Anthropic just dropped what may be the most significant AI model release of 2026. On June 9, 2026, the company unveiled Claude Fable 5 and Claude Mythos 5 — two powerhouse models built on the same underlying architecture, each aimed at a different audience and use case.
This is not just another incremental update. Fable 5 is the first time Anthropic has made its elite Mythos-class intelligence available to the general public, and the details around pricing, safety, access, and benchmark performance are turning heads across the AI world.
Here’s the full breakdown.
What Exactly Are Claude Fable 5 and Mythos 5?
Think of it this way: Anthropic built one incredibly powerful model, then shipped it in two different configurations depending on who gets to use it.
Claude Fable 5 is the publicly available version. It carries Mythos-level intelligence wrapped in a layer of safety classifiers that automatically block responses in high-risk domains. Anthropic describes it as the most capable model the company has ever made generally available — outperforming everything in their lineup, and according to their benchmarks, outperforming competitors too.
Claude Mythos 5 is the unrestricted configuration of that same model — no safety classifiers, full capability. It is not available to the public. Access is limited exclusively to organizations that have been vetted and approved through Anthropic’s Project Glasswing initiative, which focuses on trusted partners in critical infrastructure, cybersecurity, and life sciences.
The two models share an API, but they serve very different audiences.
How Did We Get Here? The Backstory on Mythos
Anthropic first introduced the Mythos model family back in April 2026, and the reaction was electric — and cautious. The model’s ability to identify software vulnerabilities, assist in biodefense screening, and handle highly sensitive scientific tasks caught the attention of both Wall Street and government officials. But Anthropic didn’t release it broadly. They were very deliberate: Mythos was made available only to a handful of partners through Project Glasswing, specifically because of its dual-use potential in cybersecurity and biology.
Now, two months later, Anthropic says it has found a way to bring that same level of intelligence to everyone — by building robust safety guardrails into the model itself. The result is Claude Fable 5.
What Makes Fable 5 So Powerful?
According to Anthropic, Claude Fable 5 sets new records on nearly every AI benchmark the company tested. The performance gap over previous models becomes even more pronounced the longer and more complex the task — meaning this model is specifically built to shine on hard, sustained work.
Standout performance areas include:
- Software engineering — Fable 5 scores 80.3% on SWE-Bench Pro, compared to 69.2% for Claude Opus 4.8 and 58.6% for GPT-5.5. On the harder FrontierCode Diamond benchmark, it reaches 29.3% versus just 13.4% for Opus 4.8.
- Knowledge work — Consistently ahead of Opus 4.8 and competing models on complex professional tasks.
- Vision — Strong multimodal performance across image understanding tasks.
- Scientific research — Fable 5 has already been used in life sciences contexts, including work on adeno-associated virus (AAV) candidates developed by Dyno Therapeutics.
- Long-horizon agentic tasks — Both Fable 5 and Mythos 5 are described as capable of working autonomously for longer periods than any previous Claude model.
Early testers have noted that Fable 5 simply feels different — broader, deeper, more knowledgeable. Developer Simon Willison described it as feeling “big,” speculating it may be the largest model Anthropic has ever shipped in terms of sheer scale.
The Safety Layer: How Fable 5 Handles Risky Requests
This is where things get technically interesting. Claude Fable 5 is not a censored model in the traditional sense. Instead, it uses built-in safety classifiers that monitor requests in real time. When a request touches high-risk territory — cybersecurity exploits, biology, chemistry, or distillation, for example — the classifier steps in and declines to respond.
When that happens, the API returns a stop_reason: "refusal" signal as a standard HTTP 200 response. Crucially, refused requests are not billed if no output was generated. Developers can also configure automatic fallback behavior, routing refused requests to Claude Opus 4.8 automatically, either server-side or client-side through SDK middleware.
This design means Fable 5 is both more powerful and more commercially sensible to integrate — you don’t pay for refusals, and you don’t have to manually handle fallback logic if you don’t want to.
Pricing: Expensive but Cheaper Than Mythos Preview
Both Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens. That’s double the price of Claude Opus 4.8 — but less than half the cost of Claude Mythos Preview, the earlier restricted model.
There are no surcharges for longer context usage. Both models support a 1 million token context window by default, with a maximum output of 128,000 tokens per request. The knowledge cutoff date for both models is January 2026.
Who Can Access Fable 5 Right Now?
Availability is being rolled out in stages, primarily due to expected high demand.
On the Claude API and consumption-based Enterprise plans, Fable 5 is fully available starting today — no waiting, no restrictions beyond normal usage limits.
For subscription plan users (Pro, Max, Team, seat-based Enterprise), Anthropic is being more conservative:
- Now through June 22, 2026: Fable 5 is included at no extra cost on all subscription plans.
- June 23 onwards: Fable 5 will be removed from subscription plans. Accessing it will require usage credits.
- Long term: Anthropic says it intends to restore Fable 5 as a standard subscription feature as quickly as capacity allows.
Anthropic has been transparent about why: demand is expected to be very high and difficult to predict, so they are pacing the rollout carefully.
On the infrastructure side, Fable 5 is available across the Claude API, Amazon Bedrock (US East and Europe Stockholm regions at launch), Vertex AI, Microsoft Foundry, Claude Platform on AWS, Claude Code, Claude Cowork, and the Claude.ai chat interface.
What About Claude Mythos 5?
Claude Mythos 5 is a different story entirely. It is the direct successor to Claude Mythos Preview, sharing the same Fable 5 architecture but without safety classifiers. This makes it significantly more capable in sensitive domains — which is precisely why public access is off the table.
Access is strictly limited to organizations already approved under Project Glasswing. For new applicants, the path is through Anthropic, AWS, or Google Cloud account teams. Approved use cases include cyberdefense, vulnerability research, drug design, and biodefense screening.
Both Fable 5 and Mythos 5 are classified as Covered Models, meaning they carry a 30-day data retention requirement and are not available under zero data retention policies.
What’s New Under the Hood: Adaptive Thinking
One of the more technically notable changes in Fable 5 and Mythos 5 is how the thinking system works. Adaptive thinking is the only reasoning mode available on both models — it activates automatically and cannot be disabled. The raw chain of thought is never returned; developers can receive either empty thinking blocks or summarized thinking, controlled via the display parameter.
Developers can also control thinking depth and cost using the new effort parameter, and task budgets are available in beta for more granular control over compute allocation.
Why Now? The Bigger Picture
The launch of Fable 5 comes at a particularly loaded moment for Anthropic. The company is widely expected to pursue an IPO later this year, and investor interest in its technology is at an all-time high. Releasing a frontier-class model publicly — especially one that beats GPT-5.5 on key benchmarks — is a significant commercial and reputational move.
At the same time, Anthropic published a statement just days before this launch warning that frontier AI systems are advancing so fast they may soon be capable of recursive self-improvement — autonomously improving themselves without human oversight. The fact that the company is simultaneously raising alarms about AI risk while releasing its most powerful model yet speaks to the difficult balance the industry is navigating right now.
Key Stats at a Glance
- Launch date: June 9, 2026
- API model IDs:
claude-fable-5/claude-mythos-5 - Context window: 1 million tokens
- Max output: 128,000 tokens
- Pricing: $10/M input, $50/M output
- Knowledge cutoff: January 2026
- SWE-Bench Pro score: 80.3% (vs 69.2% for Opus 4.8)
- Mythos 5 access: Project Glasswing only
- Fable 5 access: Claude API, Bedrock, Vertex AI, Foundry, Claude.ai, Claude Code
FAQ
What is Claude Fable 5? Claude Fable 5 is Anthropic’s most capable publicly available AI model, released on June 9, 2026. It is a Mythos-class model with built-in safety classifiers that block responses in high-risk domains like cybersecurity and biology.
What is the difference between Claude Fable 5 and Claude Mythos 5? They share the same underlying model architecture, but Mythos 5 does not have safety classifiers. Mythos 5 is restricted to approved Project Glasswing partners, while Fable 5 is publicly available.
How much does Claude Fable 5 cost? Claude Fable 5 and Mythos 5 are priced at $10 per million input tokens and $50 per million output tokens — less than half the price of the previous Claude Mythos Preview.
Can I use Claude Fable 5 on my subscription plan? Yes, until June 22, 2026, Fable 5 is included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost. After June 23, usage credits will be required until Anthropic restores it as a standard subscription feature.
What is Project Glasswing? Project Glasswing is Anthropic’s initiative for deploying its most advanced, unrestricted models to vetted organizations in critical infrastructure, cybersecurity, and life sciences research.
How does Fable 5 handle refused requests? When the safety classifier declines a request, the API returns a stop_reason: "refusal" response. Refused requests are not billed, and developers can configure automatic fallback to Claude Opus 4.8.
