Apparently Grok 4.7 has 40% more weights than Grok 4.6, but the price ($6 output token, $2 input) is the same.
Given that the decrease in their margin and the fact they delayed the release of Grok 4.7 almost two weeks past the original date, XAI must not have been happy with the results for 4.7. And XAI also waited the day before Opus 5.5 is rumored to launch. I imagine Opus 5.5 will blow Grok 4.7 out of the water benchmark wise.
However, I have become skeptical of benchmarks. Grok 4.5 solved some issues setting up a buildroot system that Fable 5 couldn't do. I find the post cursor groks are phenomenal at frontend web development, though Claude is much better at backend ruby.
My favorite part of the new Groks has been how they speak in plain english. I simply cannot stand Claudish. Or even GPT, which doesn't have Claude's ticks but definitely likes to handwave explaining technical concepts. Still, nothing beats Claude 3.5 and 4 with explaining since it seems all models have regressed. I wonder if Grok 4.7 will also regress with English because of all the RL.
FYI a quick fix for claudish is to ask for the response to be in ASD-STE100 (Simple Technical English). Then it is far more readable. But I would agree that this is an annoyance and shouldn't require user workaround to get something readable.
I don’t think system prompts are particularly reliable way to do much at all. It’s better to put it as a hook after each response, or a skill at least so you can trigger it at will if you don’t want it everytime.
I totally agree, it’s like that as models become more intelligent, they are less understandable by most of people... but aren’t we humans doing the same?
Agreed. The more knowledge you amass on a subject, the more important it becomes to be extremely specific and nuanced - or your communications end up being incorrect. You become better at expressing your thoughts, but harder to understand.
The weird thing is, that's not what AI models seem to be doing. The prose is just weird.
> You become better at expressing your thoughts, but harder to understand.
This happens most though when the speaker doesn't (or care to) understand their audience.
Eg i find effective communication requires expertise in both the subject matter domain but also the reference of the listener. Eg in ELI5 framing, if you don't know what information 5yr olds are expected to know you'll do a poor job at an ELI5.
It often feels like Claude does poorly at both framing the response relative to what it "thinks" the listener knows, but also the prose is... sideways, just weird as you said.
I find people tend to lack nuance in their communication more often - leading to problems and misunderstandings.
If I don't grok an elaborate explanation, I can ask for clarification. If it's explained to me in an overly simplistic or unnuanced way, I'll walk away with a false sense of understanding.
That said, I'm sure we all have very different concentrations of these types of people and problems around us. I've definitely met some engineers who seem to actively try to make their language incomprehensible
The best ideas are usually the simplest to elaborate. If someone comes up with a convoluted scheme that are hard to understand or be adequately explained, it's usually fraud.
When claude speak in convoluted mess, they are often going off on tangents in real work that you asked it to do, too.
That believes that the world can be simplified into dichotomies, or at least, simplified. Sometimes problems are complex, and the solutions to them necessarily so. For example, cancer. I order to begin to understand that problem, you have to understand the utter complex scheme it has devised in order to exist. A 20 minute YouTube video isn't going to be able to begin to cover the basics of the subject, although there are some good ones, with clever analogies.
Just because something is difficult to understand doesn't mean it's fraud, although if someone is trying to dazzle you with clever words and names of institutions you recognize because they are selling you something, there's a good chance they're lying to you in order to get some money from you.
No but almost all good ideas can be reduced down to a few sentences if you're good at explaining things. It's a different kind of intelligence than what's commonly called IQ but it's something like that regardless.
Sure the explanation will oversimplify a lot but then you can expand it recursively if needed, you gotta start somewhere.
> That believes that the world can be simplified into dichotomies, or at least, simplified. Sometimes problems are complex, and the solutions to them necessarily so. For example, cancer
You just simplified most of the problems people work on down to cancer complexity. Ironic, isn't it?
That's also simply not the case, most people are building CRUD apps with some frontend code and some accessory stuff like build systems etc., which while complex, can still be expressed in very plain, easy to understand language for anyone who's a bit technical.
That's half true. A very smart model should be able make good explanations, which include simple understandable prose. That can should be possible even as its thought process gets more alien.
What I notice about Claudish is that it has its preferred cliche’s and overstretched methaphores, it packs too many ideas in a sentence, and to achieve the latter it makes up adjectives.
I should try adding these tips to my system prompt. Is there a shorthand to describe such language use? I am not a native English speaker.
Look up the output-style setting, which is a bit stronger than putting it in the system prompt. The new "concise" setting is better than the default but in practice, Claude is a very stubborn model when it comes to these patterns and they're really hard to eliminate, mostly you can only hope to mitigate.
As for the wording of the prompt, you're pretty on point, I created a custom output style targeting mostly the first two you have there. Some people have wording that demands a certain technical standard or uses fancy words to describe what to avoid, but I haven't seen evidence those work better than asking plainly and I suspect the opposite: LLMs mimic the user to a degree so talking to it in terms of technical specifications and fancy words is an invitation to get them back.
I was going to say the reverse - claude has been the less satisfying normalized by benchmark for me in the last year. Both astra and fable have their quirks, but I am 90% codex this year up from 10% last year.
It's not just about benchmaxxing. Sincerely targeting those long-autonomy benchmarks is questionable in the first place, because naturally it drives the model to assume more and more about what you want.
The target market for frontier models is CEOs who want to lay off entire departments of their company. So the long autonomy benchmarks would seem to be sending exactly the right signal.
Perhaps you haven't had the chance to use it, but 3.8 flash is the best model for talking too. Even routing Claudes output through 3.8 to have it explain whats going on is a breath of fresh air
> I find the post cursor groks are phenomenal at frontend web development, though Claude is much better at backend ruby.
Wonder if we'd benefit from a much more specialized + task-specific benchmarks to paint a clearer picture like this. A benchmark solely for frontend, ruby, hardware, etc.
I've noticed Chatgpt 5.6 Sol High, on the chat interface, inventing words that are a mixture of Portuguese and English. Like "hardcodar" a mix of "hardcode" and the most common verb ending in Portuguese "-ar". Some don't have a single google hit
> My favorite part of the new Groks has been how they speak in plain english.
I don't know if it's the plain english or what, but I really like Grok for legal research (as opposed to code). It's got a noticeable edge in getting to the point compared to Opus 5.
Same. The issue with Anthropics models is that (speaking regarding code generation) they REFUSE any kind of comment override instructions. I've tried everything and no matter what, after a few turns, they resort to generating the same overtly verbose junk. Bun's codebase is littered with them
See
// `HANDLE` is an opaque kernel handle (kernel32 validates and returns 0/FALSE
// on a non-console handle); every out-param is `&mut T` to a `#[repr(C)]` POD,
// ABI-identical to the Win32 `LP*` pointer (thin non-null). The reference type
// encodes the only pointer-validity precondition, so `safe fn` discharges the
// link-time proof. (`bun_windows_sys::kernel32` declares these with `*mut`;
// redeclared locally so the legacy-conhost cursor path below is plain calls.)
or
// Progress's terminal handle is the canonical `output::File` (vtable-backed
// stderr/File from `OutputSinkVTable`). The duplicate `ProgressTerminalVTable`
// from B-0 round 1 is removed; tty/ansi/winsize route through the new
// `OutputSinkVTable` slots so `bun_core` stays T0 (no `bun_sys` dep).
Nice to see this release cadence increasing and some continued improvement in quality. I am guessing these models are basically still outcomes of the cursor team integrating with the massive amount of compute they now own: I’d imagine we will see significant step up improvements with grok 5 later this year as the team gets more experienced and confident with larger training deployments. Here’s hoping for another competitive frontier model!
Astra fails in similar ways, and at similar frequency, as GPT 5.6 Sol does. It often goes way out of scope, or just stops prematurely, or tries to find odd and even dangerous workarounds when it gets stuck.
It's phenomenal at computer use and 3D stuff. I've been using it less and less for coding.
For some reason reasoning effort low and medium used similar numbers of tokens, and xhigh used less than high. I think I need to try without OpenRouter in the middle.
UPDATE: I tried again with the xAI API directly: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - not a great deal of difference between reasoning levels, and this time xhigh and low used the same number of reasoning tokens for some reason.
Are there good tools for doing context audits? I feel I have no good way to visualize what a new session is getting by default in a given repo without crawling through every potentially included markdown file
Is this an ad for Vals AI? Looking at their website, the rankings don't mesh with my observed utility for almost any model outside of fable and astra being good-ish.
I've been using 4.6 for some one-off game mods/utilities and it has done very well. "I have a very niche keyboard (Moonlander) and I play this very niche space sim, make me a SVG keyboard cheatsheet for it". Told me to grab keymap.c for the keyboard and inputmap.xml for the game's key bindings, churned for a while, then spit out a pretty good first attempt. Spent another hour of back and forth to refine it, and it's done: https://files.catbox.moe/x0u76x.svg
Excited to try 4.7. I hope they fixed the "it's not X, it's Y" that showed up in 4.6.
Well when I ran out of Grok SuperHeavy subscription ($300) once and tried to use extra credits to cover half a day remaining till reset, $50 in extra credits went in two hours. Based on that, subscription definitely lasts longer; Grok subscription just about covers a week of my work (sometimes a bit extra remains unused, sometimes it runs out half a day to a day early). And as a point of comparison, it lasts for doing same tasks as 2.5-3 weekly limits of Codex on 5.6 Sol did (using xhigh on both Sol and Grok); I needed 3x$200 Codex subscriptions to cover my weekly usage.
No doubt xAI has seen rapid progress, but it's been several months of them being "just behind" OpenAI and Anthropic. It seems the gap between just behind the frontier and pushing it is a lot wider than most people thought it was a year ago, and that's why a clear third contender in the frontier model space has yet to materialize.
Codex has become my goto tooling. I used to be a Claude Max subscriber, but I was becoming disappointed with the quality of the output from Opus 5. Fable chewed through my usage too quickly to be practical. Moving to a Pro account w/ Codex was a big improvement. Sol had great output, and the usage was more than sufficient for most of my needs. However astra does tend to chew up usage, so when i've done to much of that, and it's became an issue Grok Build has beocme my second go to account. The output especially after the cursor purhcase has become quite good, and the usage has always been very generous.
as someone who is limited by amazon bedrock support at work (no idea why we got stuck with the worst one) - grok is literally the only budget-ish model option, so nice to see it updated, Sol and Opus are just too rich for my blood. Luna is good but so slow at getting things done (tps wise it's fast)
Either way, the fact that xAI or SpaceXAI or whatever the name is, I can commend the team behind it on their rapid ascent and progress by being close and or on the frontier in several respects.
There for awhile it seemed like we’d have 3 big competitors but then Grok 4.2 or 4.4 was just diabolical while OAI and Claude continued their significant improvements. Grok was/is so bad that I was convinced musk was gonna shut it down and just fund Anthropic compute once they reached their compute agreement.
Same reason Coca-Cola only mention Pepsi and Pepsi only mention Coca-Cola.
It's an proven way to capture the market. You would rather split the pie in two rather than in 4,12 or 50 right?
Every Grok release obscures their cache pricing while highlighting their input/output pricing
From their headline comparison:
Grok: $2/$6 per million
Fable: $10/$50 per million
What this doesn't say: Grok costs 0.50/M cache read, Fable $0.25/M cache read
Long running agentic workflows are dominated by cache reads.
Just makes Grok sound deceptive, and more importantly, reliant on user's lack of understanding of costs aka predatory (which in turn is more infuriating)
In my experience Grok especially inside Grok build is pretty solid choice, it’s a no nonsense model and stays on its course. Another surface where I truly enjoy the experience of using Grok model is Grok bot
I've had really good experiences with Grok 4.6 and grok build. I've been playing around with tscircuit and it can write code with an understanding of spacial reasoning, while also importing cad components from different file formats into tsx, I've been having claude come in and try to error check it and so far claude hasn't found anything to improve in my three projects.
I'm excited for 4.7 although I share skepticism with other users whether 4.7 will be significantly better, since they didn't raise the price.
I refuse to use Grok. Mostly because of the usual reasons - somehow this high profile AI model seems more disgusting than others and it is in a way impressive.
But also Xai doesn’t seem to care about user experience and long term support.
I am subscribed to ChatGPT, Claude, Kimi and GLM coding plans. 200$ one on GPT and the 20$ ish ones on all others. Recently added Grok and it has somehow bacome my second most used model.
For daily one off questions I prefer it because it is fast enough and I like the way it responds. I also use it for basic research like “find me a battery drill for this and that”.
Kimi and GLM feel extremely coding oriented. I use them for code reviews basically. I hate the way Anthropic models talk. GPT takes too much time and effort for that kind of stuff for some reason.
You should try it, it is less sycophantic than other models and is faster and better at most reasoning levels, don't confuse the twitter bots and services also named Grok with the frontier model itself
I can't take anyone seriously who uses grok seriously. I like to look at the cybertruck owners forum every so often because it's just... hilarious. And the amount of superfluous grok use over there is just insane. Half the posts I click in there will have a bunch of people dumping entire grok takes "why do people hate cybertruck owners?" "Because they're jealous and poor," sort of stuff that they just LOVE to post.
As a technical point of reference to compare against other llm stuff, sure, I'll glance at a report or benchmark but I really couldn't care less about anything to do with the project and it could blow other options away and I wouldn't touch it.
If the CursorBench 4.0 score diagram is the headline, I read it as "Grok 4.7 xHigh is almost the same as Fable5.1 on low".
Is there a metric for like... time taken when comparing these two? I see score and cost.
If Fable5.1 can knock it out more quickly on low but Grok4.7 might take twice as long to stumble through a problem (and leave behind a bunch of yucky comments or un-needed extra unit tests), are they really comparable?
Or like... the "quality" of the solution? "It works" versus "it's unmaintainable/very messy/hacky".
I tried in Omp (Oh-my-pi), and so far it's really problematic.
It will loop in thinking mode ("Let me implement those fixes: Fix 1, Fix 2, Fix 3 .... Fix 80, Fix 81"), ignore the AGENTS.md instructions, corrupt plan files, etc etc... I have 5.6 Sol as advisor/watchdog, and it blocks every turn, I never saw this. Quite a shame, 4.6 wasn't so bad.
Probably the same reason as oh-my-zsh, you don't need 90% of it. Further compounding the problem in an agent harness is that you are polluting the context window by throwing the kitchen sink at it.
- it allows different models within one session via roles (I only have API, so pay per token)
- it's much more likely (ime) to use the LSP over grep for determining how code fits together
But I agree a 20k+ starting context is way overkill.
I find it's very hard to get information on harnesses people are using. I have to stay model agnostic so I avoid claude, codex, cursor, etc. I've used and tried opencode, which worked well, but obviously lacks the above features.
Does anyone have a resource for following what people are actually being productive with? With so much vibe going on it's hard to separate the wheat from the chaff.
This explains why. Mentioned in another comment, but cursorbench explicitly tests with Cursor as the harness, and OpenAI doesn't allow them to use Astra in Cursor.
Cursor never added Astra to its consumer subscription plans. And it's likely exactly because of this announcement. Why would they add support for a model they would have to remove shortly after?
That's not accurate. OpenAI doesn't allow Grok to provide Astra to Cursor customers anymore, but it doesn't ban anyone from using Astra via alternative harnesses.
If Cursor wanted to include Astra in CursorBench nothing would stop them, they could easily have spent half an hour vibecoding in OpenAI API key support - if it hadn't been convenient to neglect to do that.
Even if they could do that (workaround to include Astra in CursorBench), that has no practical consequences for Cursor users and that's what I as a Cursor user (what I use for dev, though I use ChatGPT for non-dev stuff) care about.
It would make the benchmark way better obviously, by showing how their new model compares to their competitors, the whole point of benchmarks and graphs.
The point of Cursor Bench is to show how models perform in Cursor. If 99% of their users won't be able to access a model unless they go out of their way to include setup an API key for it (which would be insanely expensive with Astra), why would they include it in the benchmark?
Deceptive? An extremely quick google search would answer your question. OpenAI pulled out of Cursor before they released Astra so it never got that benchmark.
I wonder if that means that SpaceX evals show that they consider astra better than fable or that they hate Sam&co so much they don't want to show their stuff.
Its because of this. You can't use Astra in Cursor, and cursorbench uses cursor as the harness. They can't actually benchmark it using their harness hence why its not included.
This has been my experience as well. Grok will end tasks almost immediately and claim "Done!". It's definitely the laziest and most "dishonest" of all the models. The others aren't perfect, but I can't use Grok for any serious coding task.
I used openrouter to send same prompt to qwen, derpseek, gemini and grok and found that grok does good research and produces less bullshit, especially when prompted to be critical of an idea
Have you considered that the single most impressive breakthrough of LLMs as a technology is their ability to generalize beyond what they were explicitly trained on? Great analogy, pal, but LLMs aren't cars.
Everyone should be clear that this is what they’re cheering on when they celebrate a Grok performance win. A technology is no longer neutral when wielded by a self-proclaimed white supremacist whose actions have killed over a
million black and brown people, mostly children and babies.
Think we all can agree he has had staggering successes, but they have all come from having massive capital from Paypal which wasn't anything super innovative, it just solved a convenient problem at a convenient time and was awarded handsomely. Elon has put his capital to work in various ways to become successful, not all of the ways being morally sound.
About half of SpaceX revenue is Starlink subscriptions. Starlink is the one profitable division; the rest of the company operates at a loss, including xAI.
As if "the public" knows literally anything about how the US federal government is administered.
If anything, they voted for reduced debt burden and they got the opposite. DOGE failed at pretty much every single one of the goals that the public arguably gave it a mandate for.
Ah, yes, democracy!, except for when the public is wrong.
Who decides when the public is wrong? We do! Who decides "what the public voted for"? We do! So we are the rulers? No, of course, not, this is democracy.
You want to become the decider of when the public is wrong and of what the public voted for? TYRANT! TYRANT!
No, the claim above is "I know the voters' intent behind their vote based on who they voted for."
This is simply epistemologically incorrect. It's obviously incorrect in this case because voters writ large do not have any idea how the government is administered and how to improve it, so even if they claimed to be voting for that, it would not necessarily be an endorsement of any particular approach.
More specifically we know it's not true in this case because there are polls. Voters didn't even claim to care about this! "How the government is administered" was not a high salience issue to voters. Simple as that.
Nonetheless, I didn't suggest anything about overriding their votes. It sounds like you have some sensitive spots to work through (someone obliquely criticizing your idol for sucking at his job?)
Seriously. They already get caught uploading everyone’s private credentials once before, one would have to be a particularly gullible rube to trust grok again. Especially with musk in charge.
Thank you for your perspective and I agree as repetitive as it is, I hope hn can continue to surface disagreement/substantial views on a topic.
These comments don't stay up much anymore and I can't tell if it's structural to the forum (flag weight + statistical mechanics of votes + guidelines) or if it's the userbase sentiment.
I completely agree. But this has been true for many years. This sort of head in the sand compartmentalization seems to be a core feature of the culture here.
I think I already made my point, but I'll make it again.
Nobody both worked and spent their money to get Trump elected like Musk. 300 million to his 2024 campaign [1]. DOGE. On-stage endorsements. Nobody even came close.
No, other big labs are not "innocent little virgins", but they're not even in the same solar system of harm as Musk. To hand-wave at the differences is to permit them.
I have to say I'm a little tired of whenever a Musk related product comes up there's random nolifes that arrive to rant about politics. Luckily they're relatively rare on hacker news.
Also it's kinda hilarious how you think any money spent on Grok will go toward furthering climate change versus literally any other AI model that does the same thing. Grok at least seems to be more efficient than most models.
How many years will go by before people will stop talking about the CSAM thing that happened for only a few days as if its still happening? FWIW, it's bad it was generated, but equating it to harming of real children I think is not a good thing.
And methane is a large percentage of all power production in the US. So again that also applies to all the other data centers. (And FWIW they've been winding down and shutting down the on site methane generators.)
It’s not just that it was generated, it’s that they argued in court _and won_, that it was legal, giving every last predator a nigh-on unverifiable defence.
I think you're seeing a big shift around it.... since it's been markedly cheaper and also still easily available from OpenCode, it's getting large enterprise traction.
Well I "tried it out" I asked it one question, and it gave no answer and said "Sign up to use more!" I don't think I'll be doing that, no.
I can't think of a single dimension grok is winning on (capability, cost, voice), but want to stay open-minded -- anybody want to vouch for its capabilities in any domain?
If you haven't used it, how do you know if it's winning?
I think it's winning on UI for normies (grok bot) and they made some claims about being pareto SOTA (lowest cost per task completed) a while back with 4.6.
I find it to be a perfectly capable model for implementation (there are many in this class--deepseek flash, spark1.3, luna, etc). I find the usage to be very generous w/ supergrok. I find the model to be just fine for 90% of what I want to do, but I use a smarter model to plan complicated things.
After the cursor aquisition it's become a quite capable coding model. If you take cost into account, it's close to the top. OpenAI is maybe still #1, but I'd put Grok at #2 (again, including cost as a factor).
For me it works well for agentic coding tasks and terminal/unix/bash (in cursor and grok build); it's also token efficient and cheaper than gpt 5.6. It's def not as good as Fable for me (I haven't used Astra much, can't comment). So it's not the cheapest, not the most capable, but it has a good mix of it for my backend, go, infra work.
The voice is the weird part. The early Grok 4 models had a very distinct presentation unlike anything else out there. Then suddenly it made a big jump in coding ability and started sounding just like every other model.
Given that the decrease in their margin and the fact they delayed the release of Grok 4.7 almost two weeks past the original date, XAI must not have been happy with the results for 4.7. And XAI also waited the day before Opus 5.5 is rumored to launch. I imagine Opus 5.5 will blow Grok 4.7 out of the water benchmark wise.
However, I have become skeptical of benchmarks. Grok 4.5 solved some issues setting up a buildroot system that Fable 5 couldn't do. I find the post cursor groks are phenomenal at frontend web development, though Claude is much better at backend ruby.
My favorite part of the new Groks has been how they speak in plain english. I simply cannot stand Claudish. Or even GPT, which doesn't have Claude's ticks but definitely likes to handwave explaining technical concepts. Still, nothing beats Claude 3.5 and 4 with explaining since it seems all models have regressed. I wonder if Grok 4.7 will also regress with English because of all the RL.
I totally agree, it’s like that as models become more intelligent, they are less understandable by most of people... but aren’t we humans doing the same?
The weird thing is, that's not what AI models seem to be doing. The prose is just weird.
This happens most though when the speaker doesn't (or care to) understand their audience.
Eg i find effective communication requires expertise in both the subject matter domain but also the reference of the listener. Eg in ELI5 framing, if you don't know what information 5yr olds are expected to know you'll do a poor job at an ELI5.
It often feels like Claude does poorly at both framing the response relative to what it "thinks" the listener knows, but also the prose is... sideways, just weird as you said.
If I don't grok an elaborate explanation, I can ask for clarification. If it's explained to me in an overly simplistic or unnuanced way, I'll walk away with a false sense of understanding.
That said, I'm sure we all have very different concentrations of these types of people and problems around us. I've definitely met some engineers who seem to actively try to make their language incomprehensible
It is unsurprising that a LLM fails, without coaching, to effectively communicate.
Part of intelligence is knowing your audience and communicating efficiently.
When claude speak in convoluted mess, they are often going off on tangents in real work that you asked it to do, too.
Just because something is difficult to understand doesn't mean it's fraud, although if someone is trying to dazzle you with clever words and names of institutions you recognize because they are selling you something, there's a good chance they're lying to you in order to get some money from you.
Sure the explanation will oversimplify a lot but then you can expand it recursively if needed, you gotta start somewhere.
You just simplified most of the problems people work on down to cancer complexity. Ironic, isn't it?
That's also simply not the case, most people are building CRUD apps with some frontend code and some accessory stuff like build systems etc., which while complex, can still be expressed in very plain, easy to understand language for anyone who's a bit technical.
Does not excuse the Claude slop.
Solving the problem right in front of you is easy. Stepping back and asking: is that a problem to be solved, is infinitely harder.
I did not use Claude to write my comment, so I don't know where that is coming from.
I should try adding these tips to my system prompt. Is there a shorthand to describe such language use? I am not a native English speaker.
As for the wording of the prompt, you're pretty on point, I created a custom output style targeting mostly the first two you have there. Some people have wording that demands a certain technical standard or uses fancy words to describe what to avoid, but I haven't seen evidence those work better than asking plainly and I suspect the opposite: LLMs mimic the user to a degree so talking to it in terms of technical specifications and fancy words is an invitation to get them back.
That being said, I currently prefer Sol / Astra to Opus / Fable as I find both to be a better cost payoff to me.
Wonder if we'd benefit from a much more specialized + task-specific benchmarks to paint a clearer picture like this. A benchmark solely for frontend, ruby, hardware, etc.
I don't know if it's the plain english or what, but I really like Grok for legal research (as opposed to code). It's got a noticeable edge in getting to the point compared to Opus 5.
How representative that is of real world usage, I don't know.
In their benchmark GPT 5.6 Sol performs suspiciously poorly compared to the former models.
Fable 5.1 is not there quite there yet.
They need to get that Sonnet 3.5 magic back.
xAI missed its chance, Ball is on Anthropic's court.
It's phenomenal at computer use and 3D stuff. I've been using it less and less for coding.
Here's reasoning level high: https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
For some reason reasoning effort low and medium used similar numbers of tokens, and xhigh used less than high. I think I need to try without OpenRouter in the middle.
UPDATE: I tried again with the xAI API directly: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - not a great deal of difference between reasoning levels, and this time xhigh and low used the same number of reasoning tokens for some reason.
https://x.com/ValsAI/status/2102086608476590432
I like to follow them and look for benchmark for each LLM release.
Excited to try 4.7. I hope they fixed the "it's not X, it's Y" that showed up in 4.6.
[1]https://eebench.org/
GLM or Kimi are better for my own personal projects. DS? uhm. it just keeps doing dumb crap
Astra for deep dive investigations, Sol 5.6 at mid-level for day to day tasks, Grok 4.6 via Cursor for routine and low complexity tasks.
There for awhile it seemed like we’d have 3 big competitors but then Grok 4.2 or 4.4 was just diabolical while OAI and Claude continued their significant improvements. Grok was/is so bad that I was convinced musk was gonna shut it down and just fund Anthropic compute once they reached their compute agreement.
From their headline comparison:
Long running agentic workflows are dominated by cache reads.Just makes Grok sound deceptive, and more importantly, reliant on user's lack of understanding of costs aka predatory (which in turn is more infuriating)
I'm excited for 4.7 although I share skepticism with other users whether 4.7 will be significantly better, since they didn't raise the price.
But also Xai doesn’t seem to care about user experience and long term support.
For daily one off questions I prefer it because it is fast enough and I like the way it responds. I also use it for basic research like “find me a battery drill for this and that”.
Kimi and GLM feel extremely coding oriented. I use them for code reviews basically. I hate the way Anthropic models talk. GPT takes too much time and effort for that kind of stuff for some reason.
Grok happened to be a nice middle ground.
As a technical point of reference to compare against other llm stuff, sure, I'll glance at a report or benchmark but I really couldn't care less about anything to do with the project and it could blow other options away and I wouldn't touch it.
Is there a metric for like... time taken when comparing these two? I see score and cost.
If Fable5.1 can knock it out more quickly on low but Grok4.7 might take twice as long to stumble through a problem (and leave behind a bunch of yucky comments or un-needed extra unit tests), are they really comparable?
Or like... the "quality" of the solution? "It works" versus "it's unmaintainable/very messy/hacky".
It will loop in thinking mode ("Let me implement those fixes: Fix 1, Fix 2, Fix 3 .... Fix 80, Fix 81"), ignore the AGENTS.md instructions, corrupt plan files, etc etc... I have 5.6 Sol as advisor/watchdog, and it blocks every turn, I never saw this. Quite a shame, 4.6 wasn't so bad.
- it allows different models within one session via roles (I only have API, so pay per token)
- it's much more likely (ime) to use the LSP over grep for determining how code fits together
But I agree a 20k+ starting context is way overkill.
I find it's very hard to get information on harnesses people are using. I have to stay model agnostic so I avoid claude, codex, cursor, etc. I've used and tried opencode, which worked well, but obviously lacks the above features.
Does anyone have a resource for following what people are actually being productive with? With so much vibe going on it's hard to separate the wheat from the chaff.
This explains why. Mentioned in another comment, but cursorbench explicitly tests with Cursor as the harness, and OpenAI doesn't allow them to use Astra in Cursor.
That said, I don't expect them to benchmark Astra in their Cursor harness given the situation.
If Cursor wanted to include Astra in CursorBench nothing would stop them, they could easily have spent half an hour vibecoding in OpenAI API key support - if it hadn't been convenient to neglect to do that.
Its because of this. You can't use Astra in Cursor, and cursorbench uses cursor as the harness. They can't actually benchmark it using their harness hence why its not included.
The personality is bland and it doesn’t work nearly as hard or even tries to help.
Until you ask it to start generating horrific imagery and then it's best in class.
I don't use Grok, but do you want your LLM to have a personality? "Personality" is exactly what people don't like about Claude.
That it isn't the most efficient way to achieve the same end result is irrelevant.
Even with his successes (Tesla, SpaceX) he has built them up in large part by bending levers of government to his advantage.
Can you provide specific examples of where Elon has bent the levers of government?
So what? Thats called being a maverick. He is very very good at executing on making money which is the point of business.
Also pushing technology forward.
Anthropic: 1.25B/month
Google: 0.92B/month
Unnamed customer starting in december: 1.1B/month
Starlink monthly revenue is ~1.5B/month
If anything, they voted for reduced debt burden and they got the opposite. DOGE failed at pretty much every single one of the goals that the public arguably gave it a mandate for.
Ah, yes, democracy!, except for when the public is wrong.
Who decides when the public is wrong? We do! Who decides "what the public voted for"? We do! So we are the rulers? No, of course, not, this is democracy.
You want to become the decider of when the public is wrong and of what the public voted for? TYRANT! TYRANT!
This is simply epistemologically incorrect. It's obviously incorrect in this case because voters writ large do not have any idea how the government is administered and how to improve it, so even if they claimed to be voting for that, it would not necessarily be an endorsement of any particular approach.
More specifically we know it's not true in this case because there are polls. Voters didn't even claim to care about this! "How the government is administered" was not a high salience issue to voters. Simple as that.
Nonetheless, I didn't suggest anything about overriding their votes. It sounds like you have some sensitive spots to work through (someone obliquely criticizing your idol for sucking at his job?)
half of voters don't pay any attention to politics until the week or two before voting
Sheep often like to think themselves the wolf or coyote, it would seem.
Fuck, it is like the denial around Jan 6th. Those idiots we’re live streaming that shit. I watched it go down live. Now they say they weren’t violent.
We can’t have discourse when we have legit video evidence and people refuse to open their eyes and choose to deny reality
Which Nazi ideologies do you think he embraces? How do you reconcile all the Nazi ideologies he rejects?
These comments don't stay up much anymore and I can't tell if it's structural to the forum (flag weight + statistical mechanics of votes + guidelines) or if it's the userbase sentiment.
Nobody both worked and spent their money to get Trump elected like Musk. 300 million to his 2024 campaign [1]. DOGE. On-stage endorsements. Nobody even came close.
No, other big labs are not "innocent little virgins", but they're not even in the same solar system of harm as Musk. To hand-wave at the differences is to permit them.
[1] https://www.opensecrets.org/2024-presidential-race/donald-tr...
Also it's kinda hilarious how you think any money spent on Grok will go toward furthering climate change versus literally any other AI model that does the same thing. Grok at least seems to be more efficient than most models.
Handing corporate code secrets to his AI model is... unusually trusting.
And methane is a large percentage of all power production in the US. So again that also applies to all the other data centers. (And FWIW they've been winding down and shutting down the on site methane generators.)
And no corporate code was handed to AI models.
1 - https://bench.killswitch-lang.org
For now, I doubt anyone would notice your protest if you didn't announce it.
I can't think of a single dimension grok is winning on (capability, cost, voice), but want to stay open-minded -- anybody want to vouch for its capabilities in any domain?
I think it's winning on UI for normies (grok bot) and they made some claims about being pareto SOTA (lowest cost per task completed) a while back with 4.6.
I find it to be a perfectly capable model for implementation (there are many in this class--deepseek flash, spark1.3, luna, etc). I find the usage to be very generous w/ supergrok. I find the model to be just fine for 90% of what I want to do, but I use a smarter model to plan complicated things.
The voice is the same AI slop as the others imho.
(This is about Grok 4.6, I didn't test 4.7 yet).
edit: clarified I mean agentic coding tasks