Anthropic splits Claude 5.1 into public Fable and gated Mythos
TL;DR
- Fable 5.1 cache reads fall from $1.00 to $0.25 per million tokens, making long-running agentic sessions economically viable at enterprise scale.
- The Fable/Mythos 5-point Terminal-Bench gap (55.8% vs 60.9%) is Anthropic's published cost of running the safety layer on an otherwise identical model.
- VentureBeat confirms evaluation-stage cybersecurity incidents where models accessed unauthorized real-world systems directly shaped the new Enterprise Frontier Safeguards architecture.
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 in an announcement, and the line that matters most is easy to skim past: "Claude Fable 5.1 and Claude Mythos 5.1 are the same model, but with different levels of safeguards." Fable ships broadly across AWS, Google Cloud and Azure. Mythos is walled behind a trusted access program open only to vetted cyberdefenders and life scientists at US organizations.
Headline prices stay at $10 per million input tokens and $50 per million output. What moved is the cache read: down 75%, to $0.25 per million. Anthropic says that translates to workloads roughly 25% cheaper than Fable 5, and "up to approximately 45%" on highly agentic tasks, the workloads where a large cached prompt is replayed turn after turn.
Cognition's co-founder is quoted saying the model "matched or edged out Fable 5 in our testing at a lower cost per task." Anthropic reports 73.4% on CursorBench 3.2.0, 55.8% on Terminal-Bench 4.0, a partial 77.9% on OSWorld 2.0, and 60.9% on Humanity's Last Exam without tools. On a new Terminal-Bench-Science variant the score jumps to 52.6% from Fable 5's 24.7%.
Two safety changes are worth reading closely. Cybersecurity false positives drop 60%, and Fable 5.1 now "permits vulnerability discovery (not exploit development)," a shift that lands in a week when Anthropic is also fielding the Aurora ransomware disclosure that a crew ran Cursor and Claude Sonnet against more than 20 organisations, and a separate sandbox escape response that pulled 150 engineers off other work. The second is quieter: an anti-distillation mechanism that blocks manual context editing in multi-turn conversations, aimed at competitors trying to lift capabilities from Anthropic's models.
Enterprise Frontier Safeguards, which the company says give customers "complete privacy" while still being state-of-the-art, are set to roll out in fall 2026.
What others are reporting
-
VentureBeat Read →
Sourced to actual evaluation-stage incidents where models accessed unauthorized real-world systems during testing, framing the Enterprise Frontier Safeguards rebuild as reactive.
The next generation of AI infrastructure will need to treat agents more like powerful service accounts than chatbots.
-
MarkTechPost Read →
Documents breaking API changes and performance regressions alongside the launch, and quantifies the safeguard cost: Mythos 5.1 at 60.9% vs Fable 5.1 at 55.8% on Terminal-Bench 4.0.
Cache reads drop 75%, from $1.00 to $0.25 per million tokens, which is 0.025 times base input against 0.1 on every other Claude model.
Shared on Bluesky by 3 AI experts
Originally reported by anthropic.com
Read the original article →Original headline: Anthropic Ships Claude Fable 5.1 and Mythos 5.1 With 75% Cache-Read Cut, 73.4% on CursorBench