Anthropic on 9 June 2026 released Claude Fable 5, the first generally available model from its Mythos-class tier, alongside Claude Mythos 5, a more restricted-access version of the same underlying model with some safeguards lifted for approved users. Anthropic positions Fable 5 as its most capable model for general use, but the release comes with three conditions worth noting: high-risk topics are routed away from Fable to Claude Opus 4.8, the model is priced above Opus, and access comes with a 30-day data-retention requirement that, by Anthropic's own statement and independent reporting, applies even to some customers who previously had zero-retention terms.

What Fable 5 is, and how it relates to Mythos

Anthropic previewed the Mythos tier in April 2026 but limited access because of the model's cybersecurity capabilities, restricting it to a controlled set of cyber-defence and critical-infrastructure partners through Project Glasswing, in collaboration with the US government. Fable 5 is the company's route to making that capability tier broadly usable: according to Anthropic, Fable 5 and Mythos 5 share the same underlying model, but Fable 5 adds classifiers and safeguards intended to make it safe for general use, while Mythos 5 keeps more of the unguarded model and stays limited to approved partners. With the launch, Anthropic's lineup spans four tiers — Haiku, Sonnet, Opus and now Mythos.

The numbers Anthropic is reporting

The benchmark claims should be read as Anthropic's claims. In the company's reported table, Fable 5 and Mythos 5 scored 80.3% on SWE-Bench Pro, an agentic-coding test, compared with 69.2% for Claude Opus 4.8, 58.6% for OpenAI's GPT-5.5 and 54.2% for Google's Gemini 3.1 Pro. Anthropic also reported 29.3% on Cognition's FrontierCode Diamond benchmark, more than double Opus 4.8's 13.4%. Those are vendor-reported launch figures, not independent reproduction.

For a concrete example, Anthropic cited Stripe, an early-access customer, which reported that Fable migrated a 50-million-line Ruby codebase in a day — work Anthropic estimated would otherwise take a team more than two months by hand. That is a vendor case study, not a neutral benchmark. The broader caveat holds: the numbers have not been independently reproduced across ordinary enterprise workloads, and benchmark gains do not automatically translate into lower cost, lower risk or better reliability for a given team. Treat them as a reason to test Fable 5, not as proof that it should replace Opus by default.

What the safeguards change

The most important qualifier sits in how the safeguards work. Anthropic says queries touching cybersecurity, biology, chemistry and model distillation are flagged and routed to Claude Opus 4.8 instead of being answered by Fable 5, and that this fallback triggers in fewer than 5% of sessions on average. Because Fable 5 and Mythos 5 are the same model underneath, several of the headline capabilities belong to the shared model — and in the guarded domains, Fable 5 effectively performs like Opus 4.8 rather than at Mythos-class level. Anthropic itself notes that, in blocking mode, its classifiers prevent Fable from making progress on offensive-cyber tasks. Buyers should therefore expect Mythos-level behaviour mainly outside the guarded domains, and Opus-level behaviour within them.

On safety testing, Anthropic said an external bug bounty produced no universal jailbreaks in over 1,000 hours of testing and that external red-teaming organisations also failed to find one — though it noted the UK AI Safety Institute made progress towards a universal jailbreak in an early testing window. These are company-reported results, and Anthropic acknowledges that completely preventing such jailbreaks is likely impossible. Dianne Penn, the company's head of product management for research, framed the release to CNBC as a "race to the top" on safety.

Cost, availability and the data-retention condition

Fable 5 is priced at US$10 per million input tokens and US$50 per million output tokens — about double Claude Opus 4.8, and by several outlets' reckoning among the most expensive of the major commercial models. As TechCrunch noted, that price alone may deter wider adoption at a time when enterprises are increasingly scrutinising AI bills.

Access also comes with a notable condition. Anthropic said it will require 30-day retention on all traffic for Mythos-class models, on both first- and third-party surfaces, adding that the data will not be used to train models and will be deleted after 30 days. TechCrunch reported that this applies even to enterprises that previously held zero-retention agreements, and flagged it as a possible precedent in which access to more powerful models comes bundled with mandatory retention framed as a safety measure.

The model is available now through the Claude API (as claude-fable-5), Claude Code, consumption-based Enterprise plans and Amazon Bedrock. Subscription access is staged: Fable 5 is included in Pro, Max, Team and seat-based Enterprise plans at no extra cost through 22 June, after which, from 23 June, it will require usage credits until Anthropic restores it as a standard feature when capacity allows. The launch also lands as Anthropic moves toward a public listing — it confidentially filed a draft Form S-1 with the US SEC on 1 June 2026 — and days after the company urged AI labs to adopt a coordinated slowdown on frontier development, citing the risk of recursive self-improvement.

The obvious counter-read is that Anthropic is turning safety gating into product positioning. Releasing a model while stating that its most dangerous capabilities remain restricted lets the company claim both capability leadership and caution at once. That does not make the safeguards fake, but it does mean buyers should separate the safety narrative from operational evidence: independent evaluations, internal workload tests, retention terms and incident-response commitments matter more than launch-day framing.

Key Takeaways

  • Anthropic released Claude Fable 5 on 9 June 2026 — the first generally available model from its "Mythos" tier, above Opus — alongside a more restricted Claude Mythos 5 (same underlying model, some safeguards lifted) for approved partners.

  • Benchmarks are Anthropic-reported (e.g. 80.3% on SWE-Bench Pro vs 69.2% for Opus 4.8 and 58.6% for GPT-5.5) and not independently reproduced; treat them as a reason to test, not proof to switch.

  • Fable 5 routes cybersecurity, biology, chemistry and distillation queries to Opus 4.8 (under 5% of sessions), so it delivers Mythos-level behaviour mainly outside those guarded domains.

  • Pricing is US$10/US$50 per million input/output tokens — about double Opus 4.8 — and Anthropic says it will require 30-day retention on all Mythos-class traffic; TechCrunch reported this extends to customers who previously had zero-retention terms.

  • Available via the Claude API (claude-fable-5), Claude Code, Enterprise plans and Amazon Bedrock; free on Pro/Max/Team/seat Enterprise plans through 22 June, then usage-credit-based from 23 June.