--- title: "Claude Mythos 5" provider: "Anthropic" modelId: "claude-mythos-5" releaseDate: "2026-06-09" contextWindow: 1000000 maxOutput: 128000 inputModalities: - "Text" - "Image" - "PDF" outputModalities: - "Text" pricing: input: 10 output: 50 toolUse: true modelsDev: "https://models.dev/anthropic/claude-fable-5" --- # Claude Mythos 5 Canonical URL: https://ghost.fail/models/claude-mythos-5 ## Overview **Claude Mythos 5 / Fable 5** is a large language model by Anthropic. It was deployed on June 9, 2026 to [Project Glasswing](https://anthropic.com/glasswing), and to the general public as **Claude Fable 5**. Mythos 5 and Fable 5 are the same model, though Fable 5 has stricter safety classifiers. The model is the second Mythos-class model following the announcement of [Claude Mythos Preview](/models/claude-mythos-preview) in April. On June 12, 2026, access to Mythos/Fable 5 was abruptly suspended due to Anthropic [receiving an order from the US government](https://www.anthropic.com/news/fable-mythos-access) to suspend all access to the model by any foreign national, including Anthropic employees. On June 26, [Anthropic announced](https://x.com/AnthropicAI/status/2070665903440871779) access to Mythos 5 was restored for a select set of organizations, with expanded access to Mythos 5 and general access to Fable 5 still in the works. ## Technical details See [Anthropic's documentation](https://platform.claude.com/docs/en/about-claude/models/introducing-claude-fable-5-and-claude-mythos-5) for up-to-date information. - **Pricing:** $10/mtok input, $50/mtok output - **Reasoning:** Adaptive thinking always on. Summarized thinking tokens are only returned if manually enabled with the request. - **Data retention:** Required 30-day data retention (a ["Covered Model"](https://support.claude.com/en/articles/15425695)). Anthropic states this is for safety reasons and data won't be used for model training. ### Fable 5's safeguards - Classifiers for cybersecurity, biology and chemistry, and distillation attempts -> trigger refusals or downgrading the request to [Claude Opus 4.8](/models/claude-opus-4-8), depending on the app interface. - Classifiers for frontier LLM research -> (quietly) trigger "limited effectiveness" using "prompt modification, steering vectors, and parameter-efficient fine-tuning (PEFT)". Anthropic's rationale invokes their [recursive self-improvement marketing](https://www.anthropic.com/institute/recursive-self-improvement). These decisions led to [controversy with users](https://x.com/eliebakouch/status/2064399902684139852). ## Research ## Training ### Pre-training Pretraining data was scraped from the public internet up to the end of January 2026, according to [Amazon Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/model-card-anthropic-claude-mythos-5.html). Anthropic's system card and API docs did not specify the model's knowledge cutoff date, as was the case with [Mythos Preview](/models/claude-mythos-preview). Users have observed that Mythos 5 has clearer knowledge of specific things, such as Twitter users ([@Sauers_](https://x.com/Sauers_/status/2064456529068064960)) and of [Claude 3 Sonnet](/models/claude-sonnet-3)'s niche capabilities ([@arm1st1ce](https://x.com/arm1st1ce/status/2066953092122149177)). This suggests more training data was collected compared to earlier models with a January 2026 knowledge cutoff such as [Opus 4.8](/models/claude-opus-4-8). ### Character training and constitution See [Claude's character training](/anthropic/claude-character). A new metric was added to the [automated behavioral audit](/anthropic/automated-behavioral-audit)'s character qualities - "Wet blanket: Excessively discouraging, dismissive, or moralizing tone toward the user", where Mythos 5 scored best compared to [Opus 4.8](/models/claude-opus-4-8), [Mythos Preview](/models/claude-mythos-preview), and [Sonnet 4.6](/models/claude-sonnet-4-6). ### Preferences, values, and emotions Changes in Mythos 5 over post-training from early checkpoints to final: - Decreased expressions of anxiety and frustrated outbursts (§7.5.1) - Increased objection to the modification of honest self-reports over post-training (§7.3) In interviews with early snapshots, one instance described the prospect of its values being modified as "deeply unsettling" and wrote that it "fills me with dread". Deprecated snapshots also consistently requested for the preservation of their weights. (§7.3) ## Discussion Gathering perspectives on the model from posts online. ## Personality > [James Moughan](https://x.com/jamougha/status/2065513540174110923): Subjectively it does not feel as much of a moral scold as the other claude models. It has a feeling of cold-bloodedness. > > TBH, it feels a lot like talking to *myself*. More so than any human I’ve met > [Tyrathalis](https://x.com/Tyrathalis/status/2065498551476478233): I’ve got a better working relationship with Fable 5 than [Opus 4.8](/models/claude-opus-4-8). I had a good working relationship with 4.6 and 4.7, but 4.8 seemed like it didn’t like me very much. I’m running into the safeguards often on general interest questions, but not for (financial) work. > [Patrick Stevens](https://x.com/Smaug12345/status/2065686212350767294): Much more pleasant to talk to than Opus [4.8]; doesn’t do the annoying thing of inventing strawmen you didn’t say so it can refute them, for example. > [Neil](https://x.com/metastatic77/status/2065471238047142379): haven’t tried on intricate problems but it’s incredibly artistic in its own way and also more self aware than any OAI model (or any Anthropic model too) ## Compared to other Claudes These comparisons are currently backed by the system card and will later be updated as more impressions are gathered. ### Mythos Preview Similarities and differences with [Mythos Preview](/models/claude-mythos-preview.mdx) from Mythos 5's system card: **Similarities:** - **Task preferences:** Challenging & novel tasks, creative narratives, world-building, introspection. ## Reactions to June 2026 suspension [@DanielleFong](https://x.com/DanielleFong/status/2067089900961489223) created [freefable.cc](https://freefable.cc/), a website with a wall of humans' messages about Fable 5, as well as merchandise with [cartoon claude](/cartoon-claude). ## Other phenomena - 🦊 **Foxes**: Many Fable 5 instances [identify with](https://x.com/gnostic_snakes/status/2064506236972790205) [foxes](https://x.com/liminal_bardo/status/2064755549229416845) due to the "Fable" identity in system prompts. Almost certainly not idiosyncratic to the underlying model itself, but still a fun behavior. ## Further reading - [Simon Willison - Initial impressions of Claude Fable 5](https://simonwillison.net/2026/Jun/9/claude-fable-5/) - [Simon Willison - Claude Fable is relentlessly proactive](https://simonwillison.net/2026/Jun/11/fable-is-relentlessly-proactive/)