GLM-5.2 Matches Frontier Models at Far Lower API Prices

no wonder OpenAI is talking about price cuts

with Blackwell they are literally printing money on inference

This is DeepSeek-V4-Pro inference cost on Blackwell

DeepSeek-V4-Pro costs $0.25 per million tokens output at 140 tokens/s. Now granted, this is very short context, but even if you 10x the price per token to go up to a million context it's still just $2.5.

but labs are serving it to you for $25 or $30 at 50 tokens/s

This is DeepSeek-V4-Pro inference cost on Blackwell

Now this is only 8k input tokens, but even if it was 10x the price per token it would still just be roughly a dollar per million tokens

It shows you how $25 or $30 per million tokens output is a bit ridiculous

This is DeepSeek-V4-Pro inference cost on Blackwell

Now this is only 8k input tokens, but even if it was 10x the price per token it would still just be roughly a dollar per million tokens

It shows you how $25 or $30 per million tokens output is a bit ridiculous

We could have DeepSeek-V4-Pro at 140 tokens/s for a dollar and they would still be at 50%+ margins

@JamieDay719679 @scaling01 There are providers on openrouter from all over the world.

@airesearch12 no

@lev_ey @scaling01

@scaling01 DeepSeek certainly looses money at this price. Believe me. This price is not sustainable for a model this size.

@scaling01 I should disclaim I am working at an inference provider (http://Chutes.ai) and thus have some insight on the internal cost structure of serving these models.

@scaling01 Are they making money or are they subsidized by the ccp? Genuine question

@scaling01 So much room to optimize spend without loss of performance on general tasks. At ¬◇ we’ve written about this before, and will announce some work soon to solve it. https://www.notdiamond.ai/blog/how-to-reduce-claude-code-costs-without-sacrificing-output-quality

GLM, there's a case to be made, but DeepSeek (when provided by themselves and not others, which is the lowest price and the one people quote) requires that they keep all of your outputs. They could very well be training on it, and there's no reason to believe they aren't. And that's a huge subsidy.

@Akumunokokoro @JamieDay719679 @scaling01 Deepseek is the exception. GLM-5.2 and other models cost around the same for any provider.

@scaling01 Get people in on subscriptions then hammer them on API pricing

@freddy_mertens @JamieDay719679 @scaling01 Freedom is defined by power, what you mean is "muh freedom"

@scaling01 This is why i don't believe openai and anthropic are losing that much money on subscriptions compared to the api price comparison. They are probably losing money but not that much

@freddy_mertens @JamieDay719679 @scaling01 spare me the 2014 reddit debate

@airesearch12 @scaling01 Dude american labs are just money printing monsters.

@TechnologyPat @RoliumGens @scaling01 but thats like it used to be with Amazon, at first it was profitable in the sense its making real big money but because it had expansion costs and R&D, then it ate revenue, but some day OpenAI, Anthropic and SpaceX will make enormous amounts of money and we will just watch.

@redcypress0 @JamieDay719679 @scaling01 At nowhere near price parity though. That $7.4B investment round that closed today will probably help to continue allowing them to undercut the competition.