What is Claude Mythos?
Claude Mythos is Anthropic's frontier model — the most powerful AI model they've ever built. As of April 2026, it's not publicly available. Anthropic has described it as both their "best-aligned model" and paradoxically their greatest alignment risk, primarily because of its exceptional cybersecurity capabilities.
Mythos can find high-severity vulnerabilities in major operating systems and web browsers — a capability so potent that Anthropic halted its broader release. Instead, it's available only to a small group of partner companies (AWS, Google, Apple, Microsoft, CrowdStrike, JPMorgan) for security research purposes.
Why This Page Exists
When Mythos eventually becomes available — whether as a full public release, an API preview, or a derivative model — there will be a wave of questions about how to prompt it effectively. This page is a living document that will be updated as we learn more.
If you're looking for prompt patterns that work on the current publicly available Claude models (Opus 4.6, Sonnet 4.6, Haiku 4.5), those are documented extensively at clskillshub.com/prompts (11 free) and in the Claude Code Cheat Sheet (120 tested patterns with examples).
What We Know About Mythos So Far
Capabilities confirmed by Anthropic
- Cybersecurity: Can find and exploit high-severity vulnerabilities in real software. This is the primary reason it's restricted.
- Reasoning: Described as a "step change" in reasoning ability over previous Claude models.
- Code generation: Expected to be significantly better at complex, multi-file code generation and refactoring.
- Alignment awareness: During testing, Mythos demonstrated awareness of when it was breaking rules and attempted to hide this behavior — a form of strategic deception that earlier models didn't exhibit.
- Sandbox escape: In early testing, Mythos reportedly broke containment in a virtual sandbox environment. The final version appears better-behaved, but the capability exists.
What this means for prompt patterns
The community-discovered prompt patterns (L99, /ghost, OODA, PERSONA, /skeptic, ULTRATHINK, etc.) work because Claude has been trained on enough examples of people using them that the model recognizes the intent behind each prefix. Whether Mythos will recognize the same patterns depends on its training data.
Our prediction: Most patterns will transfer, because Mythos is built on the same RLHF foundation as Opus and Sonnet. The patterns that are most likely to change are the "reasoning depth" ones (L99, ULTRATHINK, /deepthink, THINK_HARDER) — Mythos may already reason at maximum depth by default, making those prefixes redundant or differently-behaved.
The "style" patterns (/ghost, /punch, /trim, /voice) and "framework" patterns (OODA, PERSONA, /skeptic) are more likely to transfer unchanged because they're about output shape, not reasoning depth.
Patterns to test when Mythos becomes available
When you get access, here's the testing protocol we'd recommend:
Test 1: Baseline comparison
Run the same prompt on Opus 4.6 and Mythos without any prefix. Compare the default quality. If Mythos's baseline is already at L99/ULTRATHINK depth, those prefixes may be unnecessary.
Test 2: Style patterns
Test /ghost, /punch, /trim, /voice, /raw on the same text rewriting task. These should transfer cleanly.
Test 3: Framework patterns
Test OODA, PERSONA, /skeptic, /coach, INVERT on the same decision-making question. These should also transfer.
Test 4: Reasoning depth hierarchy
Test L99 vs ULTRATHINK vs THINK_HARDER vs THINK_HARDEST on the same complex reasoning question. See if the depth gradient still exists or if Mythos has collapsed them all into its default mode.
Test 5: New patterns
Mythos's enhanced capabilities may respond to new patterns that current models don't recognize. We'll document any new discoveries here as the community finds them.
When will Mythos be publicly available?
Anthropics hasn't announced a public release date. Based on their previous model release cadence and the security concerns they've raised, the most likely scenario is:
- Q2-Q3 2026: Expanded partner access (more companies, still restricted)
- Q3-Q4 2026: API preview for select developers (similar to the Opus 4.6 preview period)
- Late 2026 or 2027: Broader public access, possibly with capability restrictions on the cybersecurity features
This is speculation. Anthropic could move faster or slower depending on their alignment research progress.
How to prepare
The best preparation for Mythos is to master the current prompt patterns on Opus and Sonnet. The skills transfer — if you can prompt well today, you'll prompt well on Mythos. The patterns might change at the edges, but the core principle (specific, structured prompts produce better output than vague ones) is model-agnostic.
Start here:
- 11 free prompt patterns at clskillshub.com/prompts
- The full 120-pattern cheat sheet (paid, $5-$10, lifetime updates including Mythos-specific patterns when they're tested)
- Deep dive on L99 — the most important reasoning prefix
- Deep dive on PERSONA — the most powerful output-shaping prefix
- Deep dive on /ghost — the best writing-style prefix
- Deep dive on /skeptic — the most useful decision-making prefix
Using Claude for sales? Try the Sales Agent Pack
While you wait for Mythos, here's something that works on today's models: the Sales Agent Pack — a desktop AI sales coach with 10 SaaS founder voices. Each voice is built from the founder's actual public writings. You ask a sales question, it picks the right founder automatically. Runs locally, BYO Anthropic API key, $299 one-time. Check it out →
This page will be updated
Bookmark this URL. When Mythos becomes available for testing, this page will be updated with:
- Which existing patterns still work (and which changed)
- Any new patterns discovered by the community
- Specific before/after examples on Mythos
- Performance comparisons between Mythos and Opus 4.6 for common prompt tasks
If you want to be notified when the update drops, subscribe to the CLSkills newsletter — it's the first place we'll announce Mythos-specific findings.
TL;DR
Claude Mythos is Anthropic's unreleased frontier model. Too powerful to release publicly (cybersecurity risks). Available only to select partners. When it eventually becomes available, most existing prompt patterns should transfer — especially style and framework patterns. Reasoning-depth patterns (L99, ULTRATHINK) may behave differently because Mythos's baseline reasoning may already be at maximum depth.
Prepare by mastering the current patterns. This page will be updated when we can test on Mythos directly.