Glonce

Claude's system prompts reveal new Mythos tier above Opus 5

modelsThe Glonce editorsHacker News🔥68Glonce rating68Impact40Novelty25Relevance100Surprise20Players80Credibility90
Claude's system prompts reveal new Mythos tier above Opus 5

Anthropic's platform.claude.com "System Prompts" page publishes the system-prompt text behind each Claude model, organized by dated revision. Its newest entry, dated July 24, 2026, is for Claude Opus 5, and it is the first to disclose a new model tier called Mythos that sits above Opus in capability. The tier's current generation consists of Claude Mythos 5 and Claude Fable 5, which the page says share the same underlying model; Fable 5 carries additional safety measures for biology, cybersecurity, and LLM R&D that Mythos 5 does not. An earlier model in the tier, Claude Mythos Preview, is not public: it is used only by a small number of trusted organizations as part of what Anthropic calls Project Glasswing. Fable 5's own system prompt, dated June 9, 2026, describes Fable 5 as the most intelligent generally available model, and states that Mythos 5 is available without Fable 5's extra safety measures only to approved organizations.

Fable 5 and Mythos 5 first released on June 9, 2026. Three days later, on June 12, 2026, Anthropic suspended access to both models to comply with U.S. Department of Commerce export controls. The Department lifted those controls on June 30, 2026, and Anthropic restored access the next day, July 1, 2026. The Opus 5 entry states these facts plainly. An earlier entry, for Claude Opus 4.8 and dated May 28, 2026, had explained Mythos Preview's non-public status by citing cybersecurity concerns specifically, a reason the later Opus 5 entry no longer includes.

The Opus 5 system prompt also quotes an Anthropic blog post explaining a safeguards mechanism built for Fable 5: because Fable 5's capabilities in areas such as cybersecurity could be misused if left unprotected, some queries sent to Fable 5 are instead answered by Opus 5, described in the quote as Anthropic's "next-most-capable model." Anthropic says the safeguards were tuned conservatively so the model could ship both safely and quickly, which means they also catch some harmless requests, triggering on average in less than 5% of sessions, and that the company is working to cut those false positives as more capable models arrive.

The page also shows Opus 5 carries a reliable knowledge cutoff of the end of May 2026, later than the end of January 2026 cutoff stated in the Fable 5, Opus 4.8, Opus 4.7, and Sonnet 4.6 entries on the same page. Separately, the page's own explanatory text notes that starting with the Claude 4.6 generation, each model ID is a single fixed snapshot, so those models get only one dated entry rather than the multiple revisions earlier model versions accumulated over time.

Key facts

  • The July 24, 2026 Opus 5 entry is the first on the page to disclose a new Mythos tier above Opus, made up of Claude Mythos 5 and Claude Fable 5.
  • Fable 5 and Mythos 5 share the same underlying model, but Fable 5 adds safety measures for biology, cybersecurity, and LLM R&D that Mythos 5 does not carry.
  • Anthropic suspended both models on June 12, 2026, three days after their June 9 launch, to comply with U.S. Department of Commerce export controls, then restored access on July 1 after the controls were lifted June 30.
  • Fable 5's safeguards route some queries to Opus 5 instead, and by Anthropic's own account also catch harmless requests, triggering on average in less than 5% of sessions.
  • Opus 5's stated knowledge cutoff is the end of May 2026, later than the end of January 2026 cutoff shared by the Fable 5, Opus 4.8, Opus 4.7, and Sonnet 4.6 entries on the same page.

Why it matters

This is the first public confirmation, in Anthropic's own documentation, that a model tier called Mythos now exists above Opus, and that two of its members, Fable 5 and Mythos 5, already shipped and were briefly pulled back for export-control reasons. The page also documents a quiet change in how Anthropic manages model versions: starting with the Claude 4.6 generation, each model ID is a single fixed snapshot, ending the practice of multiple dated prompt revisions under one version number. Together these are the clearest public signal so far of Anthropic's roadmap above Opus and of how export controls can reach a shipped frontier model within days of release.

Who it affects

Anyone using Claude Fable 5 sees Anthropic's own description of it as the most intelligent generally available model, with some of its queries silently redirected to Opus 5 instead of Fable 5 when a safeguard trips, an event the system prompt tells Opus 5 to explain if the user asks. Developers and API customers relying on the Fable 5 or Mythos 5 model strings were affected directly by the June 12 to July 1 suspension. Only a small number of trusted organizations under Anthropic's Project Glasswing currently have access to Claude Mythos Preview, the non-public model that sits at the base of the Mythos tier.

How to use it

The page itself is public documentation at platform.claude.com; developers can read the exact system-prompt text and model strings ('claude-fable-5', 'claude-opus-5', 'claude-sonnet-5', 'claude-haiku-4-5-20251001') for each currently available model. Access to Claude Mythos Preview is not available through this route: the page states it is limited to trusted organizations inside Project Glasswing, with no public sign-up path described.

How solid is it

The material comes directly from Anthropic's own release-notes page, which reproduces the literal system-prompt text Anthropic ships with each model, so the claims about model names, dates, and the export-control episode are Anthropic's own statements about its own products rather than third-party reporting. No individual person is named or quoted anywhere on the page; the only quoted voice is Anthropic's blog post, cited inside the Opus 5 system prompt itself.

Risks and caveats

Anthropic's own description of the Fable 5 safeguards concedes they are tuned conservatively and catch harmless requests along with the ones they are meant to catch, at a rate the company puts at under 5% of sessions on average. The June 2026 suspension shows that a shipped frontier model can lose availability within days over export-control changes outside Anthropic's control, then regain it just as fast once the controls are lifted. The page gives no reason for why the export controls were imposed in the first place, and does not say whether the suspension covered API access, claude.ai, or both.

“Releasing a model this capable comes with risks. Without safeguards, Fable 5's capabilities in areas like cybersecurity could be misused to cause serious damage. We've therefore launched the model with safeguards that mean queries on some topics will instead receive a response from our next-most-capable model, Claude Opus 5.”

— Anthropic, in a blog post quoted inside the Claude Opus 5 system prompt