Kimi K3 Hype

Yes, Kimi K3 is impressive. No doubt.

A Chinese open-weight model being seriously compared with Claude Fable / GPT-5.6 level models would have sounded crazy not long ago. On frontend coding, long-context reasoning, and some agentic benchmarks, Kimi K3 looks genuinely strong. Moonshot deserves credit for that.

But I think the market is focusing on the wrong thing.

The real AI war is not about building a “97% frontier model at 1/10 the cost” that dominates Twitter for two weeks.

The real AI war is about building a “95% good-enough model at 1/100 the cost” that quietly eats production workloads.

That is why DeepSeek’s strategy is much more interesting to me.

Look at OpenRouter usage. DeepSeek V4 Flash and V4 Pro have processed over 8 trillion tokens since late April. More importantly, the trend matters: OpenRouter reported that DeepSeek’s token share roughly doubled from January to June, became the top model author by token share in mid-May, and was still #1 by token share in July.

That is not benchmark theatre. That is developers voting with API bills.

And this is actually a much harder “benchmark” to max out.

To max out public benchmarks, you can post-train aggressively on known task formats: coding tests, math problems, frontend leaderboards, and agentic evals. You may get a model that looks amazing on charts.

But benchmark maxxing often creates jagged behavior: surprisingly strong in some narrow formats, weirdly brittle in others. A model can pass the eval, then fail in real-world workflows in ways that feel unnatural and hard to debug.

To max out OpenRouter usage sustainably, the model has to be stable and usable enough for day-to-day agentic work.

It has to be cheap, fast, reliable, and general enough for messy real-world tasks: coding agents, customer support bots, document pipelines, research workflows, data extraction, tool calling, summarization, translation, and millions of boring production calls that never show up on leaderboards.

That is a very different test.

And the pricing explains why DeepSeek is winning that test. DeepSeek V4 Flash is priced at cents-level per million tokens. DeepSeek V4 Pro is still very cheap relative to frontier models. Meanwhile, models like GLM, MiniMax, and now Kimi K3 may be technically impressive, but they are still not playing the same cost game.

This is also where DeepSeek’s R&D focus matters.

DeepSeek is not only training strong models. A big part of its effort is focused on efficiency gains in model serving - both infrastructure improvements and algorithmic improvements that directly reduce token cost.

That may sound less sexy than topping a benchmark leaderboard, but it is exactly the kind of work that compounds in production.

For a demo, the cost difference does not matter. For production, it matters a lot.

A model that is 2% better on public benchmarks but 10x or 50x more expensive may make headlines, but it will not automatically win production traffic.

If the task is mission-critical and requires frontier reasoning, sure, use the best model available: Fable, GPT-5.6, or whatever is SOTA at the time.

But a huge amount of real-world AI workload does not require the absolute best model. It requires a model that is reliable enough, fast enough, cheap enough, and available at scale.

That is where DeepSeek’s approach is dangerous.

It is not necessarily trying to win every leaderboard. It is trying to become the default commodity intelligence layer.

A 97% model at 1/10 cost is great for PR, fundraising, IPO storytelling, and valuation games.

A 95% model at 1/100 cost changes user behavior.

The first creates narrative. The second creates distribution.

And in AI, distribution plus cost efficiency may matter more than marginal benchmark wins.