# Claude 5 The Claude 5 generation is [[Anthropic]]'s 2026 model family. It is worth understanding as a family rather than a list, because the tiers behave differently from every previous generation: two of the models share weights, one tier is not publicly available, and one model silently answers on behalf of another. ## The lineup | Model | API string | Tier | Public | |---|---|---|---| | Claude Mythos 5 | — | Mythos | No | | [[Claude Fable 5]] | `claude-fable-5` | Mythos | Yes | | [[Claude Opus 5]] | `claude-opus-5` | Opus | Yes | | [[Claude Sonnet 5]] | `claude-sonnet-5` | Sonnet | Yes | | Claude Haiku 4.5 | `claude-haiku-4-5-20251001` | Haiku | Yes | Haiku is still on 4.5, so "Claude 5" is not a clean sweep of the lineup. ## The Mythos tier is the structural change Above Opus sits a new tier called **Mythos**, and it changes how you should read the family. **Fable 5 and Mythos 5 are the same underlying model.** Fable carries additional safety measures for biology, cybersecurity, and LLM R&D. Mythos does not. That means the public frontier model and the restricted one are not different capabilities, they are different safety configurations of one set of weights. **Claude Mythos Preview is not publicly available.** It runs with a small number of trusted organizations under something called **Project Glasswing** (anthropic.com/glasswing). I find this the most interesting thing about the generation. The published tier list has always implied capability ordering. It now encodes access policy too, and the top of the ladder is a model most people will never call. ## Safeguards routing: Fable answers as Opus When you select [[Claude Fable 5]], some queries get **silently redirected to [[Claude Opus 5]]**. Anthropic's own wording: > 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. To release the model both safely and quickly, we've tuned these safeguards conservatively. They say it triggers in under 5% of sessions on average and admit it will sometimes catch harmless requests. **The practical consequence for anyone benchmarking or building:** you cannot assume the model you selected is the model that answered. If you are evaluating Fable 5 and your task touches cybersecurity, some fraction of your results are Opus 5 results. Anyone publishing Fable 5 benchmark numbers on security-adjacent tasks needs to account for this, and mostly nobody is. The Opus 5 system prompt carries a dedicated `<fable_safeguards_routing>` block instructing the model on how to explain this to confused users. ## The export control interruption A piece of history the models themselves have to be told about, because it happened after their training cutoff: - **9 June 2026**: Fable 5 and Mythos 5 released - **12 June 2026**: Anthropic suspends access to both, to comply with U.S. Department of Commerce export controls - **30 June 2026**: The Department lifts the controls - **1 July 2026**: Access restored Three weeks of a frontier model being switched off by trade policy. Worth remembering the next time you plan a product around a specific model endpoint. Also see the vault's existing news note on the ban period. ## What changed for the people using them The capability jump is real, but the practical disruption is elsewhere: **these models want less instruction than their predecessors.** Anthropic deleted over 80% of [[Claude Code]]'s system prompt for Opus 5 and Fable 5 with no measurable loss on coding evals. Their own transcripts showed conflicting guidance fighting inside a single request ("leave documentation as appropriate" against "DO NOT add comments"), forcing the model to spend reasoning on reconciling its instructions rather than on the task. The six reversals are documented in [[Context Engineering]]. The short version: give judgment instead of rules, design interfaces instead of supplying examples, disclose progressively instead of loading upfront, stop repeating yourself, let auto-memory work, and use rich references instead of plain specs. If you built an elaborate harness for the 4.x generation, some of it is now working against you. Run `/doctor` in [[Claude Code]]. ## Reasoning effort has a ceiling Vals AI's finding on [[Claude Opus 5]] applies as a family-level lesson: performance rises from low to medium to high, then flattens or dips at the two highest levels while costing considerably more. Combined with [[GPT-5.6]]'s migration guidance to test one level below your current setting, the pattern across both frontier labs in July 2026 is the same. **Maximum effort is not a free upgrade.** ## Caveats - The tier structure and routing details here come substantially from a leaked Opus 5 system prompt, not from a formal product page. The behaviour is corroborated by Anthropic's own safeguards blog quote, but treat specifics as well-sourced rather than official - Model strings and tier names in this generation have moved fast. Verify against current docs before writing code against them - Mythos-tier availability may change ## References - [[Claude Opus 5]] announcement: https://www.anthropic.com/news/claude-opus-5 - Anthropic on Fable/Mythos access restoration: https://anthropic.com/news/fable-mythos-access - Project Glasswing: https://anthropic.com/glasswing - Thariq, "The new rules of context engineering for Claude 5 models" (2026-07-24, 4.3M views): https://x.com/trq212/status/2080710971228918066 - Anthropic blog version: https://claude.com/blog/the-new-rules-of-context-engineering-for-claude-5-generation-models - [[Hacker News]] discussion of the context engineering post (462 points): https://news.ycombinator.com/item?id=49051361 - Opus 5 system prompt leak, Pliny / CL4R1T4S (2026-07-25): https://x.com/elder_plinius/status/2080782804435251373 · https://github.com/elder-plinius/CL4R1T4S/blob/main/ANTHROPIC/OPUS-5.md - Vals AI on reasoning levels: https://x.com/ValsAI/status/2080817011157094511 ## Related - [[Claude Opus 5]] - [[Claude Fable 5]] - [[Claude Sonnet 5]] - [[Claude Opus 4.7]] - [[Anthropic]] - [[Claude]] - [[Claude Code]] - [[Context Engineering]] - [[GPT-5.6]] - [[AI Frontier Model]]