# This Week in Tech: The AI Arms Race Heats Up
If it feels like the AI industry is moving faster than ever, that's because it is. This week made it clear the race has moved past "who has the smartest model" and into something bigger: a fight over chips, data centers, energy, and the money to fund it all.
## The money is getting enormous
Big Tech's combined AI purchase commitments are now approaching a staggering $1.5 trillion. That's not R&D budgets — that's actual contracted spend on chips, data centers, and infrastructure. At the same time, chipmaker SMIC has been raising prices as its factories run close to full capacity, a sign that demand for AI silicon still isn't slowing down, even after years of expansion.
## Prices are moving in opposite directions
Here's the twist: while the infrastructure race gets more expensive, the actual cost of *using* AI is dropping fast for some players. OpenAI and Anthropic have both been cutting prices on their models, betting that cheaper access means more usage and more developers building on top of them. DeepSeek, interestingly, has gone the other way and raised its prices. It's an early sign that the "AI price war" isn't uniform — some companies are racing to the bottom, others are betting people will pay more for reliability or performance.
## Google's Gemini just got cheaper and smarter
Google quietly shipped an update to Gemini 3.6 Flash this week, just three weeks after the previous version. The update brings real, measurable coding improvements — jumping from 34.4% to 43.6% on the FrontierCode 1.1 benchmark, and from 49% to 65.3% on DeepSWE — along with a 1-million-token context window. Pricing was cut roughly in half, to $0.75 per million input tokens and $3.75 per million output tokens, at least through the end of the year. That's clearly aimed at developers building multi-step AI agents, not just casual chatbot users.
## xAI enters the ring with Grok 4.6
Not to be left out, SpaceXAI (the company formerly known as xAI) launched Grok 4.6 this week, a model built specifically for long-running agent tasks, coding, and multi-step workflows. Early benchmarks put it roughly level with GPT-5.6 on the Artificial Analysis Intelligence Index. Pricing sits at $2 per million input tokens and $6 per million output tokens, with a faster (and pricier) variant also available.
## AI is starting to do real research
Perhaps the most eyebrow-raising story of the month: an unreleased OpenAI model reportedly helped disprove a decades-old open problem in combinatorial geometry tied to mathematician Paul Erdős — the kind of result that would normally take human mathematicians years, if it happens at all. It's an early, genuinely interesting signal that AI is starting to move from "doing tasks" to "contributing original research," which is a very different milestone than another chatbot benchmark win.
## My take
The infrastructure spending numbers are almost too big to process, but the part worth watching closer to home is the price war. If model costs keep dropping the way Gemini's did this week, expect AI features to show up in more everyday apps, cheaper and faster than most people expected even six months ago.
*What's caught your eye in AI this week? Drop a comment below.*
