Frontier AI Just Got Cheaper. Use the Window

On Tuesday, Anthropic released Claude Fable 5, its new top-tier model, priced at US$10 per million input tokens and US$50 per million output tokens. Per the announcement, that is under half the price of the previous frontier tier, with output lengths up to 128,000 tokens, and the model is free to use for Pro, Max, Team and Enterprise subscribers until 22 June, with usage throttling starting on the 23rd.

Two prices halving in a headline is easy to scroll past. I want to make the case that this particular release deserves fifteen minutes of an operator's attention, for three reasons.

1. The cost curve keeps bending in your favour

Every business I speak to has someone senior who says a version of "we're waiting for the technology to settle." This release is another data point that the waiting strategy has the economics backwards. The best available capability keeps getting cheaper, faster than most planning cycles can track. What cost a given amount to run last year costs a fraction of that now, and the pattern has repeated across every major release cycle.

Waiting does not buy you a better entry point. It buys your faster-moving competitor a longer head start with tools that get cheaper for them every quarter too. The sensible response to a falling cost curve is small, early, disciplined exposure, so that when the capability crosses the threshold that matters for your specific workflow, you already know it.

2. The free fortnight is a structured evaluation, if you treat it like one

Anthropic is giving paying subscribers unrestricted access to its best model until 22 June. Most businesses will spend that window letting a few curious staff poke at it. A sharper use of the same fortnight:

- Pick your two hardest real tasks. Whatever your current tools handle worst: a messy multi-document reconciliation, a technical tender response, a gnarly customer complaint thread. Frontier models earn their price difference on hard tasks, so test on hard tasks.

- Run the same inputs through your current setup and the new model. Side by side, same prompt, same source material. Keep the outputs.

- Score them with the person who owns the work. Would they ship it? How many minutes of correction does each output need? That correction time is your real unit cost.

- Write down the verdict before the window closes. Even if the verdict is "no material difference for our work," that is a useful, dated finding you can revisit at the next release.

Two weeks, a handful of hours, and you exit with an evidence-based view of whether the frontier tier matters for your business instead of a vibe.

3. Capability is starting to be gated by trust as well as price

The quieter half of the announcement is worth noticing. Anthropic simultaneously introduced Claude Mythos 5, the same underlying model with fewer cyber-safety restrictions, and per the announcement it is restricted to approved cyberdefence partners and select researchers under a trusted-access programme rather than sold to anyone with a credit card.

That is a structural signal. Vendors are beginning to tier their most capable systems by who the buyer is and how they govern the technology, not purely by willingness to pay. For a mid-market business the immediate impact is nil; you are not bidding for restricted-tier access. But the direction of travel suggests that a business with its AI governance basics in order, meaning a usage policy, clear data-handling rules and someone accountable for the risk, will face fewer doors closed to it over time than a business that treated governance as paperwork. Cheap insurance, and it costs mostly effort.

The honest caveats

Anti-hype means applying scepticism evenly, including to releases I find genuinely exciting. So: the performance claims in this launch are the vendor's own, and independent benchmarking takes weeks to accumulate. Treat the capability claims as promising and unverified, which is exactly why the side-by-side evaluation on your own work matters more than any leaderboard. Your tasks are the only benchmark that pays your invoices.

And a price cut on the frontier tier does not mean every workflow should run on the frontier tier. Plenty of business tasks are handled perfectly well by mid-tier models at a fraction of even the new price. The skill is routing: expensive model for the hard work, cheap model for the volume work. Businesses that never develop that routing muscle end up overpaying in one direction or underperforming in the other.

Fifteen minutes to read the release, a few hours inside the free window, one written verdict. That is the whole ask, and it is a better use of June than another quarter of waiting for the technology to settle. It will not settle. It will just keep getting cheaper for whoever is ready.

Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.