It seems to be extremely economical - 4x better reasoning efficiency compared to Opus while being priced at $2/$6. For comparison, GPT 5.4 is $2.5/$15, GPT 5.5/5.6 are $5/$30, Opus 4.8 is $5/$25, Fable is $10/$50.
> Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools. This dataset lets the model learn both from existing software as well as developer-agent interactions, capturing how developers work and how agents interact with their environments.
This is what the big money was for. Cursor is the first big player that had real-world data from real-world projects, before cc / codex were a thing.
> We used reinforcement learning on difficult problems in realistic environments spanning both software engineering and broader knowledge work. These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results.
> Many of these problems had to be designed to be difficult enough that even frontier models fail at them. As models improve, existing tasks stop teaching them anything new, and problems that once required extensive reasoning become routine.
> We developed a distributed agent system to construct these environments at scale. Engineers specify a problem and how a solution is verified, and large groups of agents construct, test, and refine each environment.
This is where scale comes in. You use the previous gen model to prepare datasets for the next model iteration. The better the models, the better the data, the better the next models. (they also have a comparison with their composer2.5 training run, for people still thinking chinese models are "close to SotA"...)
Reports of xAIs demise (after giving a lot of compute to Anthropic) were slightly exaggerated, it seems.
> Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs
NitpickLawyer
First impressions:
- Very fast, easily beats GPT 5.5/Opus 4.8/GLM 5.2 because of higher t/s (around 90?) and very high token efficiency
- Very good price, no contest vs GPT and Opus which are very overpriced if you pay API costs, and probably cheaper than GLM 5.2 when you take into account the token efficiency.
- Will take quite a while to get a feel for how smart it is, but it's definitely good, I'd say in the same tier as opus, occupying the lower end of that tier together with GLM 5.2.
redox99
Of the 3 models I tried, Grok did the best at making an iOS app I wanted for personal use (a bike computer with specific qualities). (Claude just gave up and did an HTML/CSS implementation but I insisted on native SwiftUI+Metal.) Grok definitely fumbles sometimes, but I have been surprised what it CAN intuit versus me having to micromanage it.
(I am not an iOS developer, so getting something specific that I needed in a few hours/days was really helpful instead of spending months/years learning the language, APIs, etc.)
(I am absolutely not "vibe-coding" Caddy btw, just tinkering with it for personal projects.)
> Grok 4.5 and Composer 2.5 are two different model weight classes, and we're excited to support both sizes and weights. Composer 2.5 will remain offered, and we will release new models of this size going forward.
aarvin_roshin
> Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools.
This -- training on work done on hard, real-world tasks -- seems to be how most frontier models are making capability gains these days. In fact people make decent money doing that for data companies like Mercor. However it's also striking that Cursor managed to gather so much of such data.
Turns out Cursor will train on everything you do unless you opt-out, even if you're already paying for it with cash! Are that many people really not opting out?
This is why it seems like a significant concern to me: It's very clear that typical, run-of-the-mill coding has been completely commoditized, so the primary value remaining is either in novel use-cases and applications, or novel technical solutions to hard problems.
Presumably the value for novel use-cases could be captured by building a business around it via the usual moats (distribution, relationships, network effects, first mover advantage, etc.) so the code and techniques do not matter as much.
However novel technical solutions, which are already hard to monetize without building a whole damn business around it, could at least be capitalized on by simply being able to claim credit for it. I'd at least like the option of being "paid in exposure" if I'm not getting paid in cash. But having them "leaked" unwittingly via the training corpus to whosoever happens to prompt the model with the same problem removes even that option.
I know people have been calling out this risk forever, and I don't use any tool that I can't opt-out of training completely, but the scale at which this is happening -- on an ongoing basis, mind you, after training on the data of the whole world, and that too after paying for the product -- is surprising. I'm bullish on the technology but we really should be way more careful handing these AI companies even more of our intellectual crown jewels.
keeda
With each release from the the other major labs, it becomes harder for Google to tell a compelling story about Gemini 3.5.
Edit: Gemini 3.5 Pro. Expectations grow with each day it is not released.
Great model, very nice. Opus class performance at Haiku level pricing (or cheaper with the token efficiency). This seems like a GLM-5.2 killer and this is what Sonnet 5 should have been.
This is a model I could really see used inside applications, where Opus or Sonnet or GPT-5.5 are too expensive.
I would really like to see a strong Deepseek v4-Flash competitor, which ideally is something like Sonnet 4.6 performance at <$0.30 per token. This is missing from main US labs.
mchusma
Tried this for a legal use case and it was excellent, comparable to Opus in quality but much faster. AI is miles behind in law compared to coding: the output was similar to a law student intern. But coherent and directionally correct and beats starting from a blank sheet of paper. Impressed.
comments (10)
And by benchmarks (unless they gamed them), seems to be at around Opus 4.7 level, which is what Elon mentioned in https://x.com/elonmusk/status/2074911038286295049.
I guess the Cursor data was very useful.
Tiberium
> Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools. This dataset lets the model learn both from existing software as well as developer-agent interactions, capturing how developers work and how agents interact with their environments.
This is what the big money was for. Cursor is the first big player that had real-world data from real-world projects, before cc / codex were a thing.
> We used reinforcement learning on difficult problems in realistic environments spanning both software engineering and broader knowledge work. These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results.
> Many of these problems had to be designed to be difficult enough that even frontier models fail at them. As models improve, existing tasks stop teaching them anything new, and problems that once required extensive reasoning become routine.
> We developed a distributed agent system to construct these environments at scale. Engineers specify a problem and how a solution is verified, and large groups of agents construct, test, and refine each environment.
This is where scale comes in. You use the previous gen model to prepare datasets for the next model iteration. The better the models, the better the data, the better the next models. (they also have a comparison with their composer2.5 training run, for people still thinking chinese models are "close to SotA"...)
Reports of xAIs demise (after giving a lot of compute to Anthropic) were slightly exaggerated, it seems.
> Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs
NitpickLawyer
- Very fast, easily beats GPT 5.5/Opus 4.8/GLM 5.2 because of higher t/s (around 90?) and very high token efficiency
- Very good price, no contest vs GPT and Opus which are very overpriced if you pay API costs, and probably cheaper than GLM 5.2 when you take into account the token efficiency.
- Will take quite a while to get a feel for how smart it is, but it's definitely good, I'd say in the same tier as opus, occupying the lower end of that tier together with GLM 5.2.
redox99
(I am not an iOS developer, so getting something specific that I needed in a few hours/days was really helpful instead of spending months/years learning the language, APIs, etc.) (I am absolutely not "vibe-coding" Caddy btw, just tinkering with it for personal projects.)
mholt
Notably:
> Grok 4.5 and Composer 2.5 are two different model weight classes, and we're excited to support both sizes and weights. Composer 2.5 will remain offered, and we will release new models of this size going forward.
aarvin_roshin
This -- training on work done on hard, real-world tasks -- seems to be how most frontier models are making capability gains these days. In fact people make decent money doing that for data companies like Mercor. However it's also striking that Cursor managed to gather so much of such data.
Turns out Cursor will train on everything you do unless you opt-out, even if you're already paying for it with cash! Are that many people really not opting out?
This is why it seems like a significant concern to me: It's very clear that typical, run-of-the-mill coding has been completely commoditized, so the primary value remaining is either in novel use-cases and applications, or novel technical solutions to hard problems.
Presumably the value for novel use-cases could be captured by building a business around it via the usual moats (distribution, relationships, network effects, first mover advantage, etc.) so the code and techniques do not matter as much.
However novel technical solutions, which are already hard to monetize without building a whole damn business around it, could at least be capitalized on by simply being able to claim credit for it. I'd at least like the option of being "paid in exposure" if I'm not getting paid in cash. But having them "leaked" unwittingly via the training corpus to whosoever happens to prompt the model with the same problem removes even that option.
I know people have been calling out this risk forever, and I don't use any tool that I can't opt-out of training completely, but the scale at which this is happening -- on an ongoing basis, mind you, after training on the data of the whole world, and that too after paying for the product -- is surprising. I'm bullish on the technology but we really should be way more careful handing these AI companies even more of our intellectual crown jewels.
keeda
Edit: Gemini 3.5 Pro. Expectations grow with each day it is not released.
xnx
simonw
This is a model I could really see used inside applications, where Opus or Sonnet or GPT-5.5 are too expensive.
I would really like to see a strong Deepseek v4-Flash competitor, which ideally is something like Sonnet 4.6 performance at <$0.30 per token. This is missing from main US labs.
mchusma
rayiner