Agent Horizon

Anthropic introduced two new models on September 1: Claude Fable 5.1, generally available to all users, and Claude Mythos 5.1, a version with looser safeguards limited to vetted cybersecurity and life-sciences organizations. Anthropic describes them as the same underlying model, distinguished only by which safety guardrails are switched on. Read the official announcement.

Anthropic says Fable 5.1 sets a new bar for coding, knowledge work, and long-running problem-solving tasks, and that it now outperforms both its predecessor Fable 5 and Claude Opus 5 across every benchmark the company published. On Anthropic’s own numbers, Fable 5.1 tops the Artificial Analysis Intelligence Index at 66, ahead of Opus 5 at 63 and OpenAI’s GPT-5.6 Sol at 61. Anthropic is also cutting cache-read pricing by 75%, which it says lowers real-world costs for agentic workloads by up to 45%, and it added invisible watermarking to comply with the EU’s AI Act transparency rules.

What’s actually new here, and what to watch

The two-tier release structure — one public model with guardrails, one gated model without them for people who need the raw capability — is becoming Anthropic’s standard playbook for handling dual-use capability jumps in biology and cybersecurity, and it’s a sensible middle path between full lockdown and unrestricted release. The claim that safety false positives in cybersecurity work dropped by roughly 60% compared to earlier models also addresses a real, previously voiced complaint from developers who found earlier Claude models over-cautious on legitimate security research.

That said, the pricing story is more complicated than the headline suggests. Independent benchmarking firm Artificial Analysis, which helped Anthropic with pre-release testing, found that at maximum reasoning effort Fable 5.1 actually costs more per completed task than Fable 5, because it uses roughly 1.7 times as many output tokens — the cache-price cut helps, but doesn’t fully offset heavier usage on hard problems. That’s a useful reminder that headline discounts on frontier models rarely tell the whole cost story; teams should benchmark their own workloads rather than take vendor pricing claims at face value.

For developers, the practical takeaway is that Fable 5.1 is a solid, if not dramatically cheaper, upgrade for long-horizon coding and research agents, arriving the same week as comparable pushes from OpenAI, Google, and Meta — a sign that frontier competition on both capability and cost is intensifying across the board.