difference between the internal Mythos and the public Fable 5 comes down to two words: routing and sandboxing. When you send a prompt to Fable 5, a first classification layer assesses the risk level. If the request touches on sensitive domains (synthetic biology, cyberattacks, advanced social engineering), it is redirected to Claude Opus 4.8, which has more proven safeguards.

If the request is deemed safe, it lands on the Mythos engine with its superior reasoning capabilities. According to Anthropic's announcement, this routing architecture is what allowed the company to say yes to the release while maintaining control.

It is not a "secure" model in the sense that everyone means. It is a model whose inputs are filtered before reaching the dangerous engine. The distinction is important.

The scores: +10% on Opus 4.8, but with caveats

Anthropic's numbers are impressive on paper. On reasoning benchmarks (GPQA, MATH-500, ARC-AGI), Fable 5 shows scores 8 to 12% higher than Claude Opus 4.8.

In code, the gap narrows but remains significant. On SWE-bench Verified, Fable 5 reaches levels that place it above all current public models, including GPT-5.5 (98.2 in agentic) and Gemini 3.1 Pro (overall score of 92).

But there is a major methodological issue. These benchmarks are published by Anthropic itself. No independent third party has yet been able to reproduce these results as of June 9, 2026. The Guardian, in son article sur la sortie, highlights this lack of external verification as a point of tension with the research community.

The comparison context is also biased. Anthropic compares Fable 5 to Opus 4.8, not to the most recent competing models. If we look at the comparatif des LLM 2026, Gemini 3.1 Pro dominates the overall ranking with 92 points. Fable 5 has not been integrated into these third-party rankings.

Pricing: a luxury rate that excludes independents

$10 per million input tokens. $50 for output. Twice as much as Opus 4.8.

To put this into perspective: a complex code analysis that costs $2 with Claude Sonnet 4.6 will cost around $15-20 with Fable 5. A batch of 1,000 customer requests? You go from $50 to $300 in a single weekend.

This is no accident. Anthropic is explicitly targeting Fortune 500 companies with this model. The pricing acts as a natural filter: only organizations with substantial AI budgets can afford Fable 5 in production.

For independent developers and small teams, the calculation is simple. Claude Opus 4.7 Adaptive (overall score of 90, 94.3 in agentic) offers 90% of the capabilities at 50% of the price. Alternatively, you can turn to the meilleurs LLM gratuits for everyday tasks and save the budget for cases where Fable 5 is truly irreplaceable.

TechCrunch reports that cette stratégie de prix premium is directly linked to IPO preparation. Anthropic must demonstrate that it can monetize its most powerful models at high margins. Fable 5 is an argument for investors as much as it is a product.

Automatic routing: a problematic black box

The most controversial mechanism of Fable 5 is its automatic routing to Opus 4.8.

In practice, you never know which engine is processing your request. Anthropic states in its documentation that a response header allows you to check if routing has occurred, but in practice, most API integrations do not expose it by default.

This creates an absurd situation: you pay the Fable 5 rate ($10/$50), but some of your requests are processed by Opus 4.8 ($5/$25). Anthropic does not offer a differential refund. You pay the premium price regardless of the route taken.

The study AI Governance and Accountability: An Analysis of Anthropic's Claude, published on arXiv, precisely analyzes this type of architecture and points out a risk of blurred responsibility. When the model routes to Opus 4.8 without clearly stating it, who is responsible for the quality of the output? The user who paid for Fable 5, or Anthropic who chose to downgrade the request?

For companies under strict regulation (banking, healthcare), this is a potential dealbreaker. You cannot certify that a specific model produced a given output if the system dynamically routes between two different engines.

Glasswing and security: the Mythos paradox

It is impossible to talk about Fable 5 without mentioning the Glasswing project. In May 2026, Anthropic used Claude Mythos to scan the code of critical infrastructures. Result: more than 10,000 security vulnerabilities discovered in a month.

This result is spectacular from a capability standpoint. But it is terrifying from a security standpoint. If Mythos can find 10,000 vulnerabilities, it can probably exploit them. This is exactly the reason why Anthropic classified Mythos as "high danger".

Fable 5 uses the same reasoning engine, with an upstream filter. The filter blocks explicit requests like "find the vulnerabilities in this system". But what happens with an indirect request? "Analyze the architecture of this payment system to explain to me how it works" — a legitimate request that could serve as a vector.

Anthropic has not published detailed data on the bypass rate of the routing filter. The Guardian points out that the company refused to share the results of its internal red teaming on Fable 5 specifically.

This silence is all the more striking given that Anthropic just refused China access to the Mythos model, justifying this decision by citing national security risks. If the model is too dangerous for a nation-state, why is it safe enough for the public at $10 per million tokens?

Anthropic, the Pentagon and the IPO: the triple pressure

To understand Fable 5, you have to look at Anthropic's strategic context in June 2026.

The Pentagon cut its contracts. After months of negotiations, the Department of Defense ended its agreements with Anthropic, officially for budgetary reasons, unofficially because of political pressure surrounding the deployment of American models in military contexts. This loss of revenue represents a direct financial pressure.

The IPO is approaching. Anthropic needs to demonstrate two things to investors: high margins (hence Fable 5's premium pricing) and an ability to navigate regulation (hence the "responsible" routing architecture). Fable 5 is a flagship product for roadshows.

The AI cold war is intensifying. By blocking China's access to Mythos while releasing it for the Western market, Anthropic positions its models as strategic infrastructure. This is no longer tech. It's geopolitics.

The TechCrunch article on the release of Fable 5 describes this maneuver as a "balancing act between monetization and responsibility" — a journalistic euphemism for saying that Anthropic is taking a calculated risk.

Fable 5 facing the competition: who wins?

The LLM landscape of June 2025-2026 is saturated. Where does Fable 5 stand?

Against GPT-5.5 (OpenAI): GPT-5.5 dominates the agentic leaderboard at 98.2 points. Fable 5 has no public third-party agentic score. In terms of pure reasoning, Fable 5 seems to have the edge, but without an independent benchmark, this is hard to confirm. For a detailed comparison, see our Claude vs ChatGPT.

Against Gemini 3.1 Pro (Google): Gemini leads the overall leaderboard at 92 points. Its price is significantly lower thanks to the Google Cloud ecosystem. For most businesses, the price-to-performance ratio leans in favor of Google.

Against DeepSeek V4 Pro (Max): At 88 points and a noticeably lower cost, DeepSeek remains the pragmatic choice. Fable 5 must justify its premium price through capabilities that DeepSeek cannot match — which it does on highly specialized reasoning, but not on everyday tasks.

Against local models: For organizations that refuse to send their data to the cloud, Fable 5 is not an option. The best LLMs to run locally via Ollama or LM Studio remain the only alternative for on-prem. Our local LLM installation guide details the possible setups.

The truth? Fable 5 is a niche tool. It shines on ultra-complex reasoning tasks where budget is not a factor. For everything else, there are more rational alternatives.

AI agents: Does Fable 5 change the game?

The June 2025 agentic ranking places GPT-5.5 at the top with 98.2, followed by Gemini 3 Pro Deep Think at 95.4 and Claude Opus 4.7 Adaptive at 94.3. Fable 5 does not appear in these rankings — they date from before its release.

But Anthropic clearly positions Fable 5 as an engine for agents. Superior reasoning, long context management, and the ability to plan multi-step action sequences are all key criteria for the best LLMs for AI agents.

The problem is automatic routing. An agent executing a 20-step task will see some of its sub-queries routed to Opus 4.8. The consistency of the reasoning chain is potentially weakened as a result. You switch from one engine to another without control, which is exactly what an agentic system should never do.

For serious agent architectures, the current consensus remains on GPT-5.5 or Gemini 3 Pro Deep Think, precisely because they offer a single, predictable engine.

Open-source infrastructure as a counterpower

While Anthropic locks Mythos behind a premium paywall, a parallel movement is growing. The TrustGraph project, presented on Hacker News in June 2026, offers open-source AI infrastructure for building data and reasoning pipelines without relying on a single vendor.

The approach is radically different. Instead of paying $50 per million tokens for a black-box model, you deploy your own reasoning graphs on open-source models. Raw performance is lower than Fable 5. But the transparency, control, and cost are incomparably better.

This is the fundamental tension of 2026: ultra-powerful yet opaque and expensive proprietary models, versus less performant but auditable and affordable open-source models. Fable 5 is the most extreme embodiment of the first pole.

❌ Common mistakes

Mistake 1: Thinking Fable 5 is "safe"

What's wrong: the word "safe" suggests that the model itself was modified to be less dangerous. In reality, only upstream routing was added. The underlying Mythos engine is intact.

The solution: treat Fable 5 as Mythos with an input filter. Not as a different model. Apply the same precautions you would apply with a model rated "high danger".

Mistake 2: Using Fable 5 for all your tasks

What's wrong: paying $10/$50 per million tokens for requests that Claude Sonnet 4.6 (free or $3/$15) handles just as well. Most daily tasks don't require Mythos reasoning.

The solution: reserve Fable 5 for use cases where the performance gap is measurable and justifies the extra cost. For the rest, use the best LLMs for coding at reasonable rates.

Mistake 3: Ignoring automatic routing in your integration tests

What's wrong: benchmarking Fable 5 without checking which engine actually processed each request. Your scores are skewed if 30% of the requests were routed to Opus 4.8.

The solution: parse the API response headers to identify routing. Anthropic exposes them — you just have to read them.

Mistake 4: Comparing Fable 5 to existing third-party benchmarks

What's wrong: the June 2025 LLM rankings (Gemini 3.1 Pro at 92, GPT-5.5 at 91) do not include Fable 5. Comparing them directly is misleading.

The solution: wait for independent benchmarks. Anthropic's scores are a starting point, not a conclusion.

❓ Frequently Asked Questions

Is Fable 5 available in France?

Yes, via the Anthropic API. There are no geographic restrictions for Western countries. However, China is explicitly excluded from accessing the Mythos family, as we detailed in our article on Anthropic's refusal to grant China access to the Mythos model.

Can automatic routing be disabled?

No. Anthropic does not offer an option to force processing by the Mythos engine on all requests. Routing is mandatory and non-configurable. This is a deliberate security choice, but it is also a major functional limitation.

Is Fable 5 worth double the price of Opus 4.8?

For highly specialized reasoning tasks (advanced mathematics, formal proofs, complex system architecture), the performance gap may justify it. For everything else, no. Opus 4.7 Adaptive or even Sonnet 4.6 offer a much better price-to-performance ratio.

Will Anthropic release other Fable versions?

Probably. Fable 5 is a market test. If enterprise adoption is strong and no major security incidents occur, Anthropic will likely release Fable 6 with finer routing and potentially adjusted pricing. The cadence will also depend on the IPO timeline.

Are there French alternatives at this level of performance?

Not exactly. The best French LLMs still lag behind Claude and GPT on reasoning benchmarks. For French content, Fable 5 remains performant, but the premium price is hard to justify compared to optimized Francophone models.

✅ Conclusion

Claude Fable 5 is the most powerful model Anthropic has ever made public, but it is also the most paradoxical: you are being sold a "secure" Mythos engine with a filter you do not control, at a price that excludes 95% of developers, all while Anthropic prepares its IPO and manages its breakup with the Pentagon. For use cases where reasoning makes the difference between an acceptable result and a breakthrough result, Fable 5 is worth testing. For everything else, check out our monthly comparison of the best LLMs before signing the check.

#intelligence-artificielle #anthropic #modele-mythos #claude-fable-5 #claude ai

📚 Related articles

LLM & Modèles 🟢 Débutant 12 min

Best Free Llms (June 2026)

Discover the ranking of the best free LLMs in June 2026. Market analysis and comparison of uncensored AI models.

2026-06-09 04:01