🔮 Kimi K3 surprise & AI economics; the solar paradox; AI's right to learn, cancer vaccine & junior jobs++
Exponential View Azeem Azhar ● Covered by 7 sources
Moonshot AI just dropped Kimi K3, an open model that reportedly beats Claude Opus 4.8 and matches GPT-5.6 on some benchmarks. Weird part: it's not cheap, and that could shake up who profits from AI.
Based on reporting by Exponential View, Azeem Azhar — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Moonshot AI's new open-weight model, Kimi K3, is being talked about as a bigger jolt to the US AI industry than the original DeepSeek release. The early consensus, hedged with the usual benchmark caveats, is that K3 outperforms Claude Opus 4.8 and lands roughly on par with Claude's Fable and OpenAI's GPT 5.6 on certain tasks. Nobody's pretending benchmarks capture the real world perfectly, but the chatter is loud enough to notice.
What makes this release strange is the pricing. Open models are supposed to be the cheap option, the DeepSeek playbook of undercutting Western labs on cost. K3 doesn't follow that script. According to Artificial Analysis, it costs about the same as OpenAI's GPT 5.6 Sol overall, and it's roughly 24 times pricier than DeepSeek V4 Pro. Yet on a pure per-token basis, it comes in at half the price of GPT 5.6 Sol. So it's cheaper in one sense and startlingly expensive in another, far from the blunt discount Chinese models usually offer.
For companies and governments without deep pockets, this is still probably good news. It adds real pressure to the fat inference margins OpenAI and Anthropic have been enjoying, margins they've kept because their models sit at the frontier. But pressure isn't the same as losing customers. Enterprises don't just chase the lowest price tag; they pay for security, support, and the polished tooling built around the big labs' models. The harnesses OpenAI and Anthropic have constructed remain a real advantage that a strong open model doesn't erase overnight.
And there's a bigger structural story here. Token demand tends to be elastic, meaning cheaper intelligence increases usage rather than just shrinking spend. Hyperscalers and neoclouds will end up hosting these capable open models too, which only deepens demand for compute, chips, and memory. That shifts more of the revenue pool away from the model layer and toward the infrastructure sitting beneath it, reinforcing the case for continued build-out rather than undermining it. The one dampener worth remembering: even cheap, capable intelligence still has to survive a company's slow internal approval process before it actually gets used.
My take — AI-written commentary, not fact-checked reporting
Everyone keeps waiting for a cheap open model to torch the big labs' margins, and K3 is a reminder that it's not that simple: a model can be genuinely competitive and still not be the bargain people expect. The real winners here are the compute and chip suppliers sitting quietly underneath this whole fight, collecting the toll no matter which lab's logo is on the chatbot. Price wars make headlines; infrastructure payback quietly cashes the checks.
Read more about this at: Exponential View