Open-weight AI is about three months behind the frontier

Open-weight AI is about three months behind the closed frontier. Not three years. Three months — and last autumn the gap briefly closed to nothing.

That number isn’t a vibe. Epoch AI tracks it directly on their Capabilities Index: frontier open-weight models trail the most capable closed models by 3.5 months on average (a 90% confidence interval of 1.1 to 5.3). And the gap is volatile — “sometimes closing completely.” There was a stretch where Llama 3.1-405B sat level with the best closed model going (Claude 3.5 Sonnet).

This week made it concrete. Z.ai shipped GLM-5.2 under an MIT license — 753B parameters, a 1M-token context. On FrontierSWE it scores 74.4, about one point behind Claude Opus 4.8; on Terminal-Bench 2.1 it’s 81.0 to Opus’s 85.0. And the price, per Simon Willison: $1.40 / $4.40 per million tokens, against Opus’s $5 / $25.

You can now self-host a model within a point of Opus-tier coding — no licence to renegotiate, no API that can be revoked — for roughly a fifth of the output price.

The catch: the frontier isn’t where you left it

“Caught up to Opus” is not the same as “caught the frontier” — because the frontier moved. On the Code Arena WebDev leaderboard, GLM-5.2 lands at #2 — behind Claude Fable 5. Fable 5 (and its sibling Mythos) sits a tier above Opus, and that’s the tier open weights are still chasing.

Open weights have caught last year’s frontier at a discount. They have not caught this year’s.

One number the benchmarks bury: GLM-5.2 is token-hungry — about 43k output tokens per task, up from 26k in the previous version. A fifth of the sticker price is not a fifth of the bill. Measure your own workload before you bank the savings.

So what — if you’re building on this

The takeaway is not “switch to open weights.” It’s narrower: the capability you can own outright is about one quarter behind the capability you rent — and that distance keeps shrinking and reopening.

That changes one architectural decision, not your whole stack. Treat the model as a layer you can swap, not a foundation you pour: a thin interface, your prompts and evals kept model-agnostic, a fallback wired up and tested. Do that, and a model three months behind the frontier stops being a downgrade you’d have to migrate to — it becomes a fallback you flip to in an afternoon.

Which matters more than usual right now, because the top tier is also the most fragile. Days ago Fable 5 was banned under a US export control. The smartest model and the most switch-off-able model are the same model.

So here’s the question worth answering before you need to: if your product’s best feature runs on one company’s API, how many afternoons would it take to move it somewhere else? That number — not the benchmark scores — is your real exposure.