AI Economics Are Changing Faster Than Your Stack

A month ago you could pick a model and mostly forget about it. That window closed. In the past two weeks, Google shipped Gemini 3.7 Flash at half the price of the model it replaced. xAI dropped Grok 4.6 at a fraction of frontier cost. Z.ai's GLM-5.3 hit the top coding benchmarks with a third of Kimi's size. Meta open-sourced Glimmer. And DeepSeek, the cheapest name in the game, warned users it is about to get more expensive. Let me catch you up.

The deep cut

• Model choice is a standing decision, not a bet. Gemini 3.7 Flash shipped three weeks after 3.6 at half the price.
• The harness beats the model at cutting cost. Writer's own research found harness tweaks cut spend 40% across models, more reliably than swapping models.
• Cheap is a phase, not a plan. DeepSeek warned users of a "significant" price hike as it builds out data centers.

The price floor keeps dropping

The numbers moved fast and in one direction. Gemini 3.7 Flash arrived at $0.75 per million input tokens, half what 3.6 cost, three weeks after 3.6 launched. Grok 4.6 landed at $2 in, $6 out, well under frontier peers, and practitioners called it the new default for coding. DeepSeek V4 Pro came in around 57 times cheaper than a top Anthropic model.

So the model you priced into your roadmap last quarter is probably overpriced now. That is good news if you keep checking. It is a slow leak if you set it and walked away.

The knockoff scores like the original

The gap between American and Chinese labs is not what your board thinks it is. GLM-5.3 matched or beat Claude and GPT on coding benchmarks with only 750 billion parameters, a third the size of its Chinese rival Kimi K3. Nathan Lambert argues the reason is boring: Z.ai ships in days while U.S. labs spend months in pre-release testing, and that lag flatters the competition on any given release day.

Watch the license terms too. Hugging Face found that of 178 Chinese releases over 20 billion parameters, 59% carry Apache 2.0 and none carry a non-commercial restriction. The big weights are free. The money comes from API business and hardware, not licensing.

The lever your engineers control

Here is the part your infra team already suspected. Writer shipped Palmyra X6, built on Z.ai's GLM-5.2, and paired it with harness upgrades that cut customer costs up to 50%. Their research found that tuning the harness, the plumbing around the model, cut costs an average of 40%, and did it more reliably than swapping models. "The harness is the one component whose efficiency multiplies across every model an organization runs," the researchers wrote.

CEO May Habib was blunt: "The enterprise is absolutely sick of chasing the next benchmark. They want flattening cost." She also thinks the labs have a reason to keep your token use high. Worth chewing on before your next renewal.

Cheap is renting, not owning

Do not build a business plan on today's price. DeepSeek warned users of a "significant" price hike as it funds a data center in Inner Mongolia. Charging under a dollar per million tokens was never going to last while the bills for compute keep climbing.

The hedge is openness. When you run open weights, a price hike from one vendor is an annoyance, not a crisis. Meta made this pitch loud with Glimmer, an open model anyone can download, though its stronger Muse Spark stays locked behind Meta's own APIs. The pattern is worth naming: the free model is the door, and the good stuff is often still on the other side of it.

Three questions for your team

• When did we last reprice our model stack against current options, and who owns rechecking it each quarter?
• If our main vendor doubled token prices tomorrow, how many days would it take us to switch, and have we tested it?
• Are we spending our optimization effort on picking models or on the harness, and which one is actually moving our bill?