Anthropic has released its first Mythos-class model for general use. Claude Fable 5 posts state-of-the-art results on nearly every tested AI benchmark — software engineering, knowledge work, vision, and scientific research. The longer and more complex a task, the wider its margin over prior Claude models. A second release, Claude Mythos 5, shares the same underlying weights but carries safety classifiers lifted in certain areas, available only to vetted partners through Project Glasswing. Both models are priced at $10 per million input tokens and $50 per million output tokens — less than half the price of Claude Mythos Preview.
For context on Anthropic’s model trajectory and earlier frontier releases, see Giganectar’s Opus 4.8 coverage.
- Public access: Claude API, Amazon Bedrock, Google Vertex AI, Microsoft Foundry
- Safety classifiers active for cybersecurity, biology/chemistry, distillation
- Flagged requests auto-fall to Claude Opus 4.8 (<5% of sessions)
- Adaptive thinking always on — raw chain-of-thought not returned
- 30-day data retention on all traffic
- On subscription plans: free Jun 9–22; credits required after Jun 23
- Same underlying model as Fable 5 — cyber safeguards lifted
- Available only via Project Glasswing, in collaboration with the US government
- Successor to Claude Mythos Preview — comparable or stronger at lower cost
- Strongest cybersecurity capabilities of any publicly acknowledged AI model
- Biology trusted access program planned for coming weeks
- Not accessible to general API users or subscription plans
During early testing, Stripe reported Fable 5 ran a codebase-wide migration across a 50-million-line Ruby codebase in a single day — work estimated at over two months for a full engineering team by hand.
On Cognition’s FrontierCode evaluation — which tests whether a model passes difficult coding tasks while meeting high-quality production codebase standards — Fable 5 scores highest among all frontier models, even at medium effort, meaning strong results without burning maximum reasoning tokens. It also scores 80.3% on SWE-Bench Pro, 11 points ahead of the next result. On ViBench (end-to-end vibe-coding), it posts the highest performance of any tested model. For spreadsheet work, it beats Opus 4.8 at every effort level and completes runs 25–30% faster.
Using Mythos 5, Anthropic’s protein design team accelerated aspects of drug design by around ten times. With protein design and bioinformatics tools but no human assistance, Mythos 5 matched or beat skilled human operators — choosing binding sites, selecting and running tools, and recovering from failures. Of 14 protein targets studied, 9 yielded strong drug design candidates currently under investigation.
On novel hypothesis generation, scientists preferred Mythos 5’s molecular biology hypotheses over Opus-class outputs in roughly 80% of blinded head-to-head comparisons. One hypothesis — a novel mechanism for an E. coli protein — was independently corroborated by a separate lab. Mythos 5 also conducted autonomous genomics research across 138 animal species, training a custom ML model that outperformed a model published in the journal Science, despite being 100 times smaller.
Fable 5 is state-of-the-art for vision tasks. It extracts precise numbers from dense scientific figures and can rebuild a web app’s full source code from screenshots alone. Earlier Claude models required a complex helper harness to play Pokémon FireRed — Fable 5 completed the game with only raw screenshots, no maps, navigation aids, or extra game-state information.
On memory and long-context: when tested on the deck-building game Slay the Spire, access to persistent file-based memory improved Fable 5’s performance three times more than the same improvement did for Opus 4.8. Fable 5 also reached the game’s final act three times more often.
On Hebbia’s Finance Benchmark for senior-level reasoning, Fable 5 has the highest score of any model, with substantial gains in document-based reasoning, chart and table interpretation, and problem solving.
IMC noted Fable 5 aced their trading-analysis evaluations nearly across the board — factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis. On legal work, lawyers at Harvey found its redlines matched or beat their existing model in every blind review test. See Giganectar’s coverage of AI in government financial contexts for broader industry background.
Fable 5 is generally available as of June 9, 2026 on: Claude API (model string claude-fable-5), Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry. On Claude.ai — Pro, Max, Team, and seat-based Enterprise plans — it is included at no extra cost through June 22, 2026. After June 23, usage credits are required. It counts as 2× usage on Claude.ai plans.
Mythos 5 is not publicly available. It is restricted to approved Project Glasswing partners with cyber safeguards lifted, and to a limited biology trusted access program coming in the next weeks. All Mythos-class traffic carries a mandatory 30-day retention requirement for safety monitoring. No traffic is used for training. For background on platform-level AI policies, see Giganectar’s iOS 27 child safety and controls coverage.
Covers exploit development, offensive cyber tasks, agentic hacking (reconnaissance, lateral movement, defense evasion). When triggered, Fable 5 does not respond — the request is handled by Claude Opus 4.8 instead, and the user is notified.
Red-team results: External bug bounty produced zero universal jailbreaks in over 1,000 hours. One external partner found Fable 5 complied with zero harmful single-turn requests across 30 different public jailbreak techniques. The UK AISI made progress toward a jailbreak during a brief initial testing window — Anthropic acknowledged universal jailbreaks may be impossible to fully prevent, but aims to make them slow enough to detect and block before scale use.
Currently covers most biology and chemistry requests. The broad scope is intentional for the initial release; Anthropic plans to narrow the classifier as trusted access programs for biomedical researchers expand. Mythos-class models can predict properties of gene therapy delivery vehicles (adeno-associated viruses) without explicit training for those tasks — outperforming dedicated protein language models. The same capability carries dual-use risk in pathogen research, driving the broad initial coverage. The AAV candidates in Anthropic’s evaluation were developed by Dyno Therapeutics.
Detects large-scale attempts to extract Fable 5’s capabilities for training competing models. Anthropic has previously identified coordinated distillation efforts from authoritarian countries. Distillation of Fable 5 could spread near-frontier AI capabilities without the accompanying safeguards. Flagged requests fall back to Opus 4.8. For context on Anthropic’s broader AI safety framework, see Giganectar.
| TOKEN_TYPE | claude-fable-5 | claude-opus-4-8 |
|---|---|---|
| Input | $10 / MTok | $5 / MTok |
| Output | $50 / MTok | $25 / MTok |
| 5-min cache write | $12.50 / MTok | $6.25 / MTok |
| 1-hr cache write | $20 / MTok | $10 / MTok |
| Cache hits | $1 / MTok | $0.50 / MTok |
| Context window | 1,000,000 tok | 200,000 tok |
| Max output/req | 128,000 tok | — |
| Data retention | 30 days (required) | — |
// Prompt caching gives 90% discount on cached input. Fable 5 counts as 2× usage on Claude.ai subscription plans. US-only inference available at 1.1× pricing for data-residency requirements. See Giganectar’s Opus 4.8 pricing comparison.
This piece covered the June 9, 2026 joint launch of Claude Fable 5 and Claude Mythos 5 — Anthropic’s first public release of a Mythos-class model. The two models share the same underlying weights; Fable 5 carries safety classifiers for cybersecurity, biology/chemistry, and distillation, while Mythos 5 has those classifiers lifted for vetted Project Glasswing partners. Benchmark results, the fallback system, the safeguard classifier processes, pricing, and the deployment timeline were all covered above.
The new 30-day data retention requirement on all Mythos-class traffic was also discussed — used for safety monitoring only, not model training. The expansion of trusted access programs for cybersecurity organizations and biology researchers was outlined. For further reading: Anthropic’s AI safety framework at Giganectar · Opus 4.8 release context · US government AI partnerships · macOS 27 and Apple Silicon · iOS 27 platform policies.
