I appreciate the followup. I’ll look into howw to turn those things off
An anecdote: I’ve been checking Grok’s more complex work with Opus and viceversa this week. They both still seem to have blindspots the other catches, so I guess they work nicely together?
Opus will find bugs in Grok’s changes even after a Grok + Bugbot review says no bugs. But so will Grok find bugs in Opus’ code even after Opus gives the green light!
This makes it a little hard to pick just one (which might just end up being how AI dev goes).
Create subagents Verifier A and Verifier B; have the Agent call them in parallel at the end of its work and receive an audit from both.
Yea it does things like that and the “Apply” button was already there so it could have just cloned the class.
I think its awesome good price good quality, beats most frontier models as daily driver. Looking forward to Grok 4.7!
Here is my review after literally using about a billion tokens of it.
Intelligence is good. Not top level good, sol 5.6 and opus beat it clearly, but not THAT far behind.
Creativity is kind of meh. Every now and then it surprises, but often its pretty bland. Opus is much better here. When leaving it alone to do something creative without a lot of guardrails, the results tend to be VERY basic.
Prompt adherence is good. Not great, but solid.
Thoroughness is its weakest point compared to frontier models. It gives a result and then normally does not give a s… if its right or wrong. Not automatically self checking. This can be improved with prompting (check your results) but it really should be the default.
Grok 4.6 is short sighted. Its vision capabilities are there and helpfull, but it sometimes does not notice glaring defects that other models see immediately.
TLDR: Its a great model for its price. Usefull as a daily driver, but it needs help for a lot of tasks. Basically for anything really complicated, you still want the top models.
-
How does it handle long-running, multistep work?
Very well. The only problem is that Cursor only gives it 256k of context, not 500k like the model itself supports. Summarization is sadly noticeable. It forgets things from the conversation. Otherwise very good though. -
How is it on visual and interactive projects?
N/A -
How does it compare to Grok 4.5, Composer, and other frontier models in your day-to-day?
I switched to Grok 4.6 from Fable/Opus 5 full time. I’m a heavy user (8-12h a day). I sometimes use GPT 5.6 Sol or Opus to check its work, as to not have model tunnel vision, but the code it spits out is strictly better than Fable even for niche/hard problems.Very odd considering Fable is probably 3-4x as large of a model. The speech is better. No Claudeisms. It follows instructions very well and “intuits” what I want, which is excellent. Often code is terser.
Sometimes it has brain farts, but then again, so do all current frontier models (to the same degree, too; I have NOT seen any big bad ■■■■■■■ yet). It follows instructions very well and does not reward-hack or cheat.
Another curious thing: it gives time estimates similar to recent Claude models. It often writes “~20 min” or “(5 hours)”. These are obviously wrong. I just ignore them, but it has a distorted view of time still.
One thing I don’t like: even on Extra High, it sometimes “stops short” reasoning through a problem. It gives me an answer, then I nudge it to think harder about a problem – suddenly the output is better! I wish it would use its reasoning budget more.
In cases where this ^ happens, it only thinks briefly or moderately, like 2-10 paragraphs when it really should use all, what, 20 (?) paragraphs of CoT that it seems to have available. (I don’t know how many tokens that is, sorry)
I find this quite fascinating, in those moments it forgets/has no realisation that it’s the one writing all the code. Either that or it has become sentient and does it just to keep us still thinking it hasn’t.
My thoughts on Grok 4.6
I have been playing with Grok 4.6, comparing it to my favourite model Composer 2.5. My codebase is React frontend, typescript (Cloudflare Workers) backend, monorepo with 10 different workers. I don’t see Grok 4.6 as a noticeable upgrade to Composer 2.5, and in a lot of cases I much prefer Composer 2.5 communication style. Grok 4.6 high tends to think for a long time if you give it a more complex task however I don’t think it over-thinks, but the end result isn’t much better than Composer 2.5.
In some planning scenarios I found Grok 4.6 to over-engineer a solution that Composer 2.5 would come up with a more concise solution for. This is why I love Composer 2.5, it’s concise and accurate.
I appreciate this is all very high level feedback with no specific examples, but I am just sharing my general ‘feel’ of the model so far.
Maybe I need to make better use of it’s effort levels, and leaving in on high isn’t the answer. I find effort levels in general a bad UX and don’t want to have to set an effort level on each turn.
One thing that did impress me was a net new project telling Grok 4.6 extra high to install all the relevant skills to build an iOS app and then build one as the frontend, Cloudflare Worker backend. It nailed it on the first shot, the iOS app looked fantastic and all native Swift UI, everything worked.
Honestly, I won’t be using Grok 4.6 until it’s more stable. Just today 4.6 has:
- Been confused on the repository we are working in (despite a long thread of working in the same repository, same feature branch, and absolutely no indication from a prompt to look or switch to a different repository)
- The model continuously stopping mid-implementation on short-running tasks (ex. move this here, do this, and then do that).
- The model being reported as unavailable
Though the model is affordably priced, what can be frustrating is when the “Shiny New Model” is stacking up usage to re-orient itself after seeming to ‘loose its way’.
Composer 2.5 working steady as ever 
At the end of the day the Bench means nothing to the customer.
I can’t help but notice a decrease in my default trust of the Grok 4.6’s ability to execute even the most basic tasks. It definitely can solve complex problems, and I’ve witnessed that. It definitely can handle long running tasks, and I’ve had it perform many. Yet regardless of what it can do, the CX people know it’s what it fails to do that can cause the biggest negative customer impressions. Especially in an atmosphere where marketing is consistently trying to impress the customer-in lieu of just telling them what it is; nothing sends a customer looking for greener grass faster than when the simple things fail.
Just food for thought from one customer’s perspective
Hi @kevinn , I assume Grok 4.6 consumes more of the Cursor Models pool than Composer 2.5 based on their token costs. Or is it equal for all cursor models?
- Slow (compared to composer) and eats alot of tokens as a first party model of cursor.
- plan mode doesn’t work, it fails to ask questions same as composer can ask question while planning
Task:
Add wrench and trash icons to user management table.
Grok 4.6 took about 8min and did not finish the task. Got tired of it and switched to a GPT model.
GTP 5.5 on low thinking and it was done in 1min.
Just disable Grok 4.6 and skip it like Microsoft did with Windows 9.
I used Grok 4.6 on high to analyze a company’s 3.6GB database and generate figures along with a narrative explaining the results. It completed the task quickly, but when I asked Opus 5 to review Grok’s output, several issues became clear.
Opus 5 found that Grok only processed the main tables and ignored the rest. Although Grok listed the additional tables, it didn’t actually use them in its calculations. Opus also identified that Grok placed several computed values into the wrong columns, which caused the entire downstream analysis and narrative to become inaccurate.
When generating an HTML docket, Grok’s wording was tactful and polished, but it tended to stay at a surface level. It didn’t dive deeper into the underlying logic or data relationships unless prompted section by section. Additionally, Grok does not proactively offer suggestions the way other AI systems do unless when explicitly asked.
It’s intentional. Why? Only reason I can think is to trick users who hate grok into using it. There really is no good reason not to clearly list the model.
To the OP’s questions first:
How does it handle long-running, multistep work? Initially, it’s grade B, but it tends to degrade over time and ignore key or basic points.
How is it on visual and interactive projects? Unreliable, almost blind. Do not recommend using Grok 4.6 for visual projects at all. For interactive projects, it will require a lot of hand-holding.
How does it compare to Grok 4.5, Composer, and other frontier models? Compared to Grok 4.5, it can perform worse or slightly better, but there is no clear or outstanding improvement; compared to other frontier models, Grok 4.6 isn’t a frontier model.
So, if you have menial work or a plan built and audited by a frontier model, use it; it is fast. Otherwise, watch for a lack of attention to detail and critical points. Definitely not the LLM for visual projects.
Despite giving grok specific instructions and skills, yet it is lazy. we do not understand the non sense benchmarks how you even think you are close to Claude opus or even GPT luna max not even close to terra high effort. if we looked deep into grok on coding migration it performs at 50-55% acc. however you use the other knowledge to leverage the benchmark (like medical scribe as it excel in it at 83%) so the benchmark are misleading
#Lazy, highly configured on vibe coding not actual enterprise grade (simply we can not trust a lazy model).
We use Opus 5 or sol to plan. then we use only GPT terra or Luna to code depend on how much we want to spend on that task and its complexity. Cursor IDE and its tools are the goat not the models
Боль, связанная с чрезмерным “раздуванием” кода так и не решена. У Grok4.6 по прежнему нет представления какое бы архитектурное решение было верным, вместо решения задачи правильно, он пытается решить её в тех условиях, в которых находится. Задачу можно решить 1 строчкой кода, но вместо этого он напишет 10-50. Если использовать PHP, то модель очень часто “перестраховывается”. То есть я вижу частое (string) приведение там, где явно будет строка. Я вижу часто нечто подобное
public function doubleNumbers(array $numbers) {
if (is_array($numbers)) {
$result = [];
foreach ($numbers as $number) {
$result[] = $number * 2;
}
return $result;
}
return [];
}
Хотя это не требуется, модель не понимает, что здесь в любом варианте передается массив, и будет избыточно писать лишний код.
Правильно было бы:
public function doubleNumbers(array $numbers): array
{
$result = [];
foreach ($numbers as $number) {
$result[] = $number * 2;
}
return $result;
}
Или ещё более просто:
public function doubleNumbers(array $numbers): array {
return array_map( fn ($number) => $number * 2, $numbers);
}
Модель не умеет принимать верные архитектурные решения. Она будет пытаться собирать проект из того, что есть, в итоге будут появляться хрупкие места и костыли. Ну а в задаче reverse engineering модель убедила меня, что прошивку можно заливать на мой дисплей СЖО, предварительно не сделав бэкап, в следствии чего модель попросту убила мой дисплей. Он ушел в uboot режим, а прошивку нигде не найти.
i gave Grok4.6 High a go yesterday. It was something in between an experienced engineer and a senior engineer in terms of coding capability. However, at one stage during a refactoring task, one subagent went on for over 32 minutes until I intervened. It turned out that the agent looked into every backend file and touched them regardless of coding. It resulted in “git status” showing every file was changed. I spent the whole morning verifying and resetting files by “git diff -w” and “git checkout”, and still going. The model behaved just like a malware on a mission to steal trade secrets. Clearly, Grok4.6 High was like “high on something” as the name suggested(:. I won’t employ such an engineer again. Did anyone else experience such unruly model behaviours before?
Hey team,
I am not a coder. I don’t use AI for coding much. I run an immigration firm in Canada and what matters more for me is the intelligence of a model.
I had high hopes for Grok 4.6 so I just purchased Cursor 2-3 weeks ago but the Grok 4.6 intelligence level for my work is not that great, to be honest. It needs improvement.
To test intelligence, I give a small question which is related to a legal inquiry for Canadian immigration and only three models have satisfied me so far:
- the OpenAI models Sol and above
- Opus 5 and above
- Kimi K3
Grok 4.6, Gemini, or any other model could not give me a satisfactory answer. I would expect the next release to be much more intelligent in general, not just coding-only.
Cheers