It also depends on how many tokens it needs to burn through to accomplish something.
At this point, I always look at things like Artificial Analysis' total cost to run their tests. It'll take into consideration the cost of tokens, how many tokens it burns through, and how effectively it uses caching (and the price of that caching).
If a model "costs the same" but its reasoning ends up going through a ton more tokens, it doesn't really cost the same in real world usage.
Precisely. GLM 5.2 Thinking is pretty damn good - but it regularly does something nonsensical. Or even spits out what looks like a fragment of its memory cache. Or returns a bunch of Chinese.
I find myself having to resubmit a query very often...so it being a third of the cost of other AIs isn't really relevant.
GLM is actually quite expensive in actual practice because it's not very token efficient. I've yet to find a way to run it on a monthly sub reliably for cheaper than Codex.
Neuralwatt was cheap (but slow) but they cranked their price.
Ollama monthly sub is speedy but doesn't offer a lot of quota.
Right now unless you're paying by the token, there's no cost based reason to use the open weight models for daily coding work because the monthly coding plans from Anthropic and OpenAI are a better deal.
I think MiMo 2.5 seems to be better than DS4 Flash, at the same price. DS4F can write pretty advanced code but it way overthinks the simple stuff, its CoT is full of errors (immediately corrects itself), and yesterday I was shocked it reliably struggled to spot a misplaced % in a format string when reading a whole file.
Exactly, I design in glm 5.2, and then build with deepseek pro. I pay deepseek by token count, and honestly I added $20 and it just keeps going, hardly burning anything.
I'm on the Z.ai quarterly subscription plan (got in when the price was lower) and I was using it through opencode and it was like I'd only get maybe an hour of usage (if that, sometimes) before it would time out and say come back in 5 hours. Now I'm using it through their Zcode harness and I rarely hit that - they say they're giving 1.5x usage if you use it through Zcode, sometimes seems like even more than that.
> I've yet to find a way to run it on a monthly sub reliably for cheaper than Codex.
Matches my experience, I got their Pro subscription and while I enjoyed the model itself a lot and while their ZCode harness is also pretty nice, it gave me less tokens for similar amounts of money that Anthropic would give me on a subscription: https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model...
I'm yet to try out Kimi, but if their subscription were to be anywhere comparable to Anthropic/OpenAI, I might just switch over because competition is good.
DeepSeek V4 Pro is really affordable per-token but regularly kept making mistakes in the tasks I gave it. I mean I could at least afford the tokens to go over the work a 2nd, 3rd, 4th and 5th time and gradually fix most of the issues, but it was a very frustrating mode of work.
> Right now unless you're paying by the token, there's no cost based reason to use the open weight models for daily coding work because the monthly coding plans from Anthropic and OpenAI are a better deal.
Maybe. I am on a $20/month Anthropic subscription this month but I also use Claude Code frequently with Deepseek v4 flash and pro, GML5.2. For simple work Deepseek v4 flash is so nice because it is fast.
What you say is true however, the US hyper-scalers are still (desperately?) subsidizing subscriptions for market share to boost there valuations.
I really want to see AI inference costs approach zero, and I think I just need to wait a few years to see that.
For maths, it's also wrong most of the time. Generated stuff looks right, then let Opus audit and see the disaster.
DS4 is usable iff you have a way to test the generated stuff, and to convince yourself that its production is right. With successive review-fix rounds, it's obviously way more reliable too, but that can't compensate for it's lack of rigor when reasoning.
It's very smart but neither rigorous nor careful. And that is a direct result of its architecture.
> That said, Kimi is competing against GLM in my mind, and GLM 5.2 is less than 1/3 the price.
Having used GLM 5.2 extensively and K3 for a few hours now, these models are nowhere near each other. 5.2 is a great model, and I use it for a lot of things, but it's noticeably below Opus 4.8 or GPT-5.5 in real-world usage.
Tokenizers define the alphabet on which the language model is trained. I don't want people to get the impression it's a module which can be swapped out or modified on its own. Alphabet size is a design consideration related to correctly encoding the training data.
That's true, but it makes it difficult to compare pricing when it's based on tokens. Maybe we need a benchmark for price per a specific input, like enwiki8.
Yes, almost all work people share which seeks to measure the capabilities and differences of models needs to get more precise. We are clamoring to say something meaningful about these things.
It is kind of a shame we ended up comparing token pricing across models and providers when it doesn’t really make sense. Not sure what would be better though.
A better metric is price per byte. Most thinking traces, prompts, skills are in plain English, which is roughly 1 byte per character, assuming UTF-8 encoding (even code should not be much more either). As an aside, it is common to use bits-per-byte as a loss metric instead of the per token calculation, precisely because of the effect of different tokenizers.
It's going to vary dramatically based on which text you put in. Really it's hard to make one benchmark number that's relevant to all cases. But maybe we can make something a little more specific, like regular English text, code, the model's own thinking tokens, image inputs etc.
But even that isn't the whole story because the models can produce wildly amount of thinking output as well as regular output for a similar query. Sometimes you can take a cheap model and have it think a ton or an expensive model that thinks little and get similar results. But the number of tokens generated will be wildly different.
I’ve been struggling to understand the reason for the newer apparently less efficient Anthropic token encoding. If all inputs are less efficient in this encoding, why does it exist? Has Anthropic released any information that would convincingly show it was anything other than a stealth price hike? Please don’t respond if you are speculating.
I doubt you are going to get a response from an anthropic employee, but I think it is safe to assume they have swapped to a new tokenizer because it improves the performance of their models.
More tokens per same text length means more capacity to encode information. More information means model can potentially perform better.
They introduced it around the time the Mythos came so my speculation is that if you have more capable model at some level you may find the current information encoding not using its full potential.
We will see whether OpenAI also introduces new tokenizer when they come to Mythos-size models.
At this point, I always look at things like Artificial Analysis' total cost to run their tests. It'll take into consideration the cost of tokens, how many tokens it burns through, and how effectively it uses caching (and the price of that caching).
If a model "costs the same" but its reasoning ends up going through a ton more tokens, it doesn't really cost the same in real world usage.