Y'all, the AI world did not take it easy this week. Prices are falling off a cliff, everybody and their cousin wants to make chips now, and the open weight crowd out of China just dropped another bomb. Let's get into it.
DeepSeek makes coding AI dirt cheap
DeepSeek dropped V4 Flash, a new coding model that gets close to Claude Opus level performance while costing about 99 percent less to run. Ninety nine percent. That's not a typo, that's a fire sale.
This lands right in the middle of a broader price war. OpenAI cut GPT 5.6 Luna pricing hard, Google rolled out cheaper Gemini Flash models, Meta priced Muse Spark 1.1 aggressive, and SpaceXAI's Grok 4.5 is playing the same game. Everybody's racing to the bottom on price at the same time the models keep getting better.
My take: this is great news if you build software and terrible news if your business model depended on charging a premium for AI output. If you're a small shop paying by the token for coding help, go check what DeepSeek costs you right now. The savings are real. Just don't be shocked when the free tier gets nerfed in six months once everybody's hooked.
Qwen3.8-Max drops and it's open weight
Alibaba put out Qwen3.8-Max and is calling it their most capable model yet, claiming it goes toe to toe with the best stuff from Anthropic and OpenAI. They're planning to release the weights publicly too.
Whether or not the benchmarks hold up under real use, the bigger story is that another serious lab is putting a frontier grade model out there for anybody to download and run themselves. That changes the calculus for a lot of companies who don't want to be locked into a single API provider.
My take: I don't fully trust self reported benchmarks from anybody, Chinese or American. But open weights at this capability level is a real gift if you've got the hardware to run it. Worth kicking the tires once it's out and folks have run their own tests.
Everybody's suddenly a chip company
Anthropic confirmed it's building an in-house chip design team. TSMC bumped its total US investment plan up to $265 billion and raised its 2026 capex budget from $60 billion to $64 billion. AMD is buying a Canadian startup called Taalas that hard codes model weights directly into inference chips instead of running them on general purpose hardware.
Google also reorganized its whole AI org this week, which after ten years running the same setup tells you something's shifting under the hood at the biggest labs.
My take: the AI labs figured out that whoever controls the silicon controls the margins. Software companies turning into chip companies is the tell that this whole industry is settling in for the long haul, not chasing a quick hype cycle. Keep an eye on who actually ships hardware versus who just announces a team.