The Chinese models just cut your vendor's price umbrella

For a while the story was simple. The best models came from a few US labs, they cost what they cost, and you paid it. That story cracked last week. A Chinese lab shipped an open-weight model that scores near the top of the charts, and a second one showed up right behind it. Let me catch you up on what actually changed and what to do about it.

The gap you were told was safe just shrank

Moonshot AI's Kimi K3 is a 2.8 trillion parameter model that lands near the top of independent rankings. Nathan Lambert clocked it at #2 on the Vals AI index and #3 on Artificial Analysis's Intelligence Index, beaten only by Claude Fable 5 and GPT-5.6 Sol, while costing less. Alibaba followed days later with Qwen3.8, which it calls "second only to Fable 5".

The lead everyone argued about, six to nine months, is now closer to three to five. The UK's AI Security Institute put real numbers on it: recent open models "perform similarly to frontier closed models released 4 to 7 months before them," down from a 6-to-10 month gap through 2025. That is not a rounding error. It is your buffer getting cut in half.

Free does not mean cheap to run

Before you plan a migration, read the fine print on cost. Open weights are free to download, but they are not free to serve. Ben Thompson lays it out plainly: Kimi K3 runs at $3 per million input tokens and $15 per million output, cheaper than Sol's $5 and $30, but that gap can vanish. Reasoning models burn chain-of-thought tokens, and Kimi reportedly uses far more of them to reach an answer.

So the real unit is intelligence, not tokens. If two models get the right answer, the answer is the same. What differs is how much compute you burned getting there. Azeem Azhar's read is worth holding onto: Kimi is "pressure, not displacement," and it costs about the same per token as GPT-5.6 Sol, far from the usual China discount. Cheaper on paper, not always cheaper in production.

Your leverage is real, your migration is not

Here is where this touches your roadmap on Monday. Even if you never deploy Kimi, it changes the negotiation. Braden Hancock of Snorkel AI told TechCrunch that frontier-caliber open models "will place a squeeze on the margins and bring down the prices of the frontier companies." The price you pay OpenAI and Anthropic today reflects a compute shortage, not their true cost floor. A credible open alternative gives you something to point at.

The catch: enterprises do not buy on price alone. Security, support, and the tooling built around the closed models still matter. And cheaper intelligence still waits for the monthly management meeting and a slow approval cycle. Leverage in the contract talk, yes. A drop-in swap, not yet.

The fight in Washington is about your access

There is a policy layer that could reach your stack. Axios reported the Trump administration is weighing a ban on K3 and other advanced Chinese models, pushed by US labs. Dean Ball, now at OpenAI, argued open models are "inherently decelerationist" because they crush frontier margins, then walked back the idea that a crackdown was the best move.

The worry that carries weight is competitiveness, not spyware. Experts told TechCrunch that open weights running on US servers are unlikely to leak data home. The bigger concern from Hancock is ownership of innovation: US grad programs already build on Chinese open models, and half the papers students read come from Chinese institutions. PyTorch won because it was open. The same pull is happening now.

The deep cut

Do not treat this as a swap decision. Treat it as a floor-price discovery. The single most useful thing you can do this quarter is run one real workload, not a demo, on an open model like Kimi or Qwen and measure the full cost including the extra reasoning tokens. That number is your walk-away price. Bring it to your next vendor renewal. You do not need to migrate to win the discount. You need to prove you could. And keep a second model wired in for the tasks where US models refuse the job, since David Sacks has been citing US firms switching to Chinese models to close exactly those gaps.

Three questions for your team

• Which of our workloads are commodity intelligence, where any capable model gives the same answer, and which actually need the frontier? Sort them, because the commodity ones are where open models pay off first.
• What is our true per-task cost on an open model once we count reasoning and agent tokens, and how does that compare to our current contract? If we do not have that number, we cannot negotiate.
• If Commerce restricts Chinese open models, what breaks in our plan, and do we have a US open option like Nemotron wired in as a fallback?