Anthropic's new coding model cuts agentic-task costs up to 45% and false-positive security flags by 60%
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026, with Fable 5.1 running up to 45% cheaper on heavily agentic coding tasks thanks to a 75% cut on cached input reads, and around 60% fewer false-positive security flags inside Claude Code — a real move in the cost and trust economics of AI-assisted development.
13 September 2026
Anthropic released two new frontier models on 1 September 2026: Claude Fable 5.1, generally available to everyone, and Claude Mythos 5.1, restricted to vetted professionals through trusted-access programmes for cybersecurity defence and life sciences work. Both share the same underlying architecture; what differs is the strength of the safeguards around them.
The numbers worth sitting with are the economics. Fable 5.1 runs roughly 25% cheaper than Fable 5 on typical workloads, and up to 45% cheaper on heavily agentic tasks — the multi-step, multi-file, long-running work that AI coding tools are increasingly asked to do unsupervised. Most of that saving comes from a 75% price cut on cached input reads, which matters specifically for agentic sessions that repeatedly re-read large amounts of context. Anthropic also reports Fable 5.1 scoring 52.6% on Terminal-Bench-Science, and — for teams already running Claude Code — around 60% fewer false-positive security flags during code review.
Why the false-positive number matters more than the benchmark
Raw capability scores move constantly and rarely change a buying decision on their own. A 60% cut in false-positive security flags is different: it’s a direct fix to one of the most common complaints about AI-assisted code review, which is that flagging everything as a potential risk is functionally the same as flagging nothing, because reviewers stop trusting the output. If that number holds up in practice, it’s a meaningful step toward AI code review tools that development teams actually act on rather than route around.
So what
Two things worth doing with this: if you’re running or evaluating an AI-assisted development setup, ask specifically which model tier is doing the coding and the review — the gap between Fable-class and Mythos-class access is now a real distinction, not a marketing one. And if cost has been a factor in scoping how much of a build you can afford to have AI-assisted rather than fully manual, this round of pricing cuts changes that maths again. That’s exactly the kind of judgement call worth having a partner make for you rather than guessing — see how we approach it on AI-assisted development.