The AI Hub · 2026-07-06 · AI Boutique Team · Opinion

Fable-level models for everyone in eight months. Then what?

Today’s frontier is next spring’s commodity. The industry is fighting over the lid of a box whose contents are already out — and the fight is a distraction from the only question that matters: what will you do when the best model is cheap?

REGULATELITIGATEOPEN WEIGHTS DON'T GO BACK IN

Call it what the industry calls it this season — Fable-class, the top shelf, the models that cost $200 a month and headline the keynotes. Whatever the label, the pattern underneath is now four years old and has never once broken: whatever the frontier can do, an open-weight model you can download does within months. The lag used to be a year and a half. Last year it was months. DeepSeek's V4 sits eight benchmark points off the frontier today, MIT-licensed, at a fiftieth of the price. Eight months is, if anything, a conservative estimate for when this year's top shelf becomes everyone's shelf.

The shrinking lag — frontier release to open-weight equivalent
2023
~16 months
2024
~9 months
2025
~5 months
2026
~4 and falling
Our read of release history, approximate · the direction is the point, not the digits · checked 2026-07-06

The fight over the lid

Meanwhile, most of the industry's public energy goes into fighting over a box that is already open. Export controls on chips — while frontier-adjacent models train on Huawei silicon. Licensing and lawsuits — while MIT-licensed weights sit on a million hard drives, unrecallable by any court. Safety pauses argued for at podiums — while the capability being paused is downloadable behind the podium. Some of these fights matter enormously as policy. None of them will change the operational fact your business plans against: frontier-level capability is on a conveyor belt to commodity, and the belt is speeding up.

Pandora's box is the right myth but people quote it wrong. The point of the story isn't that the box was opened — it's that shutting the lid afterwards achieved nothing, and what remained inside was hope. The economic version of hope is this: when capability stops being scarce, everything scarce around capability becomes the business.

What's left when the models are equal

Depreciating — don't build on these
Model quality as a moat
"We use the best AI" as a pitch
Pricing that assumes scarce tokens
Single-vendor lock-in
Compounding — this is the business
Your data, and evals that prove quality
Workflows the model is wired into
Trust, compliance and distribution
People who supervise machines well

Cheap frontier intelligence is the most predictable event in technology right now. The unpredictable part is which businesses will be ready to spend it well. That readiness — the workflows, the baselines, the people — is buildable today, at today's prices, in weeks. Which is, of course, our favourite kind of project.

Get ready before the price drops →Related: China’s open-weight models
Key conclusions
So — what happens in eight months?
01

Nothing your architecture shouldn't already survive. If a better, cheaper model appearing would break your stack, it's broken now. Build switchable: routed models, portable prompts, your own evals as the acceptance test.

02

Every "too expensive" business case gets re-run. The use cases you killed at 2024 prices deserve a second audit at 2027 prices. Keep the kill list — it's a pipeline, not a graveyard.

03

Stop waiting for the dust to settle. There is no settled. The labs will leapfrog each other indefinitely; the only stable ground is your own data, workflows and measurement. Businesses waiting for a permanent winner will wait forever and adopt never.

04

The lid-fight is not your fight. Regulation, lawsuits and export controls will reshape the vendors, not the physics. Watch compliance obligations closely (the EU AI Act bites in August) — spectate the rest.

Book an AI Reality Check

Ready to make the AI you've already paid for pay you back?