Anthropic has launched two new AI models, Claude Fable 5.1 and Mythos 5.1. Fable is the general release. Mythos is a restricted version only available to vetted cybersecurity and life sciences organizations. Both share the same underlying model but differ in how tightly their safety rules are set.
The headline claim is a price cut. Anthropic says routine tasks now cost about 25 percent less, and heavy automated workflows, the kind where the AI runs on its own for a long stretch making many decisions, can cost up to 45 percent less. That comes from a cheaper rate on something called cache reads, a technical cost that adds up fast in these long automated sessions.
Here is the catch. Artificial Analysis, the independent firm Anthropic used to test the model before release, found that the savings mostly disappear if you ask the model to work at its highest quality setting. At that setting, the new model produces roughly 70 percent more output text than its predecessor to get the job done, and that extra text costs more than the cache savings can offset. Net result: a task that used to cost a certain amount now costs about 20 percent more, and still costs more per task than Anthropic's own Opus 5 model, which scores only a few points lower on a broad skills test.
The practical lesson for any business buying AI by the task is simple. Sticker price claims from AI companies describe a best case, and the actual bill depends heavily on which quality setting your team ends up using in practice. If your use case needs top-tier accuracy, you may end up paying more, not less, than before.
The bigger context here is pressure. Claude's prior top model struggled to win over business customers largely because of its cost, and a cheaper in-house alternative had already started beating it on price and performance since the middle of this year. This release is Anthropic trying to close that gap without giving up ground on quality.
Two other changes are worth knowing about. The new models are the first from Anthropic to embed an invisible watermark in their text, with a detection tool being rolled out to regulators, journalists, and fact-checkers so they can verify whether a document was AI-written. This mirrors a broader global push, including new Chinese rules requiring labels on AI-generated content.
Anthropic has also loosened safety filters for questions about cybersecurity and biology, which used to block harmless research questions too often. And the company is shutting down a technique where rival AI labs used thousands of fake accounts to copy Claude's abilities into their own systems, a form of AI theft that Anthropic has publicly accused several Chinese AI companies of running at scale.