Top
Best
New

Posted by ianberdin 5 hours ago

The real prices of frontier models(playcode.io)
148 points | 75 comments
SwellJoe 5 hours ago|
The fact that OpenAI documents theirs is already a big improvement over Anthropic. But, also, the OpenAI tokenizer got more efficient when they last updated it, rather than less. https://mdstudio.app/o200k-base-tokenizer
alansaber 4 hours ago||
Interesting. New models are estimated at ~5T params, so 45,000x increase over BERT base (110m). But vocab size of 200k, so only an increase of 7x over BERT base (30k).
satvikpendem 1 hour ago|||
Interesting that Grok 4.5 nears or exceeds OpenAI and Anthropic models in some benchmarks at only 1.5 trillion parameters per their announcement post.
vlian2088 21 minutes ago||||
BERT is not an LLM, it's an encoder-only model, i.e. it doesn't generate new text. the first somewhat useful publicly accessible LLM was GPT-3 with 175B params, and it was also the last frontier model whose parameter count was disclosed.
kamranjon 3 hours ago|||
Where do these estimates come from?
cyanydeez 3 hours ago||
a dark anal cavity somewhere next to marketing.
ianberdin 4 hours ago||
[dead]
Tiberium 4 hours ago||
Yeah, Anthropic's current tokenizer in Sonnet 5/Opus 4.8/Fable 5 is much worse than OpenAI's. Also, OpenAI has been using their current o200k_base from the day GPT-4o came out over two years ago. Just a few of my own tests:

- A ~2000-2002 legacy C++ game codebase at about ~90kloc: GPT 1.12M, Claude 2.2M

- A ~30kloc TypeScript codebase: GPT 260K, Claude 437K

In the end, GPT's current tokenizer is ~1.6x-2x better than Claude's current one, depending on your data. And you can check for free for both, for OpenAI just use the open-source libraries, for Anthropic - you have to use their count_tokens endpoint as they don't publish the tokenizer, but the endpoint is free (and allows requests over 1M tokens as well).

onlyrealcuzzo 4 hours ago||
Interesting... Naively I'd assume you'd have a pretty unfair advantage on quality if you have materially more information dense tokens.

That doesn't really appear to be the case as GPT and Anthropic models appear evenly matched despite Anthropic encoding the same text into almost ~2x the tokens...

I'd also - naively - assume this would make training their models more expensive. Though inference now dominates, and they'd probably rather have more tokens than less (to charge you for them at future 80% margins).

recursivecaveat 2 hours ago|||
If a given paragraph gets encoded into twice as many tokens, that means the model gets twice as many matmuls to process it. The amount of compute thrown at the problem is increased (everything else constant), which may improve the quality of the result. This is believed to be one of the reasons that 'thinking' tokens improve quality. For long tasks it will lead to more context compactions though which will harm the quality to some degree as well.
MPSimmons 55 minutes ago||
It would be nice if inference could somehow perform token generation using "contractions" of "fluffy" tokens, where combining those tokens doesn't decrease nuance but provides additional efficiency. That may already be happening - I haven't looked at the most modern methods of inference in a long long time.
not-a-llm 4 hours ago|||
more dense tokens means more stuff to fit into the embedding space which is per token, so more work to disentangle later
not-a-llm 4 hours ago||
you use the wrong word

the Anthropic tokenizer is not worse, its more expensive/verbose

nullsanity 4 hours ago||
So, worse? Because we benchmark off token use when talking about token use, and everyone else understood that.
marcosdumay 36 minutes ago|||
The most important feature of a tokenizer is dividing the inputs into independent values that the neural network can work with. It's not the size.
jascha_eng 2 hours ago||||
I mean it might lead to better performance on the model side. So the tokenizer is better but more expensive.
otikik 4 hours ago|||
It’s better for them
wgd 24 minutes ago||
> DeepSeek and GLM are left out of the tables entirely: we only have rough characters-divided-by-four estimates for them, not real tokenizer counts, and this post is about measured numbers.

lolwut. The open-weight models are inscrutable black boxes for which we can't possibly get real token counts? Typical lazy clanker, BSing their way out of doing the whole job.

foota 3 hours ago||
It's a bit unrelated, but I've been wondering if LLM providers are using cache read costs to preserve the illusion of constant output token prices. In reality, the longer your context the more expensive output tokens are, but anthropic and openai both have flat output token pricing.

In practice though as a result of cache reads over multiple turns you will end up paying quadratic pricing anyway.

ianberdin 4 hours ago||
We've switched the default model in playcode.io among Opus 4.8, Opus 4.6, Sonnet 4.6, and Sonnet 5. I must admit, Opus 4.8 is quite expensive, and the costs accumulate quickly. Opus 4.6 is about 50% cheaper, while Sonnet 5 is significantly more affordable. According to the data, Sonnet 5 is about 2-3 times cheaper. Fable 5 is unaffordable at all...

Today, I tested Sol 5.6 on various tasks. It performs similarly to Opus 4.8 but is still noticeably more expensive than Sonnet 5. Although Sonnet 5 isn't the top model, it's quite effective for creating typical websites for small and medium businesses. However, they will increase the price starting September 1, as their free offer is ending.

I'm also actively testing Grok 4.5. There's something promising about it. The design is mediocre, in my opinion, but it operates quickly and reliably without any deadloops. Usually, Grok models would fail or loop, but this one is stable.

Overall, I really want a benchmark based on real tasks.

jnwatson 4 hours ago||
The real elephant in the room is pricing for KV cache writes and reads. That makes all the difference for tasks with large context.
petercooper 3 hours ago||
Cache writes/reads are the majority of the picture for agentic development. When you see these people saying a coding run "used 200 million tokens" or whatever, most of that is cache reads, so it should be the headline price IMO (and is one reason why DeepSeek API is so striking with its minuscule cache pricing).
ianberdin 4 hours ago||
[dead]
luciana1u 1 hour ago||
Anthropic's tokenizer being 2x less efficient means they're essentially charging you for whitespace. premium whitespace, mind you — each space character gets its own attention head.
MPSimmons 58 minutes ago|
I don't know if it's fair to say that a tokenizer is being less efficient if it generates more tokens per text. I think it's more fair to say the tokenizer is more nuanced. The question is whether the additional nuance permits better model output, which could justify the additional token cost in inference.
ianberdin 25 minutes ago||
@dang, why is this flagged?
ricardobeat 4 hours ago||
I have reduced usage of Fable and Sonnet 5 to a minimum. Fable in particular is amazing at creative tasks, but not worth the cost for almost everything else. I can have Opus 4.6/4.7 running non-stop without hitting quota, vs maybe 20 minutes of Fable usage.
arjie 3 hours ago||
After getting used to Opus 4.8, Sonnet 5 is nigh unusable for coding. I much prefer Opus 4.8 + DS V4F for routing. Sonnet 5 is just not useful in the price/performance anywhere. I still get some use of Fable for planning because it can comprehend agent-built codebases (which have repetition of local patterns to a grand degree).
colechristensen 4 hours ago|||
Fable can solve the problems Opus couldn't. BUT most of the time I'm not having those kinds of problems.

I wouldn't say I'm doing anything groundbreaking but definitely at times obscure and that's when Fable has been able to dig me out of the rut. (the alternative I was actually following was reading textbooks myself to understand the domain better)

lolinder 3 hours ago||
The weird twist here is that I've found that there are times where Fable is actually worse than Opus at a straightforward task, not just more expensive. It'll launch off into its own little world for an extended thinking, then start editing files, and I'll already have spent $5 in tokens before I see enough response to know that it's on the wrong track.

Opus's verbosity is actually a boon sometimes for catching false starts early.

ianberdin 4 hours ago||
Well, I both agree and disagree with you.

On one hand, the price is just astronomical for Fable, well, not exactly astronomical, but I would say unaffordable. That is to say, so expensive that it is impossible to use.

But on the other hand, Fable is simply incomparable to anything else. I mean, it is just amazing. There is nothing even close to being equal to it.

gervwyk 3 hours ago||
I have to agree on fable being in another league. I’ve given it my most challenging problems and almost always comes back with a functional, solution 2-5 prompts away from a finished pr. Literally smashed our backlog - very impressed. What i found most efficient is to add “use sonnet agents for research” gets you really far, and on large not so novel tasks “use opus for tasks” by adding this it spins up many agents, works for 2+ hours in a usage window and completes A LOT of work.
iLoveOncall 2 hours ago|
Very unpalatable completely LLM-written article, but on top of that a lot of the fundations and conclusions are completely wrong, the main one being this one:

> You will see people claim Claude uses 2x to 4x the tokens of GPT. Our measurements do not support that, and overstating it would undercut the real point.

It's not because a single prompt represents only 1.7x the number of tokens that a model doesn't use 4x as many tokens as another, when running as an agent. This doesn't take at all the number of tokens of the output into account, and the number of tokens of the potential tool calls from this output, which directly feeds back to input tokens.

The article also has a very small test set (16 documents), all of very small length (15K tokens at most, when models go up to 1M in context and agents routinely exceed this and have to summarize).

Complete garbage article.

ac29 1 hour ago|
> You will see people claim Claude uses 2x to 4x the tokens of GPT.

My guess is something to this effect was in the prompt and the LLM made a point of correcting it

More comments...