The Chinese dropped two reasoning models: V3.2 and V3.2‑Speciale. 685B parameters. The basic V3.2 catches up with GPT‑5, Sonnet 4.5 and Gemini 3 Pro on math, tool use, and code. Speciale grabs gold medals at the International Math Olympiad
The price is fantastic: $0.28/0.42 per 1M input/output tokens (GPT‑5.1 $1/10, Gemini 3 Pro at $2/12, Sonnet 4.5 $3/$15). Numbers laugh!
Frontier performance + open weights + predatory pricing. So what now?
- US AI platforms: You’ll have to sell product and ecosystem, not brand mystique
- Startups: you can slash AI feature costs overnight. Run DeepSeek experiments. Move to prod. Self‑host if you love sovereignty
- Investors: the money migrates to verticals and last mile distribution [by the way, read my article about it]
#AI #DeepSeek #OpenSource #LLM #GPT5 #Gemini #SelfHost #ML #Startup #AIInfra
Discussion