Claude Fable 5 Arrives With Mythos-Class Power and Prices

Technology Note By: Mark Tauschek, Info-Tech Research Group

Two months ago, Anthropic did something unusual. Instead of shipping its most powerful model to everyone, it locked Claude Mythos behind Project Glasswing, a gated program for vetted cybersecurity partners including Microsoft, Amazon, and Cisco. The justification was that Mythos was too dangerous to release widely because it was frighteningly capable at finding and exploiting existing software vulnerabilities. The effect was both curiosity and fear about a model most enterprises could read about but not touch.

That changed on June 9. Claude Fable 5 is the first Mythos-class model to be generally released, and Anthropic claims it exceeds every model the company has previously made available. The headline capability is that it can work autonomously for longer than any previous Claude models, so you can run it in an agent harness like Claude Code or Claude Managed Agents and it can work for days, planning across stages, delegating to subagents, and checking its own output. Anthropic also shipped Claude Mythos 5, the unrestricted original, but that one stays inside the Project Glasswing program. No Mythos for you – for now, at least.

The release comes a week after Anthropic filed confidential IPO paperwork. Read into that what you will.

What Makes Fable Different

Fable 5 and Mythos 5 share the same underlying model with the difference being a safeguard layer in Fable 5. Queries that trip Anthropic’s guardrails in high-risk domains, primarily cybersecurity and biology, get automatically rerouted to Claude Opus 4.8, and you pay Opus prices for those rerouted requests rather than Fable prices.

This routing behavior could trip you up. For instance, if your security team plans to use Fable 5 for threat analysis, detection engineering, or malware research, some portion of their queries will automatically move to a less capable model. Anthropic ran an external bug bounty with over 1,000 hours of testing and reported that no universal jailbreak was found, which is reassuring from a safety standpoint but maybe a bit frustrating from a SOC analyst’s standpoint. Security-heavy shops should pilot with realistic workloads and measure how often the model falls back before committing to using it for internal red teaming.

There is also a potential data concern for some organizations because Mythos-class models carry a 30-day data retention requirement for safety monitoring. Anthropic says the data will not train future models, but if your data governance policies require zero retention, this will trigger an internal policy review.

Now, About That Price

Fable 5 costs $10 per million input tokens and $50 per million output tokens. Of course, Anthropic positions this as a bargain at less than half the price of Mythos Preview. Sure, but it is also double Opus 4.8 and the most expensive frontier model on the market today.

The per-token sticker price is part of the risk, but the real risk is the multiplication. Fable 5 is built for exactly the workloads that consume tokens at industrial scale, like multiday autonomous sessions, subagent delegation, and self-verification loops where the model rereads and retests its own work. A chatbot conversation burns thousands of tokens, but an agent running a three-day migration burns tens of millions, and most of those are output tokens at the $50 rate. If your developers have been treating model selection as a simple drop-down menu selection, a handful of them defaulting to Fable 5 for routine work can blow through a monthly AI budget in a week. We have already seen this movie with credit-based licensing surprises elsewhere, and it doesn’t end with a refund.

Three mitigations you should consider before anyone gets API keys:

  • Route by task, not by enthusiasm. Sonnet 4.6 and Haiku 4.5 remain the right answer for the bulk of enterprise workloads. Reserve Fable 5 for problems other models demonstrably cannot solve, which is exactly the use case Anthropic itself describes.
  • Exploit prompt caching. The 90% input token discount for cached prompts is the single biggest lever on agentic workload cost. Make caching a build requirement up front, not an optimization for later.
  • Watch the subscription cliff. Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost only through June 22. From June 23 it will shift to usage credits unless Anthropic extends the window. If your users adopt it during the free window and build it into their workflows, you’ll get the sticker shock on July’s bill, so set expectations now.

Also consider that these models are US-only inference for workloads with data residency requirements and run at 1.1x on both input and output. That may seem modest, but it compounds on already premium rates.

Our Take

Fable 5 is a real capability jump, and the safeguard architecture is a reasonable approach to shipping frontier-model power responsibly. But this is a specialist tool, and it’s priced like one, coming in with a billing transition in two weeks that you’ll have to be paying attention to.

  • Treat Fable 5 as a premium tier for long-horizon agentic work, not a default upgrade. Publish internal model selection guidance/policy before granting access, not after the first invoice.
  • Pilot security and life sciences use cases specifically to measure fallback frequency to Opus 4.8. If a third of your queries reroute, you are paying for a capability you can’t reach.
  • Review the 30-day retention requirement against your data classification policy and any contractual zero-retention commitments you have made.
  • Mark June 23 on the calendar and communicate the subscription-to-credits transition to Claude users now so Finance isn’t unpleasantly surprised in July.
  • Mandate prompt caching standards for any team building on Fable 5. The 90% input discount could be the difference between viable use cases and indefensible costs.

Want to Know More?

Latest Technology Notes

All Technology Notes