Anthropic's Fable 5.1 Arrives Two Months After Export Shutdown
Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1, the first successors since Fable 5 was disabled by US export controls in June and restored in July. Fable 5.1 costs $10 and $50 per million input and output tokens, with cache reads 75 percent cheaper, and Anthropic says safeguards, not capability, set its benchmark limits.
On this page
What shipped on Monday
Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1 on September 1, calling Fable 5.1 its "most capable model for coding and knowledge work" and "a Mythos-level model built for your most ambitious, long-running projects" [1]. Reuters covered the launch the same day [2]. Fable 5.1 is available to Pro, Max, Team, and Enterprise subscribers, and developers can reach it through the Claude Platform, marketplaces, AWS, Google Cloud, and Microsoft Foundry under the model ID claude-fable-5-1 [1].
Mythos 5.1 ships alongside it at the top of the family, while the public page positions Fable 5.1 as the developer-facing workhorse. The split matters less than what the pair represents: both successors to the model family that spent June in regulatory limbo are now generally purchasable [1][2].
Pricing: cheaper cache, cheaper agentic runs
The pricing is the most concrete change. Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, and cache reads drop to $0.25 per million tokens, which Anthropic says is 75 percent less than Fable 5 [1]. Estimated end results: typical workloads cost about 25 percent less than on Fable 5, and agentic workloads up to around 45 percent less [1]. Inference pinned to US-only infrastructure carries a 1.1x multiplier on input and output tokens [1].
For agent-heavy setups the cache cut compounds: agentic chains re-send the same context on every step, so a three-quarter discount on repeated reads shrinks the bill more than the headline input price suggests [1].
The safeguards trade-off, stated out loud
The benchmark section is unusually honest about what the safety stack costs. Anthropic states Fable 5.1 was tested with production safeguards switched on, and that where those safeguards intervened, the model scored zero on OSWorld 2.0. Flagged cybersecurity tasks were completed by Opus 4.8 instead, and flagged biology tasks by Opus 5 [1]. On Terminal-Bench-Science, the public leaderboard records 21.4 percent for Fable 5.1 against Opus 5's 30.0 percent, with Anthropic's own setup reproducing 24.7 and 29.0 percent [1].
Publishing those numbers inverts the usual marketing pattern, where guarded behaviour stays invisible until users hit it. It also sets expectations for buyers: the headline model and the model that actually answers a flagged request may not be the same system [1]. It tells developers something practical too: if your workload brushes flagged categories, the flagship may hand the task to a sibling model mid-run [1].
From June's export-control whiplash to general availability
The release closes a strange loop. Fable 5 launched on June 9, and access was disabled on June 12 after a US government order limited foreign access [4]. The Commerce Department lifted the controls on June 30 and access returned in July [3]. Two months later, the successor ships to general availability alongside Mythos 5.1 [2].
For anyone choosing models right now, two things follow. The June disruption did not permanently remove these models, but it demonstrated that availability can hinge on a single regulatory decision, a risk worth weighing alongside price. And with cache reads cut by three quarters, teams running repeated agentic workloads have a concrete reason to re-price their Claude usage before assuming frontier costs are fixed [1].