Anthropic just dropped a massive update, but you need to read the fine print before celebrating. The company launched Claude Fable 5, labeling it as a "Mythos-class" model built for the general public. It boasts stunning capabilities in coding, logic, and scientific research. Stripe allegedly used it to finish a two-month codebase migration in a single day.
If you are paying for Claude Pro, Team, or Enterprise, Fable 5 is already sitting in your dashboard. You don't have to pay anything extra to test it right now.
But there are two major catches that Anthropic buried in the announcement. First, this free ride ends on June 23. After that, you will have to pay for usage credits. Second, you are not actually getting the full, unadulterated power of Anthropic's best technology. The real powerhouse is Claude Mythos 5, and Anthropic is actively keeping it away from you.
The Two Tier Reality of Mythos Class AI
Anthropic is running a split-screen release strategy. They built a highly advanced foundation model and split it into two distinct versions based on who is asking.
The version you can access is Claude Fable 5. It excels at software engineering and complex vision tasks, but it is heavily restricted. Anthropic is terrified that the general public will use these advanced reasoning capabilities to build malware or attack infrastructure.
The unrestricted twin is Claude Mythos 5. This model possesses terrifyingly effective cybersecurity capabilities. During early testing under Project Glasswing, partner organizations used it to scan their software and instantly flagged over 10,000 critical security flaws. Because Mythos 5 can map out vulnerabilities across major operating systems and power grids in seconds, Anthropic refuses to release it publicly.
Instead, Mythos 5 is guarded behind a government-vetted corporate wall. Only about 200 vetted organizations across 15 countries have access. This group includes the US government, Apple, Amazon Web Services, and JPMorgan Chase. If you are an independent developer or a small business owner, you aren't getting in.
How the Stealth Downgrade System Works
The most controversial part of the Fable 5 release is how Anthropic handles safety blockages. Usually, when an AI model hits a safety guardrail, it refuses to answer. It gives you a canned response about how it cannot fulfill the request.
Fable 5 doesn't do that. It uses a silent redirect system.
If you ask Fable 5 a question that triggers its safety filters—especially anything involving network security, code optimization that looks like a vulnerability exploit, or sensitive biological data—the system doesn't stop. It quietly passes your prompt down to Claude Opus 4.8.
"We've therefore launched the model with safeguards that mean queries on some topics will instead receive a response from our next-most-capable model, Claude Opus 4.8." - Anthropic Official Statement
You might think you are getting the cutting-edge performance of a Mythos-class model, but you are actually getting a downgraded response from an older model generation. Anthropic admits these guardrails are tuned conservatively. They trigger in roughly 5% of all user sessions, meaning you will face silent downgrades on completely harmless prompts just because your phrasing looked suspicious to the safety filter.
The Token Pricing Shock
If you plan to use Fable 5 through the API after the free trial ends on June 23, prepare your wallet. This model class is incredibly expensive to run.
Anthropic set the price at $10 per million input tokens and $50 per million output tokens. While that is cheaper than the initial Mythos Preview pricing, it remains a premium cost structure. For context, if you run massive agentic workflows that ingest thousands of pages of documentation and output dense code repositories, your daily API bills will spike significantly.
The performance might justify the cost for enterprise teams. Fable 5 scores at the top of Cognition's FrontierCode evaluation, meaning it produces production-grade code with fewer errors than competing models. It also utilizes dynamic workflows to spin up hundreds of smaller subagents that verify each other's work before presenting it to you. You are paying a massive premium, but you get autonomous execution that actually works.
What You Should Do Before June 23
Don't let the corporate safety drama stop you from maximizing the free trial window. You have until late June to use Fable 5 without touching your API credits or paying extra on your subscription.
Throw your most complex, multi-step engineering tasks at it right now. Feed it your largest code repositories to see if it can handle refactoring tasks that usually take your team weeks. Test its vision capabilities on dense architectural blueprints or financial charts.
Just keep an eye on how it responds. If the output suddenly feels slow, generic, or lacks deep analytical nuance, you probably triggered the 5% safety bucket. Treat this period as a high-intensity test drive to determine if a $50 per million output token model actually delivers enough return on investment for your specific business workflow.
