AI
Three frontier AI models launched in 24 hours—all competing on price, not just smarts
When models converge on capability, pricing power shifts to whoever controls where AI work actually gets routed and executed.
The New Stack
Grok 4.6, Qwen 3.8-Max, and DeepSeek V4-Pro all rolled out within roughly a day of each other this week—and each was marketed first and foremost on cost savings rather than raw performance gains.
The shift is telling. Two years ago, frontier AI launches centered on what a model could do. Today, they're about how much you pay to use it. Every new benchmark exists to justify a cheaper bill.
Training, not scale, drives the improvements#
SpaceX's Grok 4.6 exemplifies the pattern. The company announced it with minimal fanfare: frontier intelligence and a "significant improvement over Grok 4.5 at the same price." Benchmark scores backed that up—Grok 4.6 hit 61 on the Artificial Analysis Intelligence Index, five points ahead of its predecessor and level with GPT-5.6 Sol. On performance leaderboards, it briefly took the top spot.
What's crucial: those gains didn't come from a larger model. Grok 4.6 runs on the same 1.5 trillion parameters as version 4.5, meaning engineers squeezed the improvements purely through post-training work. The API pricing stayed flat at $2 per million input tokens and $6 per million output tokens—because serving the same weights costs the same. The performance boost is, in effect, free.
Grok 4.6 undercuts Fable 5 Max by roughly 80% on input token costs and 88% on output tokens, making it what one analyst called "Pareto dominant."
Open weights set the price ceiling#
Two of the three launches this week came with freely downloadable model weights, and that detail reshapes the economics entirely. When a developer can grab a fully functional model from Hugging Face, a proprietary lab can no longer simply charge more for better performance. It must either release a better model at the old price or risk customers switching to the free alternative.
Alibaba's Qwen 3.8-Max initially pitched itself as open-source without releasing weights—which observers fairly called the API business dressed up for launch day. The weights arrived shortly after. DeepSeek followed with V4-Pro a day later, complete with native OpenAI API compatibility and a pricing structure that drops costs during off-peak hours.
The result: frontier-level intelligence now has variable pricing that shifts with time of day. The commodity pricing floor, once set by a single lab, is increasingly determined by whoever gives their work away.
