Grok is now the most popular model on OpenRouter(openrouter.ai)
openrouter.ai
Grok is now the most popular model on OpenRouter
https://openrouter.ai/rankings#market-share
49 comments
> most of Grok's traffic is through their free endpoint
https://openrouter.ai/x-ai
Code Fast 1: 100B
Grok 4 (free): 84.6B
https://openrouter.ai/x-ai
Code Fast 1: 100B
Grok 4 (free): 84.6B
I was hesitant to use Grok but have actually found it to be excellent and it has totally replaced ChatGPT for me now.
The biggest difference is that it isn’t sycophantic and will often tell me I am wrong when I am. This makes way more of a difference than I thought it would as I feel I can trust the results more (perhaps naively)
The biggest difference is that it isn’t sycophantic and will often tell me I am wrong when I am. This makes way more of a difference than I thought it would as I feel I can trust the results more (perhaps naively)
Hrm, I've never used it, but maybe I should. I've really only used claude and chatgpt, but it's annoying with how they agree with whatever you feed them.
I switched from OpenAI to Gemini a couple of months ago, and was impressed how it sticks to its guns if it thinks I'm wrong.
Interesting, will check it out
I would recommend trialing it for a month instead of ChatGPT.
I still use Claude code for coding
I still use Claude code for coding
Claude has been super annoying in the last few weeks, I ask it a question and it always immediately starts to write or update code instead of answering.
Same experience here. Codex has been a nice alternative when Claude Code is being dumb
Lol, I've just installed the grok cli and did exactly the same thing..
> I've started to merge scripta and scriptb into scriptc, how would you proceed?
> I've merged the scripts into scriptc
If I have to tell the AI with every prompt not to run ahead it stops being useful..
> I've started to merge scripta and scriptb into scriptc, how would you proceed?
> I've merged the scripts into scriptc
If I have to tell the AI with every prompt not to run ahead it stops being useful..
I don’t mind it because I can enable planning mode and it will (mostly) not modify things.
Claude will sometimes dump very exact file content into a plan.md when I ask it to write its plan down which is annoying.
Claude will sometimes dump very exact file content into a plan.md when I ask it to write its plan down which is annoying.
ChatGPT truly has no moat. I already stopped using it and cancelled my payment. The got left behind so quickly.
As things tend toward commodity (e.g. one item being replaceable with another with not much, or no, difference) the thing that matters most is brand.
So to the contrary, ChatGPT has the ONLY moat so far. 5.8B page visits vs 148m and growing much faster on an absolute basis. https://www.similarweb.com/website/chatgpt.com/vs/claude.ai
So to the contrary, ChatGPT has the ONLY moat so far. 5.8B page visits vs 148m and growing much faster on an absolute basis. https://www.similarweb.com/website/chatgpt.com/vs/claude.ai
The Grok 4 Fast and Grok Code Fast models have really impressed me. The only issue I've had is reaching a rate limit on the Grok 4 Fast model. Amazing pricing for highly capable models with good tool calling support.
They're definitely amazing for the price. I like that you can do quick back and forth with them. but they're not very smart. When I need something to actually analyze or write good code and not just refactor and move things around, they're not good for that.
Agreed. For harder tasks, I like to go to GPT 5 thinking mode, but I'm considering other options.
Some times I've had faster success with some of the larger Qwen3 models (480B and 235B variants). I like them in combination with the Repomix CLI to copy an entire project into context and get a response very quickly with some of the accelerated providers like Cerebras.
Some times I've had faster success with some of the larger Qwen3 models (480B and 235B variants). I like them in combination with the Repomix CLI to copy an entire project into context and get a response very quickly with some of the accelerated providers like Cerebras.
Now that Cursor has moved towards a credits system, grok code fast is making the plan last while still being reasonable in inference time. GPT 5 and GPT 5 Codex actually moves my "amount remaining" bar in realtime while being incredibly slow.
Little miss leading as it is sorted by number of output tokens, and I’ve heard Grok is rather verbose.
Sorting by revenue would definitely move Grok down. The top 2 Grok models listed are very cheap or even free right now.
Maybe it's verbose internally? When I run it, it's just extremely fast. So the verbosity doesn't seem to affect things.
It's very verbose even when prompted to remain brief, but it's still not a bad model and I use it a lot.
Those stacked bar charts are data visualization malpractice. They would tell a much clearer story (who's #1, what are the trends, when did one overtake another) as regular line charts.
50% of their tokens on grok code fast were consumed by kilocode users.
Kilocode has been giving away free grok code fast usage.
When the price comes up for 0 I will be curious to see if this trend holds.
Kilocode has been giving away free grok code fast usage.
When the price comes up for 0 I will be curious to see if this trend holds.
It's free on OpenRouter, but paid on the official xAI API, for the moment.
So, I'd imagine that is inflating the numbers just a bit lol.
So, I'd imagine that is inflating the numbers just a bit lol.
Grok 4 fast is a legit model. Their code models, including supernova still aren't smart enough. Claude and Codex are ahead. Its definitely fast, but who cares if you have to re-prompt it or it hits issues it can't fix.
Is open router used a lot? I just use claude code, so this is a misrepresentation of what is actually the most popular model right?
I am using it more than the disaster that is gpt5
I’m curious as to why, since inside GitHub Copilot GPT-5 has been stellar lately. Are you using it directly? (I assume that the prompting strategy inside Copilot is the reason why it’s so good right now).
It’s very slow. Extremely marginal improvement over o3. Sometimes falls short (after taking a lot longer). Lower the thinking amount and all marginal improvement evaporates.
Because it's free for now.
Workaccount2(2)
Zambyte(4)
And even then this wouldn't matter because most devs don't use OpenRouter (commissions!, incompatible API in edge cases, etc.), and most LLM enthusiasts who want to run free models do it on their own machine.
---
Edit: Most → Much