---
title: "It's Not the Models. It's the Clock."
description: "One anonymous model on OpenRouter changed my mind about how fast this goes. Not a benchmark. A nameless Chinese model that went to number one before anyone knew who made it."
canonical: https://www.askgib.com/articles/its-not-the-models-its-the-clock
last-updated: 2026-08-24
lang: en
---
# It's Not the Models. It's the Clock.
> One anonymous model on OpenRouter changed my mind about how fast this goes. Not a benchmark. A nameless Chinese model that went to number one before anyone knew who made it.
By Gibril Bachouchi. Filed 2026-08-24. 5 min read. Web page: https://www.askgib.com/articles/its-not-the-models-its-the-clock
Languages: [en](https://www.askgib.com/articles/its-not-the-models-its-the-clock.md) · [fr](https://www.askgib.com/fr/articles/its-not-the-models-its-the-clock.md) · [ar](https://www.askgib.com/ar/articles/its-not-the-models-its-the-clock.md) · [es](https://www.askgib.com/es/articles/its-not-the-models-its-the-clock.md) · [ru](https://www.askgib.com/ru/articles/its-not-the-models-its-the-clock.md) · [pt](https://www.askgib.com/pt/articles/its-not-the-models-its-the-clock.md) · [it](https://www.askgib.com/it/articles/its-not-the-models-its-the-clock.md) · [de](https://www.askgib.com/de/articles/its-not-the-models-its-the-clock.md) · [ja](https://www.askgib.com/ja/articles/its-not-the-models-its-the-clock.md) · [ko](https://www.askgib.com/ko/articles/its-not-the-models-its-the-clock.md) · [hi](https://www.askgib.com/hi/articles/its-not-the-models-its-the-clock.md) · [he](https://www.askgib.com/he/articles/its-not-the-models-its-the-clock.md) · [zh](https://www.askgib.com/zh/articles/its-not-the-models-its-the-clock.md) · [tr](https://www.askgib.com/tr/articles/its-not-the-models-its-the-clock.md) · [kab](https://www.askgib.com/kab/articles/its-not-the-models-its-the-clock.md)
Everyone argues about which model is best this month. Wrong argument.
The thing that changed my mind about how fast this goes wasn't a benchmark, and it wasn't a release. It was one anonymous model on OpenRouter.
A nameless model shows up. No lab, no paper, no launch post. Nobody knows who made it.
It goes to #1 by usage, more than doubling DeepSeek.
Two weeks later Z.ai says it's theirs, releases the weights as GLM-5.3-Flash, and mentions almost in passing that it had been serving the thing on Chinese-made chips the whole time.
Read that again. Frontier-adjacent quality. Open weights. Not on Nvidia. And developers found it and voted for it with their traffic before anyone told them whose it was, what it cost to train, or which flag was on it.
That is the whole story. Not the leaderboard. The fact that the gap was already closed and nobody noticed until the scoreboard said so.
For context, this didn't come from nowhere: GLM 5.2 shipped in June with open weights and an MIT licence and landed a few points off the closed frontier, and GLM 5.3 followed in August with no new pre-training at all. Every gain came from post-training. But the releases are not the point. Ox Alpha is the point, because it was the one that arrived unlabelled.
There is a pattern here, and it is older than AI.
China enters a market. China succeeds in that market. And then the thing gets cheap enough that everyone can have it. Solar panels. Batteries. Phones. Drones. EVs. Every time, the same shape.
The Western story about this is that it is unfair. Subsidies, dumping, a rigged table. That story is mostly wrong, and it has been wrong for a while. The price falls because the manufacturing gets good, and for the last few years the quality has come with it. They are not only cheaper now. They are competitive on the product itself: quality, volume, and price at the same time. That is a different competitor than the one people are still arguing about.
Point that at intelligence and you get the obvious conclusion. Their coding plan is already priced at roughly a tenth of Claude Code. That is not the floor. That is a waypoint.
I think it ends where social media ended. Free to use. Not free to build. Free to use. Somebody eats the cost because attention, distribution or data is worth more than the inference bill. We have watched this movie.
Epoch AI tracks this properly, so I'll use their figures rather than mine.
**The open-closed gap has collapsed.** In 2024 the lag between the best open model and the best closed one was about a year. By late 2025 it was around three months. Through the first half of 2026, four months. A year became a quarter in eighteen months.
**Frontier capability reaches consumer hardware in about a year.** Epoch's own framing: models that fit on consumer GPUs match frontier performance within a year or less.
**Efficiency compounds.** Pre-training compute efficiency improves about 3x per year, the same capability for a third of the compute, doubling every seven and a half months. Chip performance per dollar is up 49% a year since 2023. And frontier capability itself sped up: the capabilities index climbed around 8 points a year before April 2024, and about 15 a year since reasoning models and RL scaling took over.
Stack those. In 2024 you needed a 70B model to get GPT-4 quality. In 2026 Phi-4 runs on 8GB of VRAM and matches models three times its size on math. A 27B Qwen variant hits 77% on SWE-bench on 18GB. A 3B model runs 30 tokens a second on a MacBook Air and fits on an iPhone.
The frontier is getting better fast. The floor is rising faster.
Three years.
Whatever counts as frontier intelligence at that moment, Opus-level today and something better by then, running locally. On the phone in your pocket and the laptop on your desk. Not "good enough for simple tasks". The real thing, handling 99.9% of what anyone actually asks a machine to do: writing, code, analysis, planning, tutoring, translation, the lot.
Run the compounding and three years is not brave. 3x efficiency a year is roughly 27x. Add chip gains and you are at roughly 89x on cost per unit of capability, before anyone invents anything clever. The binding constraint on-device is memory, not FLOPs, and distillation is exactly the trick that gets around it. That's how 2024's 70B became 2026's 8B.
What stays in the cloud is the last 0.1%. Frontier research. Long-horizon agents that need a million tokens of context and a day of wall-clock. Anything where a five-point benchmark gap is the entire product. That market will be real and big and it will be a specialist market, the way supercomputing is a specialist market. Everything else is a commodity you already own.
If intelligence is on the device and the price of the hosted version is walking to zero, the model stops being the moat. Three things become the moat instead.
**Distribution.** Whoever ships the default local model on the phone, the laptop and the browser wins the way the default search engine won.
**Data.** Every model will be roughly as smart as every other. The difference is what it knows about you, your company, your customers, and whether you had the right to use it.
**Verticals.** The value moves up the stack into the last mile: the workflow, the integration, the domain. The way it always does once the underlying thing gets cheap.
If the model runs locally, so does everything it knows about you.
Your data stops being something you upload and starts being something that never leaves. The context that makes an assistant useful, your files, your messages, your calendar, your habits, is exactly the context you have the least reason to send anywhere once the model is sitting on the same chip as the data.
Good for you. A problem for an entire industry.
Because the value of that data does not fall. Only the access does. So expect the acquisition to get uglier. Permission prompts worded to be agreed to. Free apps whose actual product is the sync. SDKs buried three dependencies deep. "Help us improve our service" toggles that default to on. The harder the data is to reach, the more inventive the reaching gets, and the less of it happens in daylight.
Then the second-order effect, which I think is underrated.
Search stops being you.
Right now, when you search, the query arrives attached to a browser, a cookie, an account, an IP, and a profile that has been accumulating for fifteen years. When your local model does the searching, the thing arriving at the other end is a machine asking a question. It does not need your identity to get an answer and it has no particular reason to carry one. Twenty prompts from you become five queries that look like nothing in particular, from a model, on your device, with no profile behind them.
That breaks the link the ad-funded web is built on, and nobody had to legislate it. The agent is a natural anonymiser.
It does not have to work that way. Anonymity is a choice the client makes, and plenty of clients will keep the profile attached, because the profile is the business. But for the first time the private option is the easy one, and whoever ships the local model gets to pick the default.
Which loops straight back to the first moat.
A year ago I'd have said "someday". One anonymous model at the top of the leaderboard moved that timeline into a more confident band.
---
*Figures: Epoch AI (open/closed lag, efficiency trends, capabilities index), Artificial Analysis Intelligence Index, Z.ai release notes, OpenRouter usage data, August 2026.*