On September 1, Anthropic released Claude Fable 5.1 and, buried in the release notes, one of the more consequential pricing changes in recent AI history: a 75% cut to cache-read token costs, from $1 per million tokens down to $0.25 — confirmed across multiple outlets covering the release.
That sounds like developer-speak. It isn’t. It’s a meaningful part of the reason AI agents were too expensive for most small businesses to run at any real scale — and that constraint just eased considerably.
Why Cache Costs Were the Hidden Killer
Most conversations about AI cost focus on input and output tokens — what it costs to send information in and get a response back. That’s the number on the pricing page. The one that actually determines whether an AI agent is affordable for ongoing business use is cache reads.
When an AI agent is doing real work — processing your customer records, re-reading your product catalogue, reviewing a long document repeatedly across a multi-step workflow — it isn’t starting fresh each time. It’s reading from cache, over and over, across every step of the task. For agentic workflows that touch the same data repeatedly, cached reads can make up the bulk of the total bill, not the input or output tokens most pricing conversations focus on. A 75% cut there changes the maths for exactly the kind of repetitive, document-heavy work an SME would actually want to automate.
What This Actually Changes for a Singapore SME
It doesn’t make AI free, and it doesn’t make every use case suddenly viable. What it does is move a specific category of task from “too expensive to justify” to “worth testing” — anything that involves an AI agent re-reading the same material repeatedly: reconciling a long transaction history against invoices, cross-referencing a product catalogue against supplier price lists, or reviewing a lengthy contract against a standard checklist every time a new version comes in.
If you looked at agentic AI six months ago and the running cost didn’t pencil out for a task like that, it’s worth looking again rather than assuming the maths hasn’t moved. Prices in this category have been falling in a fairly consistent direction through 2026 — this is one data point in a broader pattern, not a one-off event.
The “Too Expensive” Argument Was Already Weaker Than It Sounded
Even before this cut, “AI is too expensive for a business our size” was often a proxy for a different, more honest objection: nobody had actually costed out a specific task against a specific tool. Cost objections that aren’t attached to a specific number are usually cover for uncertainty about where to start, not a real budget constraint — and a 75% price cut on the most expensive part of the bill is exactly the kind of thing that’s easy to miss if nobody’s tracking it in the first place.
The practical move isn’t to adopt AI broadly because the underlying compute got cheaper. It’s to take the one task you already ruled out on cost grounds and re-run the numbers now that a meaningful chunk of that cost has come down.
Where the Money Actually Comes From
Singapore’s own adoption numbers back this up from the demand side: IMDA’s 2025 Singapore Digital Economy Report shows SME AI adoption jumping from 4.2% to 14.5% in a single year (2023 to 2024), part of a digital economy IMDA now puts at 18.6% of GDP. Falling compute costs and rising local adoption are two sides of the same trend — the tools are both getting cheaper to run and easier to justify at the same time.
None of this changes the starting point I’d recommend to any SME: pick the single most repetitive, document-heavy task in your business and test it narrowly before rolling anything out broadly. If cost was the objection last time you looked, that’s now a smaller barrier than it was in August. Enterprise Singapore’s Productivity Solutions Grant also co-funds a range of automation tooling, but only through pre-approved vendors on the Business Grants Portal (50% of cost or S$30,000, whichever is lower) — and it stops taking new applications on 29 September 2026 ahead of the EDGE Grant, so it’s worth checking what the PSG grant covers now, while it’s still open.
About the Author
Keith Kwai is an independent consultant helping Singapore SMEs with digital transformation and AI adoption. He has 25 years of marketing and digital experience across global MNCs including Motorola, Singtel, Creative Technology, Epson, and Scholastic International. He builds agentic AI systems — not just advises on them. Connect at keithkwai.com or LinkedIn.
Sources: VentureBeat, Anthropic Claude Fable 5.1 pricing coverage; IMDA Singapore Digital Economy Report 2025; Enterprise Singapore Productivity Solutions Grant (PSG) guidelines.