Mythos-Class Models
Mythos-Class Models
A tier of Anthropic Claude models that sit above Opus in capability — a new generation name (Fable) alongside Haiku / Sonnet / Opus. Announced/launched ~June 9, 2026 and covered here via two YouTube reactions: Claude Mythos is Finally Here (Nate Herk) and MYTHOS MYTHOS MYTHOS (Matthew Berman).
Everything here is sourced from creator reactions to Anthropic's announcement/blog, not independently verified. Vendor benchmarks; treat as claims. Note this supersedes older "Opus is current SOTA" lines elsewhere in the vault.
The three models in the tier
| Model | Access | Guardrails | Cost vs Opus |
|---|---|---|---|
| Mythos preview | Project Glasswing (April 2026) | cyber safeguards lifted | ~5× Opus |
| Fable 5 | Everyone, today | "made safe for general use" | ~2× Opus |
| Mythos 5 | Glasswing partners only (security community, US gov, infra/cyber defenders) | cyber safeguards lifted | ~2× Opus (same model as Fable) |
Fable 5 and Mythos 5 are the same model under the hood — the only difference is that Mythos has the cyber guardrails removed. Glasswing preview holders can upgrade to Mythos 5. Berman claims it's a 10-trillion-parameter model, "first of its kind."
Pricing & availability
- $10 / million input, $50 / million output (≈2× Opus; less than half the price of Mythos preview).
- Fable 5 free on Pro Max / Team / Seat-based Enterprise plans until June 22, 2026; from June 23 → usage-credit (token) billing. Anthropic aims to restore subscription access "when capacity allows."
- Fable 5 available in the Claude API; Anthropic expects very high demand.
Capability claims
- State of the art on nearly all tested benchmarks (SE, knowledge work, vision, scientific research). The longer/more complex the task, the larger Fable's lead. Headline numbers (vendor-reported) tabulated in MYTHOS MYTHOS MYTHOS (Matthew Berman) — e.g. SWE-Bench Pro 80% (Opus 4.8 69%, GPT 5.5 58%), Terminal Bench 88%, new #1 on Humanity's Last Exam.
- Mythos 5 has "the strongest cybersecurity capabilities of any model in the world"; internal protein-design experts accelerated drug-design steps ~10×.
- Stripe early test: months of engineering compressed into days on a 50M-line Ruby codebase (codebase-wide migration in a day).
- Pokémon Fire Red completed with vision alone, no helper harness.
- More token-efficient than past Claude models, but very verbose and information-dense.
- Effort levels: both creators advise running at low effort — even that often overkill; Nate pegs Fable-low ≈ Opus-4.8-X-high.
Quirks (Berman, first-hand)
Extreme verbosity, over-eager clarifying questions, and a slow "thinking" start (minutes at ~1,500 tokens, then a burst to ~1.5M) — see MYTHOS MYTHOS MYTHOS (Matthew Berman).
Safety & strategy
- Conservative safeguards triggering <5% of sessions; distillation-flagged requests fall back to Opus 4.8.
- New 30-day data-retention policy for Mythos-class traffic (logged human access, deletion after 30 days, not used for training).
- Both creators read the staggered Fable/Mythos release as Anthropic protecting a lead and accelerating its own research (see Recursive Self-Improvement, Project Glasswing).
Live controversies (per Fable 5 Raises the Bar for AI Ambition)
- Bio guardrails over-trigger. Users hit "chat paused" on basic biology — "mitochondria," "DNA → RNA," even the word cancer rerouted to Opus 4.8. Anthropic owns this in the blog post: bio/chem/cyber/distillation requests auto-route to Opus; 95% of sessions don't fall back. Real biologists' complaints are substantive; "looking-for-something-to-complain-about" Twitter noise should be filtered out separately.
- AI-research distillation filter is invisible — WALKED BACK within 24h. Buried on p13 of the 319-page system card: Fable is "deliberately bad" on frontier LLM-development tasks (pre-training pipelines, distributed training infra, ML accelerator design) — and degraded without telling the user. Critics (Prime Intellect's Ellie Bau / Will Brown, Nathan Lambert, Dean Ball) read this as a response to Chinese low-cost distillation work; collateral damage on legitimate research is real. Anthropic blog framing: "using Claude to develop competing models already violates our terms of service."
warning Superseded per Why Fable 5 is the Most Controversial AI Release Ever: after the firestorm, Anthropic told Wired "We made the wrong trade-off, and we apologize" and made these safeguards visible (explicit refusal or announced rerouting). See Silent Model Degradation for the precedent question that outlives the walkback.
- 30-day retention is an enterprise NDA issue — STILL OPEN. Mythos-class prompts/outputs retained 30 days for trust-and-safety including human review. Mike Taylor: "if you used Claude Fable 5 today with memory turned on, you just violated all your NDAs." The Information's reporting on Glasswing partners shows the same friction at the eye-watering token-cost level. This is a real GA blocker, not theater. Escalation per Why Fable 5 is the Most Controversial AI Release Ever: the policy applies precisely to zero-data-retention customers, "flagged for potential serious harm" is at Anthropic's sole discretion, and Microsoft restricted employee use of Fable 5 in Copilot within a day (The Verge). Whittemore: this, not the research filter, is the unresolved revenue risk going into the IPO.
- Anthropic walked back the "public Mythos in a few weeks" line (The Next Wave of Enterprise AI): from Opus 4.8 launch "Mythos-level coming in the next couple of weeks" to "safeguards… we, and to our knowledge all other AI developers, have yet to develop." Confused messaging; Glasswing meanwhile expanded to 150 partners.
Felix Ryberg's "third era" frame
Felix Ryberg (Claude Code lead) on the Fable 5 launch: "With Fable 5, I've personally moved on to responsibilities or loops. I no longer tell Claude to investigate a particular crash report. It runs a loop watching every crash report that comes in." See Tasks to Responsibilities Shift and Loops as Core Primitive — Fable's capability ceiling is what makes the responsibility-mode work.
Connects to your work
- The launch is the concrete "frontier just moved" event behind the vault's loops / Claude Code Workflows / Token Maxing threads — a model explicitly built for long-horizon, parallel, token-heavy autonomous work.
- The cost/routing angle is the practical enterprise takeaway: reserve Fable for the hardest problems, route the rest down-tier.
Sources
- Claude Mythos is Finally Here (Nate Herk)
- MYTHOS MYTHOS MYTHOS (Matthew Berman)
- Fable 5 Raises the Bar for AI Ambition — third reaction; biology/distillation/retention controversies + use-case classification + Felix Ryberg's third-era framing
- The Next Wave of Enterprise AI — Glasswing expansion to 150 partners; walked-back Mythos-public timeline
- Why Fable 5 is the Most Controversial AI Release Ever — the backlash week: silent-degradation walkback, Microsoft Copilot restriction, lab-power discourse