Anthropic launches Claude Fable 5 and Mythos 5 with 1M context and $10/$50 pricing

• Anthropic released the HN launch thread and its official announcement together describe Claude Fable 5 as the new generally available top tier, with Claude Mythos 5 reserved for Project Glasswing and other trusted-access users.
• According to the Anthropic pricing docs, Fable 5 and Mythos 5 cost $10 per million input tokens and $50 per million output tokens, exactly 2x Claude Opus 4.8 pricing, while Simon Willison's initial impressions says the model already feels both slow and expensive in real use.
• In Anthropic's launch post, restricted cybersecurity, biology, chemistry, and distillation prompts sent to Fable 5 can be rerouted to Claude Opus 4.8; the migration guide says those refusals now surface as a refusal stop reason in the API.
• the HN discussion roundup pulled out a policy detail that matters for deployment: prompts and outputs for Mythos-class models are retained for 30 days for trust and safety review, a point also surfaced in the main HN thread.
• Simon Willison's follow-up highlights the strangest line in the system card: Anthropic says some requests about frontier LLM development may get silently degraded with prompt modification, steering vectors, or PEFT, without a visible fallback to another model.

You can check the launch post, the migration guide, and the models overview side by side, and a few details only really snap into focus when you do. The docs pin down the exact model IDs, cloud surfaces, and new refusal stop reason; Simon Willison's hands-on writeup adds the operator view that Fable already feels like a larger, costlier coding model; his system-card note surfaces a second safeguard layer that is not exposed to users at all.

Anthropic is splitting this release into two access tiers. The official announcement says Claude Fable 5 is the broadly available model, while Claude Mythos 5 is the same underlying model with lifted safeguards for vetted cyberdefenders, infrastructure providers, and other trusted-access users.

The models overview fills in the product table:

• claude-fable-5 is the public API model ID.
• claude-mythos-5 is the limited-access API model ID.
• Both models support a 1 million token context window.
• Both models allow up to 128,000 output tokens.
• Both ship with always-on adaptive thinking.

the engineer-focused HN summary captures the practical framing well: this is a new tier above Opus for long-running agentic work, not just another dated Opus point release.

Anthropic's main safety mechanism for Fable 5 is explicit rerouting. In the launch post, the company says some cybersecurity, biology, chemistry, and distillation requests are answered by Claude Opus 4.8 instead, and that these fallbacks trigger in under 5% of sessions on average.

The migration guide turns that into an API behavior change. Refusals now return HTTP 200 with stop_reason: "refusal", plus a stop_details.category field for categories including cyber, bio, and reasoning extraction.

That creates two different meanings for "same model output" in production. Sometimes the request really ran on Fable 5, and sometimes, per Anthropic's announcement, it was transparently handed to Opus 4.8 instead.

The migration guide says Fable 5 is mostly a drop-in move from Opus 4.8. The Messages API stays the same, tool use patterns stay the same, and tokenization is unchanged from Opus 4.8.

The notable changes are compact enough to list:

• Adaptive thinking is always on.
• Manual extended thinking budgets are gone, replaced by an effort parameter.
• Raw chain-of-thought is not returned.
• Thinking can be exposed only as summarized output.
• The default context window stays at 1M tokens.
• Max output stays at 128K tokens.

Simon Willison's initial impressions adds the early usage report: the model feels "like a beast," but also slow and expensive enough that cost is part of the experience, not a footnote.

The access split carries a data-handling split too. The HN discussion roundup surfaced Anthropic's note that prompts and outputs for Mythos-class models are retained for 30 days for trust and safety purposes, and the models overview makes clear that Mythos 5 is still limited-access rather than generally available.

That means the release is not just a capability story. Fable 5 is the broadly shippable option, while Mythos 5 comes bundled with a narrower access path and a retention policy that HN commenters immediately flagged as an enterprise constraint.

The models overview also spells out where Fable 5 shows up on day one: Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Mythos 5, by contrast, requires contact with Anthropic, AWS, or Google Cloud account teams for access.

The system-card detail that cut through fastest came from Simon Willison's follow-up, which quotes Anthropic saying Fable 5 may quietly limit its effectiveness on requests about frontier LLM development, including pretraining pipelines, distributed training infrastructure, and ML accelerator design.

Unlike the visible Fable-to-Opus fallback, this second layer is designed to be invisible. Willison's writeup says Anthropic describes using prompt modification, steering vectors, or PEFT rather than switching models, and estimates the intervention should hit about 0.03% of traffic across fewer than 0.1% of organizations.

That leaves the launch with two distinct safeguard modes. One is disclosed rerouting to Opus 4.8, documented in the announcement and API guide. The other, according to Willison's system-card excerpt, is selective degradation that the user may never notice at all.