Anthropic launches Claude Fable 5.1 with lower cache-read pricing and a restricted Mythos 5.1 variant
- Claude Fable 5.1 and Claude Mythos 5.1 use the same underlying model, but Fable is generally available while Mythos is limited to trusted-access programs with safeguards for cybersecurity and life-science work.
- Anthropic says Fable 5.1 costs about 25% less than Fable 5 for typical token-billed workloads by cutting cache-read prices, with savings of up to about 45% for highly agentic work.
- On Terminal-Bench-Science 0.1, Fable 5.1 scored 52.6%, versus 24.7% for Fable 5, 29.0% for Opus 5, and 22.4% for GPT-5.6 Sol in Anthropic's setup.
- Anthropic says its upcoming Enterprise Frontier Safeguards store data entirely in customer-controlled cloud infrastructure, claiming zero-data-retention-equivalent privacy; enterprise rollout begins in phases later this fall, with eligible customers getting zero retention until then.
- The new cyber safeguards produce 60% fewer false positives, according to Anthropic, and permit vulnerability discovery but not exploit development; Mythos 5.1 biology access will begin with a US-government-partnered program for scientists.
Hacker News opinions
My weekly reset came early, but it moved my barely used 5-hour window from 2 hours away to 5 hours away. I effectively lost one 5-hour reset this cycle.
This was great for me. I was at 98% Fable usage and would not reset until Thursday, so getting it early was exactly what I wanted.
I had mostly stopped using Fable because the alignment checks were so sensitive that it kept throwing me back to Opus. They say it is less annoying now, so I am curious to test it again.
I work in defensive cybersecurity and detections, including through the CVP, and I still get downgraded to Opus every time. I cannot use it for my actual job.
I have not seen that problem in my work or among people I know. I mainly use Fable for planning, then implement with Opus, and it has been very good at that.
At first it would reject innocuous work involving migratory-bird biology and seasonal phenomena. A month or two ago it became more willing to process those prompts, so the bio blacklist seems less aggressive.
I use it for web development and it is good, but it was useless for reverse-engineering a ten-year-old video camera and DVR system.
Fable can one-shot many complex problems if I describe them clearly and give it a way to verify the result. I only get punted to Opus occasionally for security-adjacent work.
Cache reads at one quarter of the previous cost should matter a lot for long-running agents, since later calls can reuse cached transcript context.
We saw about a 30% reduction in real-world task cost versus Fable 5 in our Viktor.com evaluations. Caching makes a big difference.
I still think the price is too high for a regular coding agent. I already get few errors with 4.8-class models such as Grok 4.6 and GPT 5.6 Sol or Terra behind a router.
The lower-cost model gap is getting hard to ignore. Open models are increasingly capable for cents on the dollar, and a cache-read discount may not close that gap fast enough.
This is only a cache-read discount. Artificial Analysis shows real usage costing 15% more than Fable 5 for marginal gains.