# A Macro Look at the AI Model Landscape

> A race to the middle?

People are starting to notice that closed American models from OpenAI and Anthropic are getting less expensive in a way.  You can complete tasks for cheaper today than you could in the past.  There's a good reason for this: open weight models from company like DeepSeek, Alibaba, and Z.ai.  All 3 Chinese counterparts offer very strong performance for a fraction of the cost.  They are not Astra level, yet, but they offer more than enough for every day use up and up to advanced software engineering.  

The interesting part is trajectory of model development.  American companies have had access to the greatest and latest hardware, where the Chinese have been heavily restricted, so their paths have been very different.  American companies keep pushing to have better and better models, but on a massive scale; they would cost a fortune to run on your own hardware.  Models like Qwen make large-ish models, but focus a lot being able to run smaller variants on consumer hardware, because that's what they largely have access to.  That is all to say, innovation for capability has been pushed by American company and innovation to run models cost-effectively has been pushed by Chinese companies.

If you have noticed, the cycle for OpenAI and Anthropic seems to be: create massive and expensive models and eventually distill it down into something more affordable, with minimal quality loss.  For example, most of Fable's capabilities can now be found in Opus, for a fraction of the cost.  This is because they are losing market share to open weight models.  A very large percentage of American companies now heavily rely on companies other than OpenAI and Anthropic.  That's a problem.  The amount of money invested into creating models like Fable and Astra and the path to get to that point are off the charts.  We are talking trillions of dollars.  Companies like DeepSeek have only spent about $200,000,000 in contrast.  This poses a big risk for American AI companies and their investors.  Needless to say, creating the best of the best comes a high cost.  

To win market share American companies need to bring costs down, which is difficult with investors begging for returns.  And Chinese companies already figured out the cost, so they need to continue down the path of making more capable models.  Benchmarks might show GLM 5.3 can compete with Fable, but it's only on paper.  GLM is great and very capable, but they can't match the highest end markets.  That's okay though, they don't need to.  Tier 3 models (like Sonnet) are more than enough for day to day use, even for software engineers.

All of this is to say, I believe we are racing to the middle.  Both countries are capable of serving the tier 2, 3, and 4 markets, but China is winning on price.  American companies need to keep pace on their capabilities and continue to trickle down the benefits to the lower tiers, while reducing costs for end users.  The latter will be difficult, with all of the cash investments made.  Chinese companies started at the bottom and are moving their way up.  American companies started at the top, keep raising the ceiling, but also need to move their way down market.

Read online: https://davidoelfke.dev/blog/a-macro-look-at-the-ai-model-landscape-34f18Fo4b6AUFiXAnqMuQm