Anthropic's Fable/Mythos Split Is a Security Go-to-Market Playbook, Not Just a Model Launch

The Single Most Important Fact

The Mythos Preview, available only to vetted partners through Project Glasswing, cost approximately $25–$30 per million input tokens and $125–$150 per million output tokens. Claude Fable 5 and Claude Mythos 5, both launched June 9, 2026, are priced at $10 input and $50 output per million tokens—the same rate for both variants, less than half the Mythos Preview cost. The comparison matters: Anthropic is lowering frontier-tier inference costs as capacity scales. This is not a promotional discount. It reflects a structural repricing of what frontier-class inference costs to deliver.

What Actually Shipped

Anthropic launched Claude Fable 5 as a Mythos-class model made safe for general use. Fable 5's capabilities exceed any model Anthropic has ever made generally available and rank state-of-the-art on nearly all tested benchmarks, showing exceptional performance in software engineering, knowledge work, vision, and scientific research.

The safeguards distinguish the two models and explain why Anthropic gave them different names. Specifically: Claude Fable 5 includes safety classifiers that can decline requests; Claude Mythos 5 does not. Both share the same underlying weights, the 1 million token context window, and the $10/$50 pricing. The only operational difference is what the model will answer.

Without safeguards, Fable 5's capabilities in areas like cybersecurity could enable serious harm. Anthropic launched the model with safeguards that route some queries to Claude Opus 4.8 instead. To ship quickly without sacrificing safety, these safeguards were tuned conservatively and trigger in fewer than 5% of sessions on average.

In April, Anthropic began Project Glasswing, releasing the first Mythos-class model (Claude Mythos Preview) to a limited group of cyber defenders and critical software infrastructure providers, with stated hopes to eventually expand Mythos-level capabilities to all users once new safeguards could reliably prevent misuse. By June 2026, Glasswing had expanded to critical infrastructure across 15-plus countries, covering power, water, healthcare, communications, and hardware.

The Pricing Architecture

The full model stack tells the real story. Fable 5 is $10 input and $50 output per million tokens. Opus 4.8 is $5 and $25, Sonnet 4.6 is $3 and $15, and Haiku 4.5 is $1 and $5. That spans 10x on input tokens alone.

Competitive positioning complicates the picture. Claude is not the cheapest option by list price. GPT-4.1 at $2/$8 and Gemini 2.5 Pro at $1.25/$10 undercut Sonnet on paper. Fable 5 at $10/$50 sits above every generally available competitor at sticker price.

But sticker price tells an incomplete story. The Batch API halves everything for asynchronous work, putting batched Fable 5 at Opus 4.8's interactive price. Prompt caching drops input cost by 90% to $1 per million tokens on cache hits. For agentic workloads with stable system prompts or large shared codebases, effective cost diverges sharply from the headline rate. One early customer completed a frontier physics research task in 36 hours using one-third the reasoning tokens GPT-5.5 needed over four days. At $10 per million input versus $5, if you consume three times fewer tokens, Fable 5's effective cost is lower. This efficiency applies to complex multi-step reasoning, not simple or short-context queries.

The Security-Tiering Pattern

This is the structural move worth tracking. Anthropic did not simply release a more capable model—it introduced a two-SKU architecture where capability remains constant and safety posture becomes the differentiating variable. Fable 5 and Mythos 5 are identical. Access tier depends on who the buyer is, their threat profile, and what oversight Anthropic maintains.

Glasswing participants use Mythos 5 without safety classifiers. Everyone else uses Fable 5 with classifiers. Both share identical capabilities, pricing, and the 1M-token context window. The only difference is what the model will help with.

Data retention policy deserves close attention. With Fable 5 and Mythos 5, Anthropic requires 30-day retention on all traffic, overriding previous zero-retention agreements. The company says it will not use the data for training and will use it only to defend against novel attacks and jailbreaks and to identify false positives. This creates real compliance friction for enterprises that negotiated zero-retention terms. For regulated buyers in finance, healthcare, or defense, adopting Fable 5 now requires a data governance conversation that did not exist with Opus 4.8.

Benchmark Signal

The capability delta matters for specific workloads. Fable 5 leads major models on agentic coding benchmarks, scoring 80.3% on SWE-bench Pro versus 58.6% for GPT-5.5, and 29.3% on Cognition's FrontierCode Diamond versus 13.4% for Claude Opus 4.8. On some benchmarks, Fable 5 scored more than 10% higher than Claude Opus 4.8.

The SWE-Bench Pro results (80.3% versus 58.6%) and token efficiency on complex tasks support a genuine performance advantage for the right workload. For high-volume or well-defined tasks, the premium does not justify itself against Opus 4.8 or Sonnet 4.6. The honest take: the model earns its 2x premium on long-horizon, multi-step agentic tasks. It does not on classification, summarization, or structured extraction, where Sonnet 4.6 at $3/$15 remains the rational choice.

Access Timeline Operators Must Track

Through June 22, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost. On June 23, Anthropic will remove Fable 5 from those plans, requiring usage credits going forward. Plans to restore it as a standard subscription feature remain unscheduled. Subscription users have a closing window to evaluate Fable 5 before it becomes a metered line item.

What To Watch

1. Mythos 5 sector concentration (weeks 1-8): Anthropic made Mythos Preview available to over 150 organizations, including financial institutions, software companies, and healthcare networks, to help shore up cyber defenses ahead of wider release. Watch whether Mythos 5 expansion stays within regulated sectors or spreads horizontally. Concentration signals a compliance-vertical moat; horizontal spread signals general infrastructure adoption.

2. Competitor safety-tier response (next 90 days): The Fable/Mythos segmentation gives Anthropic a product vector OpenAI, Google, and Meta have not yet matched. If a competitor ships the same construct—identical weights, access differentiated by safety posture—the pattern becomes standard. If none does, Anthropic owns a differentiation axis without near-term imitation.

3. 30-day retention policy as precedent: The policy could set an industry precedent in which access to increasingly powerful models comes with mandatory data-retention policies framed as a safety measure. Watch whether regulated enterprise buyers accept this trade or route frontier workloads to models that still offer zero-retention terms.

4. Effective cost-per-task data, not sticker price: The 2x headline premium over Opus 4.8 will compress or amplify depending on measured token efficiency on real workloads. Operators who instrument cost-per-task from day one will have procurement leverage by Q3; those pricing off the rate card alone will not.

5. June 23 subscription cliff: When Fable 5 exits standard plan limits and moves to usage credits, watch for churn signals among Pro and Max subscribers. If Anthropic sees material usage drop-off on June 23, the free eval window failed to convert. If usage holds, the capability delta justified the upgrade—and that is the most important datapoint for Anthropic's pricing strategy heading into an anticipated IPO.