The number hit my screen at 3:47 AM: 2.4 million tokens per second. Not a spike. A plateau. For seven straight days, an unnamed model on OpenRouter had been out-consuming DeepSeek by a factor of two, and nobody knew who was behind it. That kind of silent dominance isn't marketing. It's a signal buried in compute. I watched fortunes bloom and wither in real-time as developers flocked to this ghost, and I knew the game had changed before the official announcement ever landed.
This is not a review. This is a field report from the front lines of an anonymous model that just became the largest launch in OpenRouter's history.
The Merge That Matters
For two years, Zhipu AI ran a dual-track architecture. GLM handled text. GLM-V handled vision. Two models, two pipelines, two deployment headaches. Ox Alpha changed that calculus in one decisive move: the unified multimodal model. Text, image, and video now flow through a single architecture, a design philosophy that aligns Zhipu with the architectural choices of GPT-4o and Gemini.
But here's what the press release didn't tell you: this merge is not a trivial engineering convenience. It's an architectural declaration. When a company collapses its separate model lines into one, it's admitting that the old division of labor was an inefficiency, not a strategy. The visual encoder now has to handle video-frame sequences, not just static pixels. That requires temporal understanding baked into the architecture, not bolted on.
Code was the law, and I was its restless guardian. I've audited enough model deployments to know that a video input pipeline is where most multimodal systems fail. Ox Alpha didn't just claim video support. It shipped it to anonymous users, free, for weeks.
Speed is survival, but empathy is the signal. And what I saw in the developer response was not hype. It was desperation for a tool that actually works.
The architectural shift matters because it signals a strategic reorientation. Zhipu is betting that a single, unified model will win developer mindshare faster than a fragmented portfolio. That's a defensible bet, but it comes with an unspoken risk: the multimodal tax. Every parameter dedicated to video understanding is a parameter not dedicated to pure text reasoning. The question is whether the text degradation is acceptable to developers who primarily use these models for coding.
The Free Market Gambit
Let's talk about the economic reality of free. Ox Alpha launched anonymously, stayed free for a week, and then got extended for another week. That's not generosity. That's a deliberate acquisition strategy. The cost of serving 2.4 million tokens per second for two weeks is staggering, especially when video input is in the mix.
I watched fortunes bloom and wither in real-time during the DeFi summer, and I recognize the pattern. This is the classic land-grab: acquire market share first, figure out monetization later. The question is not whether Zhipu can sustain this. It's whether the free period is designed to establish a habit loop that persists after pricing arrives.
Liquidity is leaking. Watch closely. If Zhipu prices Ox Alpha too high post-free, the developers who came for the price will leave for the performance. DeepSeek is right there, and they've proven they can win on cost. The real battlefield is not model quality. It's developer inertia.
And here's the counter-intuitive angle that nobody is covering: the anonymous launch itself is a strategic weapon. By removing the brand bias, Zhipu forced developers to evaluate the model on raw merit. That's a high-risk play that only works when you're confident in your model. The fact that they did it with video support, a notoriously hard task, tells me they knew something the market didn't.
The Open Source Catch-22
The announcement says model weights are coming tonight. But here's the trap that the market is missing: open source is not a strategy. It's a liability.
Weights released under a permissive license mean that the model can be fine-tuned, reused, and deployed anywhere. That's great for adoption. But it's terrible for competitive differentiation. Zhipu is not OpenAI. They don't have the brand gravity to command premium pricing on API calls when a company can just deploy the open weights themselves.
The code didn't care about your marketing strategy. It cares about the license you choose. Apache 2.0? MIT? A restrictive research-only license? The license type will determine whether Ox Alpha becomes the next Llama or the next Mistral. And the market hasn't even started talking about this because it's too busy counting the tokens.
Stability isn't a product. It's a prerequisite. And the stability of the Zhipu ecosystem in the West is still unproven. The company's compliance with Chinese regulations may create friction with Western enterprise deployments. That's a risk that doesn't show up in a benchmark table but will absolutely show up in a procurement review.
The Benchmark Gap
Here's the uncomfortable truth. We have zero benchmark numbers for Ox Alpha. No MMLU. No HumanEval. No MATH. No Arena Elo. The only data we have is usage data. And usage is not proof of quality. It's proof of curiosity.
The developers on OpenRouter tried it because it was free and because it was anonymous. The mystery itself drove adoption. Whether that adoption converts to retention will depend entirely on the model's ability to handle long-horizon agent tasks and video reasoning without hallucinating.
I watched the same pattern in 2021 with the NFT mania. A product that's easy to mint and free to trade gets volume. But volume is not value. The value appears only when the project survives its own hype cycle.
Based on my audit experience, the most critical unknown is the context window. Long-running agent tasks require a context length of at least 128K tokens. If Ox Alpha has that, it can become a serious competitor to the closed-source leaders. If it only has 32K, it will fade into the background.
The signal is promising, but the evidence is missing.
The Watchlist
So what do you watch?
First, the license announcement tonight. That single document will tell you more than all the usage charts combined. If it's Apache, Ox Alpha is the new foundation model. If it's a restricted license, it's a demo.
Second, the pricing announcement. Watch whether Zhipu undercuts DeepSeek or positions as a premium. That will tell you whether they're playing offense or defense.
Third, the benchmark community's response. The LMSYS Arena rankings will be the true test. Not the hype. Not the token counts. The actual Elo rating.
Human fear is the only asset I trust right now. Fear of missing out drove the anonymous adoption. Fear of being left behind will drive the enterprise adoption. And fear of a bad model will drive the churn.
The market will decide in the next 30 days. I'll be watching.
The Takeaway
Zhipu has executed a flawless launch. The anonymous unveiling, the free window, the silent dominance of OpenRouter — that's a masterclass in product marketing. But the launch is only a fraction of the game. The next chapter is about whether Ox Alpha can hold the line when it's no longer free, no longer anonymous, and no longer the new thing.
Speed is survival, but empathy is the signal. And the empathy that this market needs right now is honesty about what we don't know. The model might be good. It might be great. It might be a flash in the pan. The data will tell the story. And the data starts rolling in the moment the free lunch ends.
I watched fortunes bloom and wither in real-time. I know the pattern. The real fortune isn't in the token count. It's in the developer mindshare that survives the bill.