Anthropic has released Fable 5.1 and Mythos 5.1. In many ways, this is an iterative update befitting of the x.1 label: higher benchmark scores, comparable functionality, and a slightly lower price than its predecessor. Nevertheless, this marks the first time a new state-of-the-art LLM has been subject to comprehensive EU AI Act regulations, with all that entails.
To start with the usual, Fable 5.1 excels in both internal benchmarks and the suite of tests from the well-known AI testing firm Artificial Analysis. The same is partly true for Mythos 5.1, the variant of the same model that is available only to a select group of validated organizations for security purposes. Whereas Opus 5 surpassed Fable 5’s scores but fell short in real-world use, Fable 5.1 will indeed feel like a meaningful, albeit minor, upgrade to many. Artificial Analysis scores 66, compared to 63 for Opus 5 and 62 for Fable 5.
Cheaper, but with an asterisk
In practice, Fable 5.1 can be up to 45 percent cheaper than Fable 5 for agentic tasks. This is primarily due to 75 percent cheaper cache reads, that is, reading information already stored in Anthropic’s memory. Since longer prompts, chats, and agentic sessions must process a large amount of input that has been used before, this can logically reduce costs significantly.
On the other hand, the input and output prices are still $10 and $50, respectively, significantly higher than those of the competition. With options like DeepSeek-V4 Pro, GLM 5.3 Flash, and OpenAI’s subscription plans, price is still not a distinguishing advantage for Anthropic.
Additionally, the EU AI Act has an impact on Fable 5.1 that will only become apparent later. AI watermarks will change the considerations surrounding AI use. For example, anyone who simply wants to check a thesis, legal document, or other text for spelling errors will (according to Anthropic) receive an output that is traceable as having been generated by AI. This isn’t evident from reading the text itself, but rather through a secret set of considerations regarding the chosen tokens in an algorithm that only Anthropic possesses.
Fable 5 had such invasive data retention policies that many large companies absolutely refused to use it. Each session could be stored for up to 30 days without exception, based on alleged security considerations. Now, customers can opt for an Enterprise Frontier Safeguard (EFS), which allows large enterprise customers to store their data within their own cloud infrastructure rather than on Anthropic’s systems. 100 customers have adopted the new policy.
Performance
As usual, we should take Anthropic’s benchmarks with a grain of salt. This isn’t because the scores of Fable 5.1, Mythos 5.1, and earlier Claude models are in question. Rather, it is striking that OpenAI’s GPT 5.6 Sol so often fails to achieve a score. This is largely because Anthropic either chooses not to accept the GPT model’s chosen method, on the grounds that it would undermine the test, or because the model fails to comply with another basic condition of the benchmark without being explicitly told to do so.
In any case, according to Anthropic, Fable/Mythos 5.1 represents another breakthrough in some areas. For example, investment firm Millennium, an early user of the model, discovered the cause of a persistent crash issue that had been occurring for years. Anthropic’s other charts show that Fable 5.1 matches the performance level of Fable 5 at a low level of reasoning, and with “xhigh” and “max” settings, it even outperforms it by a wide margin.
Also read: Claude Fable 5 is Mythos for the masses