Kimi K3 and Qwen 3.8 Match Anthropic's Fable 5, Threatening Model-Only Labs' Economics
- Kimi K3 (Moonshot Labs) and Qwen 3.8 (Alibaba) launched this week claiming performance close to Anthropic's Fable 5, with open weights to be released publicly in coming weeks
- Frontier labs split into three infrastructure strategies: leasing data centers and power (Anthropic, Knowledge Atlas, Moonshot Labs), building owned data centers but leasing power (Meta, Alibaba), or owning both data centers and power generation (SpaceX)
- Labs that lease everything turn inference costs into variable costs that scale with revenue, while labs owning infrastructure convert those costs into fixed costs, letting margin grow with usage
- Figure 1 in the article shows Fable 5 costs nearly 3x as much per completed task as OpenAI or open models, raising doubt about whether users will keep paying the premium as competitors close the performance gap
- Anthropic is flagged as uniquely exposed because it leans on a regulatory strategy and a race toward recursive self-improvement rather than owning infrastructure, leaving it squeezed by both cheaper open models and specialized vertical labs
Hacker News 의견들
How is OpenAI not a model-only provider just like Anthropic? Seems like they're just as vulnerable to the same defensibility problem.
They're diversifying into consumer hardware and are further along owning data centers. If that hardware bet pays off, it puts them in a really different spot than the other labs.
Also they've been building their own inference chips, not just consumer gadgets. Having a good model means nothing if you can't serve it fast enough, look at Anthropic six months ago, or Moonshot suspending coding plan subs because they couldn't meet demand.
ChatGPT is basically synonymous with LLMs for non-technical users now. All that accumulated user history makes switching costs real, people aren't jumping ship because some other model is 10% cheaper.
Anthropic gets squeezed from both sides: open models eat the 80% of use cases that don't need frontier capability, and vertical specific labs (bio, finance, math) beat them on cost and speed for the high value stuff. Their AGI hail mary could also get wrecked overnight by a Taiwan blockade or Nvidia export cutoff.
None of my use cases need frontier capability but I still pay $200/month because the time saved is worth way more than that to me. If I had to pay raw API rates though, I honestly don't know what I'd do.
I think the risk here is overstated, people will pay a real premium for even slightly better models, and honestly a huge chunk of the value isn't the model itself but the harness around it like Claude Code, which open source alternatives like OpenCode just can't match yet.
That $200 is heavily subsidized though, real inference cost at max plan usage could be closer to $10k.
Easy for you to say from a rich country. Where I live a monthly wage is around $200, so no amount of theoretical 'value' matters if paying for it means you can't cover rent.
That margin people are willing to pay for a better model is shrinking fast. Fable was worth it a week ago, but with K3 closing the gap, paying 5x more for a 2% real world improvement isn't worth it anymore for me.
Excel has real network effects, if everyone else uses it you're forced to too, and the UI stickiness keeps people from switching. Tokens don't have that, if I use Fable and you use something else, there's zero friction between us.