The unbearable cheapness of open weight models
摘要
文章对比了 DeepSeek V4 与 Anthropic、OpenAI 等 “前沿模型” 的 token 定价差距，认为价格可能相差接近 50 倍，并质疑这种成本结构是否会让高价厂商难以与低价模型竞争。作者提出 “open weight 模型更便宜” 的原因可能包括更广泛的社区部署与测试降低成本，或作为低价策略的竞争手段，同时也探讨了厂商如何在 “商品化模型” 上维持高价，例如通过品牌溢价与制造稀缺性（类似奢侈品逻辑）。 文章进一步提出一个担忧：未来是否可能通过政策或 “安全叙事” 限制 open weight 模型，以减少竞争压力。最后对比了不同组织的开放策略，包括 Google 的 Gemma、Meta 的 Llama，以及 Anthropic 尚未发布 open weight 模型，并提到更 “真正开源 AI” 的方向，如 Allen Institute for AI 的 OLMo 项目以及美国 NSF 与 Nvidia 合作推动更开放的 AI 研发。
荐读理由
无任何寻求结果可被兑现
原文
Today I was setting up Hermes to see how it does with web research. I chose DeepSeek V4 because I know it is cheap, but seeing it’s pricing next to Anthropic and OpenAI ‘frontier’ models is crazy. Nearly a 50x price increase based on tokens alone, let alone how much pondering any of their models might fall into (using more tokens for the same task).
What worries me about this is that Anthropic and OpenAI seem to have backed themselves into a corner of high costs. Can they reasonably decrease their prices by 20-50x to compete with DeepSeek or Xiaomi’s Mimo?
Open Weight vs Low Cost
Are these models cheap because they are open weight and having hundreds or people stress test running them on different hardware helped to lower the cost? Or is it that they are being provided as loss leaders to drive the prices down?
How do you keep prices high for commodity products?
You manufacture scarcity. You sell luxury and premium branding. This is what OpenAI and Anthropic seem to be doing by gating ‘frontier’ model usage behind higher walls.
This is how luxury brands have sold cars and hand bags forever. They are clubs and status symbols for the rich and not meant to be widely distributed.
Will Anthropic & OpenAI lean on China fears to push bans on open weight models?
This has been my fear for a few months now and each week that goes by seems to support this. How do you manufacture scarcity? One easy way is to fear monger and get the government to help restrict access to competition.
Why not compete?
The US used to be such a champion of open source, and I would hope that serious open source competition can come out of the US to prove that open weight and open source models are ultimately the future.
Google Gemma 4 was released in April 2026
Meta had llama which hasn’t had a release
OpenAI last released open weight gpt models in 2025
Anthropic to my knowledge has never released any open weight model
True Open Source vs Open Weight
I think the leap frog scenario for Open Source will be the true Open Source models where the data pipeline for training is also open sourced.
https://allenai.org/olmo -> You can download these models now and they’re seeing increasing popularity. That being said, they are a bit out of date, with data cutoffs in Dec 2024
Looking to the future, the US NSF partnered with Nvidia to enable Allen AI to develop a true fully open AI: https://www.nsf.gov/news/nsf-nvidia-partnership-enables-ai2-develop-fully-open-ai
Bonus:
Curious to dig more into Claude / ChatGPT tech stacks? Check out the tools they used to build their iOS and Android apps:
Claude Android ChatGPT Android
You can navigate to SDKs to view even more detailed breakdowns of specific parts as well as unmapped SDK paths.
这条对你有帮助吗？