Anthropic released Claude Fable 5.1 on 1 September, alongside a second model called Claude Mythos 5.1 that most people will never be able to call. The line doing the rounds is that the new model costs about 25 per cent less than Fable 5 for typical workloads, and closer to 45 per cent less for heavily agentic work.
Both of those figures are from Anthropic's own announcement and there is no reason to doubt them. They are also not a price cut, and the difference matters more than the percentage does.
The per-token price did not move#
Fable 5.1 costs $10 per million input tokens and $50 per million output tokens. That is exactly what Fable 5 cost. Nothing about the headline rate changed.
What changed is the cache. Reading from the prompt cache now costs $0.25 per million tokens, a 75 per cent reduction. That single line is the entire saving, and everything else follows from it.
This is why the two advertised numbers are so far apart. A discount that lives in cache reads is worth exactly as much as your workload re-reads context. Long agentic runs do that constantly: the same system prompt, the same tool definitions, the same accumulated transcript, fed back in on every step of a loop that might run for hundreds of steps. Those workloads see something like 45 per cent. A single-shot call that sends a fresh prompt and takes one answer touches the cache barely at all, and pays the same $10 and $50 it paid last week.
So the honest version of the headline is conditional. If your usage looks like an agent loop, this is a substantial cut. If it looks like a classifier or a one-shot summariser, it is not a cut at all. Neither of those is a criticism of the model. It is just a question you can answer about your own traffic before you assume a saving that may not be yours, and if you want the arithmetic rather than the vibe, our prompt caching calculator takes the discount as an input so you can put your own numbers through it.
Worth knowing if you have not used caching before: it is the single largest lever on most API bills and it is routinely left switched off. A 75 per cent cut to a line you are not billed for is not a saving.
The benchmark gains are not evenly spread#
Anthropic published five comparisons against Fable 5. Read as a set rather than as a headline, they say something specific about what actually improved.
The top row more than doubled. The bottom row moved just under three points. That is close to a tenfold difference in improvement across the same release, and the ordering is not random: the biggest jumps are on long-horizon terminal and computer-use tasks, and the smallest is on the one that most resembles ordinary day-to-day coding assistance.
If you are choosing a model to sit in your editor, CursorBench is the row that describes you, and it moved 2.9 points. If you are running unattended multi-step jobs in a terminal, the first two rows are yours, and they moved a great deal. The same release is a major upgrade for one reader and a rounding error for another.
This is the ordinary reason leaderboards mislead even when every number in them is honest. Nothing here is being hidden. The averaging is done by the reader, and the reader usually averages a set of benchmarks that do not describe their own workload.
Mythos 5.1 is the same model with different rules#
The second release is unusual enough to note. Anthropic describes Claude Mythos 5.1 as identical to Fable 5.1, with different safeguards, and it is not generally available. It goes only to trusted access programmes, and to US organisations first.
Two models with the same weights and different guardrails, sold to different customers, is a product structure worth watching. It is a clearer admission than most that the model and the policy layer around it are separable, and that what a lab sells is increasingly the second one.
On that note, Anthropic also says the cybersecurity safeguards in this release fire around 60 per cent less often per session than Fable 5's did. Fewer interruptions on legitimate security work is a real improvement for anyone who has fought a refusal loop while doing their job. It is also, unavoidably, a loosening, and the company is publishing it as a headline feature rather than burying it.
Getting at it#
Fable 5.1 is on the Claude API as claude-fable-5-1, and on Amazon Web Services, Google Cloud and Microsoft Azure. Vercel has it on its AI Gateway as anthropic/claude-fable-5.1, with the cybersecurity and biology safety classifiers on by default and model fallback available for refused requests.
Anthropic's page does not state a context window for the new model, so this piece does not either.
The thing to check first is not a benchmark. It is your own cache hit rate. That number decides whether this release saves you 45 per cent, 25 per cent, or nothing, and it is the one figure the announcement cannot tell you.
Discussion