Everyone talks about how these models are getting closer and closer to OpenAI/Anthropic and how they're much cheaper at API pricing, which is great.
But then I compare it to my $200/month subscription - which I believe is how most developers are actually using these - and it actually looks way more expensive from my quick calculations.
Has anyone else calculated it? Is there any way to get Kimi K3 or Qwen 3.8 Max at a similar cost to what we're all paying by subscription for Claude or Codex?
If not, I'd push back and say these Chinese models are actually more expensive, for most developers.
These subscriptions look far more limited and possibly also more expensive than the equivalent Claude/Codex ones at similar prices.
It's tricky to do a proper calculation because everyone obscures what a token or a prompt costs in subscriptions and on top of that we have the complexity of how much different models think.
For example, GLM5.3-flash runs slightly slower than Qwen3.8-Next-Flash but uses fewer tokens to think and thus finishes tasks faster overall.
On top of that we have complications like Anthropic apparently being extremely misleading about what the 20x plan really means (it isn't 20x the weekly limit of the base plan). I would not be surprised if everyone's doing something manipulative like that. The constant changes to promotional periods, frequent limit resets, harness updates etc make this even more difficult.
I don't know about kimi, but my napkin-math shows that qwen and glm give you at least an order of magnitude more usage than anthropic, and that's before you factor in any off-peak discounts.
If by most developers you mean individuals then sure. But Anthropic and OpenAI have gotten rid of the subscriptions for most, if not all, corporate contracts. So companies looking to give their employees access are paying api rates.
for those who skip to the comments: Qwen released an updated checkpoint of this model today (0902) with significantly higher benchmark results that appear to place it much closer to Fable/Sol
Everyone talks about how these models are getting closer and closer to OpenAI/Anthropic and how they're much cheaper at API pricing, which is great.
But then I compare it to my $200/month subscription - which I believe is how most developers are actually using these - and it actually looks way more expensive from my quick calculations.
Has anyone else calculated it? Is there any way to get Kimi K3 or Qwen 3.8 Max at a similar cost to what we're all paying by subscription for Claude or Codex?
If not, I'd push back and say these Chinese models are actually more expensive, for most developers.
These subscriptions look far more limited and possibly also more expensive than the equivalent Claude/Codex ones at similar prices.
https://www.kimi.ai/resources/kimi-k3-pricing
https://www.alibabacloud.com/en/campaign/ai-landing-page-tok...
It's tricky to do a proper calculation because everyone obscures what a token or a prompt costs in subscriptions and on top of that we have the complexity of how much different models think.
For example, GLM5.3-flash runs slightly slower than Qwen3.8-Next-Flash but uses fewer tokens to think and thus finishes tasks faster overall.
On top of that we have complications like Anthropic apparently being extremely misleading about what the 20x plan really means (it isn't 20x the weekly limit of the base plan). I would not be surprised if everyone's doing something manipulative like that. The constant changes to promotional periods, frequent limit resets, harness updates etc make this even more difficult.
Are you comparing their subscription to yours? How do you know how many tokens you're getting for your subscription or for theirs?
I believe the answer is: You don't.
I don't know about kimi, but my napkin-math shows that qwen and glm give you at least an order of magnitude more usage than anthropic, and that's before you factor in any off-peak discounts.
Are you calculating based on API prices? Because I'm talking specifically about subscriptions here.
If by most developers you mean individuals then sure. But Anthropic and OpenAI have gotten rid of the subscriptions for most, if not all, corporate contracts. So companies looking to give their employees access are paying api rates.
> Anthropic and OpenAI have gotten rid of the subscriptions for most, if not all, corporate contracts.
Do you have any source for that?
for those who skip to the comments: Qwen released an updated checkpoint of this model today (0902) with significantly higher benchmark results that appear to place it much closer to Fable/Sol
But not open weight I guess, right?
No (current) mention of weights being released, so apparently not.
Announcement / benchmarks:
https://x.com/Alibaba_Qwen/status/2094968708288680276