Anthropic unveils Claude Fable 5 with new safety, free until June 2026

Anthropic has launched Claude Fable 5, a public AI model with new safety features and extra guardrails, and it will be free for most users until June 22, 2026. The company says internal tests have not found any universal jailbreaks, but warns there may be false positives with its new safety checks. Early users, like Firefox engineers, suggest the model may help fix software bugs and improve security, though broader evidence is still limited. After June 22, users will need usage credits to keep using Fable 5. Anthropic plans to share more about its safety methods after more public testing.

Anthropic shipped Claude Fable 5 and Claude Mythos 5, with Claude Mythos Preview being superseded by Mythos 5 generation. The public model name is Fable 5, featuring advanced safety controls as a rewritten version of the Mythos announcement. Fable 5 is described as being on the public API and Claude Code at $25 per million input tokens and $125 per million output tokens. The model introduces a pricing structure distinct from other Claude tiers, with access limited to approved organizations rather than being a temporary complimentary subscriber benefit.

The model's official API pricing is set at $25 per million input tokens and $125 per million output tokens. Access is limited to approved organizations and is not described as inclusion in Pro, Max, Team, or Enterprise plans. Organizations will need to purchase usage credits through the standard API pricing model, a transition confirmed by TechCrunch.

Claude Fable 5: Advanced Safety and Early Results

Claude Fable 5 is Anthropic's public-facing Mythos-class AI, equipped with enhanced safety guardrails to prevent misuse. While more powerful, it automatically routes sensitive queries on topics like cybersecurity and biology to a different model. The system is designed to balance advanced capabilities with responsible deployment.

Anthropic states that extensive internal red-teaming has yet to discover any "universal jailbreaks." The company cautions, however, that the new safety classifiers may produce false positives, occasionally routing safe requests to the Opus 4.8 model. Traffic logs for both Fable and Mythos are retained for 30 days exclusively for abuse investigations and are not used for model training.

Early results from adopters are promising. Security engineers at Mozilla, who had access to a Mythos preview, reported significant gains. According to industry reports from the team, the AI helped fix a substantial number of security bugs. This indicates that Fable 5, despite its safety layers, can provide substantial value for software security and development teams.

Meanwhile, access to the unrestricted Mythos model remains limited to partners in Project Glasswing, Anthropic's security initiative. Original-source summaries indicate roughly 12 founding partners plus about 40 additional vetted organizations, with later reporting suggesting the program expanded toward nearly 200 organizations using the Mythos preview for tasks like code scanning and threat response. Anthropic has not announced similar access for biology researchers.

Availability, Pricing, and Future Access

API Pricing: $25 per million input tokens, $125 per million output tokens.
Caching Discount: Prompt caching discounts may be available.
Subscription Access: Limited to approved organizations through API access.
Future Access: Requires usage credits through standard API pricing model.

Independent feedback is still developing. While the results from Mozilla suggest potential productivity boosts for developers, broader evidence has yet to materialize. Anthropic plans to release a detailed safety paper after a period of public testing to explain its new guardrails. Developers can access Fable 5 immediately via the Claude Platform and major cloud providers.

How long will Claude Fable 5 stay available to approved organizations?

Claude Fable 5 is available through API access to approved organizations at standard pricing; access is not described as a temporary complimentary benefit but rather through the pay-per-use credit model. Anthropic has not set specific subscriber inclusion dates, focusing on approved organizational access.

What happens to cybersecurity, biology and chemistry prompts?

Those topics are automatically forwarded to Claude Opus 4.8 instead of Fable 5. Anthropic warns the routing classifiers can still false-positive, so a harmless biology question might briefly hit Opus even if it is safe.

How much more expensive is Fable 5 than Opus 4.8?

API buyers pay significantly more: Fable 5 lists at $25 per million input tokens and $125 per million output tokens, while Opus 4.8 remains at $5 / $25 for the same volumes. Prompt-caching discounts on input tokens may be available for developers who cache repeated context.

Does Anthropic keep my Fable or Mythos traffic for training?

No. Both models' traffic is retained for 30 days strictly for abuse investigations and is explicitly excluded from model-training pools, mirroring the policy already used for Opus.

Who still gets the unrestricted Mythos model?

Access to the full-strength Mythos remains invite-only: Project Glasswing launch partners plus a handful of biology labs scanning for drug-interaction side-effects and genetic-disease variants. The public Fable 5 release is the safety-capped variant of the same weights.