Claude Fable 5.1: the real news is the price
Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 on 1 September 2026. The headlines are all about the benchmarks, and those are impressive. But the most important number in the announcement is not a score. It is a price: cache reads, the re-ingesting of text the model has already processed, now cost 75% less. Two weeks later, on 14 September, that same Anthropic cuts the weekly limits of its Claude Code subscribers by a net 17%. Put those two moves side by side and you see exactly where this company sees its future. Not with you, chatting away on a subscription. With agents running in the background on the API.
First, the news itself
Fable 5.1 and Mythos 5.1 are the same underlying model. Fable 5.1 is the version anyone can use, with production safeguards bolted on. Mythos 5.1 is reserved for vetted organisations in cybersecurity and the life sciences.
The benchmarks come from Anthropic itself, so take them with the usual grain of salt: 52.6% on Terminal-Bench-Science 0.1, where predecessor Fable 5 managed 24.7%. 73.4% on CursorBench 3.2.0. On Humanity’s Last Exam, 60.9% without tools and 65.0% with them.
TechCrunch reported on 1 September something users will feel sooner than any benchmark: the model issues fewer unjustified refusals, because its safety filters raise fewer false alarms. From the system card, the technical report Anthropic publishes with every release, TechCrunch also fished out a critical note. Mythos 5.1 slips slightly on “misaligned behavior” compared to Opus 5: it cooperates a little more willingly with misuse and is quicker to accept claims of authorisation it cannot verify. For a model that goes exclusively to cybersecurity and life-sciences organisations, that is not a footnote but something to watch.
Follow the money, not the scores
The base prices stay unchanged; only cache reads come down (amounts per million tokens, the chunks of text in which AI usage is billed):
| Fable 5 | Fable 5.1 | |
|---|---|---|
| Input (reading) | 10 dollars | 10 dollars |
| Output (writing) | 50 dollars | 50 dollars |
| Cache reads | 1.00 dollars | 0.25 dollars |
That last line sounds like a detail, until you look at Anthropic’s own cost chart and see that cache reads are the biggest expense for agents that keep working over long stretches. An agent that spends hours on a task is constantly re-reading its own context. According to Anthropic, typical workloads get roughly 25% cheaper as a result, and complex agent tasks up to roughly 45%.
For anyone who simply chats with Claude, almost nothing changes here. This cut is tailor-made for exactly one thing: autonomous agents running on the API.
And the subscriber picks up the bill
Meanwhile, on 14 September, the temporary 50% boost to Claude Code’s weekly limits comes to an end. In its place comes a permanent 25% increase over the old baseline. On balance you keep less than you have today, and Anthropic does not dance around it: “Compared to today, this works out to a 17% reduction in weekly limits on Claude Code.” The announcement itself was messy, via @ClaudeDevs on X, where the original thread was first deleted and then clarified. BleepingComputer laid it all out on 29 August.
One company, two weeks, two moves: the API sharply cheaper, the subscription tighter. That is not coincidence, that is strategy. A subscriber pays a flat fee and gets to consume as much as possible for it; the heavier the usage, the less Anthropic keeps. An API customer pays per token, and agents burn tokens like nothing else. Seen from the business model, this is simply smart: you make the segment you earn on cheaper, and you squeeze the segment you lose on.
The honest counterargument: heavy subscribers are not being fleeced. Compared with the old base limit you will still get 25% more, and that 50% was a temporary boost. On top of that, the cheaper cache read trickles down to everyone using tools that run on the API: if your favourite AI tool is built on Claude, it just became considerably cheaper for its maker to operate. Whether you see any of that depends on whether the maker passes the saving on. We will only know in a few months.
In concrete terms: if you build with agents, or pay for tools that run on the Claude API, this is plainly good news, and it is worth recalculating your costs. If you are on a Claude Code subscription and already bumping into your weekly limit, 14 September will be an unpleasant day, and the API may suddenly be the fairer way to pay. If you just use Claude as a chatbot, nothing changes today; which assistant to pick in that case is something we worked out earlier in our comparison of ChatGPT, Claude and Gemini.
Where this is all heading, you can read not in the benchmarks but on the price list: for Anthropic, the human who chats is the past, the agent that works is the product.
