According to the latest Latent Space newsletter, Anthropic has released a new model internally codenamed Claude Fable/Mythos 5.1. The newsletter dubs it a new state-of-the-art (SOTA) model, suggesting it outperforms previous versions on key benchmarks, though specific scores are not provided in the source.
The most striking details are the pricing and output trade-offs. The cache price is reduced by 75%, which should make it far cheaper to reuse prompts and context across requests. However, the model also generates 70% more output tokens for the same input, implying that while caching is cheaper, the total cost per response could rise if longer outputs are the default.
The source presents these facts as a direct trade-off: a significant discount on cached reads offset by a surge in token generation. Since this is the only source, no independent verification or contrasting analysis is available, but the numbers suggest Anthropic is shifting cost structure toward higher raw output while rewarding developers who leverage caching.