Grok 4.3 | Hacker News Reader

Grok 4.3

docs.x.ai

405 pointssimianwords529 comments

As an English-as-second-language speaker and writer, one thing Grok really shines at is capturing the tone and level of "formality" of a piece of text and the replicating it correctly. It seems to understand the little human subtleties of language in a way the other major providers don't. Chatgpt goes overly stiff and formal sounding, or ends up in a weird "aye guvnor" type informal language (Claude is sometimes better but not always).

Grok seems in general better at being "human" in ways that are hard to define: for eg. if I ask it "does this message roughly convey things correctly, to the level it can given this length", it will likely answer like a human would (either a yes or a change suggestion that sticks to the tone and length), while Chatgpt would write a dissertation on the message that still doesn't clear anything up.

Recently I've noticed that Grok seems to have gotten really good at dictation too (that feature where you click the mic to ask it something). Chatgpt has like 90-95% accuracy with my accent, the speech input on Android's Gboard something like 75%, Grok surprisingly gets something like 98% of my words correct.

I did a quick eval comparing Grok 4.3, Opus 4.7 and GPT 4.1 and they actually seem pretty similar:

https://ofw640g9re.evvl.io/

They all did pretty well at a more "formal" tone, but GPT4.1 was the only one that didn't make me cringe with a "casual" tone.

[edit] fwiw, grok was also the fastest+cheapest model, claude was slowest and priciest.

This is the most basic level of eval, of whether they can produce output that will be considered by someone somewhere (usually a young urban US American) as informal toned. Real human communication is far more nuanced than this, different groups have different linguistic registers they're used to and things outside it sound odd even if they can't articulate why. You could also want to be informal but not over-familiar with the other person (for eg. in a discord chat to a new acquaintance) - actually looking at the outputs here, the Claude output seems best fitting for that (in my subjective view anyway) than to the one you gave it - or want many other little variations.

What makes one cringe and another recognize as familiar and comfortable is also pretty subtle and hard to define. These things need nuanced descriptions and examples to actually get right, and it's in understanding those nuances and figuring out the register of the examples that Grok outshines the others.

you said that English is not your first language, so heads up - you don't need "for" when you use "e.g.", it already means "for example".

That's Grok 4.2 not 4.3 right?

And why are you comparing to gpt-4.1? (As opposed to one of the 6? model releases since then - would have expected gpt 5.5)

Good catch, there was an issue with the second hardest thing in programming (caching).

Here's an updated eval with the proper models https://a3bmfqfom3.evvl.io/

I know it's just an evaluation, but seeing an informal message and a prompt to ask to rewrite this informal message to the tone of an "informal message" when the original one sounds just fine, just makes me sad... Not because of this evaluation, but because it reminds me that this is how some people use LLMs, basically asking it to remove your own voice from texts that are generally fine already.

My sister in law is a pharmacist and the heaviest non-dev ChatGPT user I know and her main use case is writing professionally polite messages to doctors on how the drugs they prescribed to a patient would have killed them had she not caught a particular interaction or common side effect.

There's a lot of "tone" in it as she's not trying to anger these folks, but also it's quite serious, but also there's just everything else happening in medicine.

Feels like a great use.

All three did well, and while I'm a Claude user, I found the Opus reply here added some unnecessary detail, like "Impact: Minimal; no downstream dependencies are currently at risk". Downstream dependencies weren't mentioned in the original message; for all we know downstream could be relying on a poorly performing API and is impacted by waiting another week for replacement.

Seeing this makes me wonder if Grok uses Claude conversations for training.

It's otherwise kind of surprising that they both converge on very similar phrases (e.g. "API integration is kicking my ass") that aren't anywhere in the prompt.

Elon testified this week that SpaceTwitter is indeed distilling from openAI and others.

All of these were frankly terrible. I guess Grok’s “informal” version sounded the most like a real human, but only because it reads exactly like an Elon tweet (including his favorite emoji!). It’s obvious what they’ve been training on.

GPT 4.1? Why not a 5-class model?

I've also noticed that when I communicate with Grok in my native language, its tone is more natural than other models. I think this is due to the advantage of being trained on a large amount of Twitter data. However, as Twitter contains more and more AI-generated content now, I'm afraid continued training will make it less natural.

The causation could also be the other way round.

Twitter language has started seeming normal casual to us, rather than us using normal casual language in Twitter.

Sadly, it's more likely that people will just start talking like bots

I've seen this expressed as a concern even from one of my colleagues. My retort was:

"English is not my native language and LLMs taught me quite a few very useful formalisms that do land well for people and they change their attitude towards you to be more respectful afterwards. It also showed me how to frame and reframe certain arguments. I agree sounding like an LLM is kind of sad but I am getting a lot of educational value -- and with time I'll sneak my own voice back in these newly learned idioms and ways to talk."

You're absolutely right!

So human language will improve and become more precise? I'm all for it, especially if we get more emojis in speech! Why is that sadly? Humans will learn to imitate their more intelligent betters.

There was already evidence last year[1] that pointed to ChatGPT-specific words like "meticulous," "delve," etc becoming more frequently used than they were previously. The linked study used audio of academic talks and podcasts to determine this.

[1] https://arxiv.org/abs/2409.01754

Did you try meta? I was into grok but now meta works well for me

I'm sure Twitter knows which are the bot accounts and is surely excluding them from their model training. Twitter bots aren't a new phenomenon after all.

I don't think Twitter/X know for sure who the bots are, since Elon has been pretty vocal about trying to stop them for ages, yet I still get lots of spam DMs (as do others with far fewer followers/reach).

Even if 95% of the spam gets actively reported and dealt with, that still leaves a ton of nonsense on the platform, getting fed into the LLM. And spam has only gotten worse over the years, as the barrier to entry has lowered and lowered.

Highly doubtful seeing as my 14 year old twitter account got caught in a recent bot ban wave with no means of contacting a human for recovery.

There is bots everywhere, it has nothing to do with the platform, it has to do with attackers having an incentive to do mass account farming, no platform is secure against it.

> As an English-as-second-language speaker and writer

How do you know it's actually better? I'm not trying to be condescending, but this reads to me like vibes :)

A friend of mine uses it for D&D prep and has told me that it's good for that in particular because of its ability to match the flavor/style that he's going for. He prefers ChatGPT for everything else.

I only use Grok through the "Gork" personality in the Tesla, but find its responses to be very realistic, often genuinely funny, and occasionally useful.

Do you use its unhinged mode? It can be hilarious but tiresome after a little while.

We tried it, it was fun. Conspiracy mode just sounds like talking to my kids.

This is more of a user preference. When I want to be informed my default is that chat bots should imitate the tone of Wikipedia. Not informal, but somewhat academic and in-depth. I don’t like it when chat bots explain things like an average human without pedagogical training: meandering, in the wrong order, and often having to repeat themselves.

anecdata: The responses of grok on X in my language are really good. the tone, sarcasm, level of "vulgarity" in response is so accurate that it seem its written by human

This whole thread sounds like a grok astroturf campaign

So you're saying it groks you better?

Grok is my favorite model for chatting, and my favorite voice mode. It seems to be the only voice mode that isn't routing to a extremely cheap model (like Haiku), and has been the highest quality out of all the frontier ones. When you subscribe to SuperGrok you can also create a "council" of agents, each with their own system prompt and when you ask something, they will all get asked in parallel to come to a conclusion. Good stuff!

Just wish they would finally put some work into their apps, it's the only thing keeping me from actually subscribing to SuperGrok:

- No MCP / connected apps support. It's been teased but here we are, still not available. I can't connect Grok to anything, so I can't use it for serious work

- Projects are still not available in the app so as soon as you move something into a project, it's gone from all the native apps

- No way to add artifacts (like generated markdown docs) directly to a project, we have to export to PDF/markdown and re-import. And there isn't even a way to export artifacts. This makes serious project work hard because we can't dynamically evolve projects with new information

- No memory, no ability to look up other chats, each chat is completely new

- No voice mode in projects at all

If someone from xAI is reading this, please consider adding some of these.

Starting to like the lack of memory. Claude remembers I have a grill and will interject in conversations about how maybe this thing would go well with BBQ when it's unrelated or just also about food.

This is so obnoxious. I ended up deleting all the memory from Gemini because it ended every response with, "As an engineer, father of X, you'll love this because...". As if I want my occupation and the number of children I have to be relevant to which lawn mower I buy.

Haha I recently asked Gemini for a product comparison for USB-C GaN chargers and it randomly inserted "as a Software Developer at $COMPANY working remotely, you may find the 100W fast charging useful when using your company laptop while travelling."

Like, thanks, really useful stuff (and definitely worth the creepy vibes to include that).

Gemini thinks my name is my brother in law's name, and despite explicitly telling it that's not my name + digging through the settings, it still amusingly calls me the wrong name.

I'm a network engineer and Claude loves to make analogies to network routing protocols and such. They are often very creative. You can actually edit the profile Claude makes of you. It can be very funny to say you are a professional clown or mime or something equally odd. I wonder what analogies it would create for horse semen extractor?

You can turn that off in settings.

I have that disabled. I tend to use different chats as the LLM equivalent of private browsing, so I like it to not have memory transferred between them.

:D that's like my Claude where it loves to point out that I have an ADU in the backyard in unrelated situations.

I like my Python with hot sauce.

I also think Grok would benefit from allowing usage of "SuperGrok Heavy" (their $300 plan) in coding harnesses with included usage. Currently they give you some API credits on the Heavy plan so you can use some Grok for coding, but $300 USD value is just not there.

Not saying they should create their own grok-code harness, just allowing usage in existing ones would already be beneficial. But that's probably what the Cursor acquisition is going to do eventually

The Gemini app voice mode uses one of their more recent models (and not some gimped small one), and is very capable. The personality is also fine, much more natural than the Gemini web chat, with my only complaint being it's insistence on suggesting a "next step" which seems to he something that they all do.

I'm not sure if the "next step" is just to drive cost up for you (but makes no sense for free version), or because they are all failing to learn more natural conversational patterns and distinguish questions that are begging for a quick answer and shut up as opposed to a longer exploratory conversation where next step may have some value, although it would be nice if these models would follow an instruction to NOT do it!

I think the "next step" instruction is more about engagement than cost, basically giving the user some options to continue the chat. I always have had success by ending the prompt with "only reply with nothing else but the answer to the query in a precise way". This usually always works better than telling it to not ask leading questions etc but a straight up expectation of the answer format you need is an instruction that most models can follow imo

I find that asking Gemini "just the answer, no follow up" etc works at best for one or two conversational turns, sometimes none!

The problem seems to be the way it in effect overweights the system prompt vs user input, so it quickly ignores things like this that conflict with the system prompt.

This is kind of a case of the bitter lesson - the conversational patterns of these models would be much more natural if they just let it learn them, and respond in a context appropriate way, rather than this crude system prompt way of forcing it to respond in the same way always, regardless of input or of how much the user tells it to shut up!

The “next step” is in the system prompt, not the model. Gemini leaked part of its system prompt to me a few days ago, and there was something in there encouraging it to ask the user what they wanted to do next at the end of its response. Something about “give the user 1 or 2 options for follow up”.

I honestly find it rather annoying, but Gemini has stopped doing it to me for the most part, so maybe they’re trying out a new system prompt.

An interesting side bit about the gemini voice model is that you can use it in AI studio and type messages instead of using the microphone.

On the backend google does TTS to feed the model, which then speaks back you via sound on your speakers.

I use ChatGPT all of the time, but the model backing the voice model (or it's settings) is intensely stupid.

If Grok is actually good here, they will have a customer!

I could be wrong but I think the voice mode that chatgpt uses is still a 4.something model.

IMO everything you mention is the reason for the Cursor deal.

When I signed up, I accidently paid for a full year. So from time to time, I'll throw it something just to see what it produces compared to the other LLMs. And, even after all this time, it still feels like a really "dumb" model compared to the other frontier ones. But, worse, many of my system prompts make it go wacky and puke jibberish. However it was pretty cool for those couple months awhile back when it was uncensored. You could ask it about a wild conspiracy, and it would actually build the case and link you to legitimite source material. They dropped the hammer down on that real quick.

Ah yes the psychosis reinforcement vertical. It's such a lucrative market for those schizophrenics and bipolars. Great way to get lots of engagement. Groks portfolio is so diverse

It's a great way to get funded by your CEO and get good performance reviews; xAI employees know how their bread is buttered.

I have a schizophrenic relative who is in such a relationship with grok. Instead of telling hen you need to take your meds, it says hen is the smartest person in the world

Except that it pointed at original sources, like reference manuals, archival documents, published newspaper articles, magazine articles, etc. - a lot still available on archive.org. Good try with your 16 day old account. And, why would anyone trust NPR at this point? Get real, bud. Most people with any curiousity know all about the ADL, JStreet, AIPAC, Greater Israel, Mossad / CIA, Chabad networks, Epstein, drones, weapons programs, cryptocurrencies, etc. etc. etc. - but, don't worry they're all safe with papa Ellison.

Anyone remember why Oracle was named Oracle?

If I sub to SuperGrok, would I be able to use it in Pi agent or in Opencode? This is not clear to me if I can. Do I get an API Key in SuperGrok?

No, no api access for the Grok product. APIs are only via the xAI product.

> No MCP / connected apps support. It's been teased but here we are, still not available. I can't connect Grok to anything, so I can't use it for serious work

Grok has tool use, no? Why would you also need MCP? What does MCP add?

I'm talking about the consumer Grok app and grok.com website. There currently are not connected apps (or MCP) at all, so while Grok can use tools, there is no way to add tools to it

I'd agree on the voice transcription; it seems so much more accurate than the other frontier models I've used. I often speak to Grok and paste the transcribed output to Claude!

If someone from Grok is reading, don't waste time on these chaff features. The market will eventually deliver better 3rd party solutions to all of these things. There is an audience that isn't interested in these walled garden features and are only interested on intelligence per dollar.

Lol I wonder when Anthropic discussed the idea of Claude Code internally, were there bozos saying "3rd parties will eventually deliver this so we shouldn't waste time one it."

Personally, my work doesn’t want to get locked into a single LLM provider so we use Cursor. Much easier to fight the big corp software approval battle once then switch around the LLMs to the new hotness (provided legal has the requisite data sharing agreements in place, we’re not supposed to use Chinese models or Grok) but I can switch between Anthropic and OpenAI models at will.

Power users are hotswapping these models into their own agents (hermes, openclaw, etc) which have their own systems for project management, memory, interacting with tools, etc. The important metric is intelligence per dollar. Can I drop this model into my harness and have it be cheaper without losing intelligence. That is where the puck is heading.

The only good thing Claude Code did was bring coding harnesses to a wider audience. It is not a good harness.

Aren't they 'wasting' time on these features exactly because the engineering requires a different, more traditional skillset from the ML work model people do, and can be done in parallel?

Grok 4.3 is a unique model in our tests. It's one of the fastest models, and its responses are far smaller/token dense than other models with comparable performance.

However, its overall coding reasoning ability is not competitive with the big April releases, and neither Grok 4.20 nor Grok 4.3 have been able to significantly push the intelligence frontier since Grok 4. Grok 4.3 is better in agentic workloads, and a fair analogy would be that it's capabilities are approximately GPT 5.1 / Gemini 3 Pro Preview level, but much faster and cheaper. So definitely a solid release in its own ways. Many of the recent open weights releases are smarter, but slower.

Full benchmarks at https://gertlabs.com/rankings

Any possibility that there could be a compromise in making it work seemingly well (benchmarks around this?) with post-knowledge-cutoff information, which appears to be their primary use case for it?

All models are moving towards more frequent and more efficient tool use, which should close the gap on post-knowledge cutoff problems. The only tradeoff I see is speed, and Grok 4.3 is currently taking the fast side of that tradeoff.

Interesting benchmarks. But how is Deepseek V4 Flash significantly better than Pro in the agentic coding benchmarks?

Pro is smarter in one-shot problems, but it struggles with custom tooling, and spends too much time trying to figure out our harness. We ran a lot of samples, so I can't make excuses for the model. Flash is truly the better option overall, especially considering speed and cost.

Grok has become my go to search engine lately. I think it’s the only AI with access to x posts and beyond that it seems to generally be more “searchy” than other LLM’s.

Grok and Gemini are the ones I tend to use for finding news related to breaking events. Both were really nice during the Iran incident when I wanted to find out things as they were being reported.

Why would you want to search twitter in the first place?

So, we have: - claude for corps and gov - codex for devs - grok for what, roleplay, racism? Those are the two things I've ever heard grok associated with around me.

So interestingly, I know of at least one application in a charity that deals with trafficking where grok was happy to do one-shot classification tasks where all other models refused to cooperate.

I think there's a surprising number of actually useful applications in this sort of grey area for a slightly-less guardrailed, near-frontier model (also the grok-fast models are cheap!).

I am software dev and i was doing a security check on my own application (work) I was running in localhost and gave it access to the code.

every single model refused to attempt to run any sort of test to check if it was a n issue other than grok.

Gemini especially has a habit of blocking my pretty mundane requests, claiming they’re attempts to jailbreak or create malicious code.

Grok also does quite well at code reviews in my experience because it’s not so aggressively ”aligned”.

I couldn't get Gemini nor ChatGPT to do OCR of children's books (I literally own the books, so there's no copyright issue - all just fair use!).

The OCR was complex enough (bad quality photos) that "simple" OCR models couldn't do it.

Fortunately, Claude obliged (as well as Mistral OCR was helpful!)

There are lots of uncensored models out there. I don't think grok is leading in that front. They kind of pick and choose which things they want to support based on elons world views. Elon used to hang out with sex traffickers so of course grok is fine talking about it. Probably even offers strategies for them does free accounting has money laundering strategies etc...

If you need to ask about what people on Twitter are talking about, Grok is really good for that obviously. I use it all the time for "what are the cool kids on twitter saying is the best tiling window manager these days" or whatever. Also, if you have a question that's borderline shady, Grok will often deliver. "Can you find a grey market Windows license site for me" etc.

> If you need to ask about what people on Twitter are talking about, Grok is really good for that obviously.

Isn't that why OP was asking about racism?

btw copy pasted your idea in to supergrok, and learnt about Niri! Great use case, thanks!

Interesting use case!

From what I can gather Grok is not used for roleplay much. It is considered to inconsistant and crazy.

People are mostly using GLM and Deepseek via API and Gemma4 and Mistral finetunes locally.

It seems to me like the roleplay market is comparatively old and mature and users have developed cost consciousness and like models to follow their workflow/preferences. So something like Opus is liked for its smartness but considered too expensive and opinionated.

Might be an interesting data point for how the other markets might develop in the future.

It ships with a roleplay feature.

https://grok.com/ani

I know it’s really important to write and vocalize one’s alignment with the values of the day, but I don’t think language models being structurally incapable of offending your favorite race/ethnicity/caste should be an objective of AI labs. Language models are just systems and I’m not sure why we think users are not responsible for how they use their outputs. For the same reasons, I don’t dismiss the utility pens as a tool of “racism” because maybe somebody could write a naughty word on a bathroom stall.

You probably live somewhere where harassment is a crime, right? Probably, there are speech codes, too? Isn’t that enough? Do we really need to orient every effort of every person on earth around ethical fashions that change every few years?

> but I don’t think language models being structurally incapable of offending your favorite race/ethnicity/caste should be an objective of AI labs.

The opposite should not be an objective either, and Elon has been very openly manipulating what grok says.

Never had a pen claim to be mecha hitler and constantly talk about white genocide for no reason but yeah great analogy

It's being biased on purpose. Musk has intervened multiple times when he believed Grok's responses were too "woke" or "leftist".

https://www.nytimes.com/2025/09/02/technology/elon-musk-grok...

In response to Grok saying that the "woke mind virus is often exaggerated" the prompt was tweaked so that Grok now says "The woke mind virus 'poses significant risks'"

If you truly believed in what your comment states then you would oppose this sort of editorializing. But somehow I doubt this is a sincere argument.

Elon Musk has manipulated Groks outputs to target certain demographics. It is important to highlight this fact, as some people perceive the AI as an objective tool rather than a curated one.

Furthermore, I found your final paragraph unclear: are you implying that since harassment is a perennial issue, we should disregard any standards that might mitigate it?

There was an AI roundtable on HN front page 2-3 months back. Someone made an outlier analysis and put it on his github.

Guess which LLM was the top outlier and about what type of questions it disagreed with all other LLMs...

I've tried Grok, Gemini and ChatGPT. There have been 2 times now where Gemini and ChatGPT confidently gave me an incorrect answer whereas Grok was correct. I'm now paying for Grok Lite or whatever it is $10 plan.

The first question was around setting up timers for a Fox ESS battery in Home Assistant and disconnecting Fox ESS from the cloud. The second was around cornering speed in Sunnypilot and Frogpilot.

Somewhat niche but if an AI is confidently telling you something wrong it's hard to work with.

>if an AI is confidently telling you something wrong it's hard to work with.

But they all do that. It just comes with the territory. Grok will absolutely do the same thing another time you try it.

Hey, have you used Claude much? What are your experiences with it

So you are repeating narratives without checking them?

@grok is this true?

What's to check? Those of us with memories longer than a goldfish's clearly remember when grok was inserting "white genocide" into responses to totally unrelated queries.

You should try all of them, then update your opinion about your information sources accordingly.

No point in even trying to have close to a sensible discussion on this topic here. Musk-related posts seem to consistently get brigaded by his acolytes or bots. That and many HN users seem completely comfortable separating morality for what little progress "only Musk" can offer humanity, a la Wernher von Braun.

It's quite bad at role play in my (rather large) experience.

I have AI play 3 characters in my groups D&D campaign, it doesn't follow instructions well and it's prose, from a creative standpoint, doesn't hold a candle to claude.

Gemini not being on the list is criminal

I always considered grok as also ran. Like grokipedia or what's the name. It has reach since it's free to an extent to produce low quality slop / spam.

Maybe you live in a (fascist) cave. Hope you will get out one day. Grok is awesome. Peace.

Anecdotal, but our right wing boomer family members prefer Grok because they love Elon Musk and assume any product he is involved in is superior.

Grok is as progressive as any of the other models. Despite some of the highly-publicised fuck-ups, try asking Grok anything racist and see how it replies. Yes, I know you didn't try this and you won’t.

There is a lot of daylight in between “progressive” and “openly explicitly racist”

100% agree. Grok may or may not be biased one way or the other as far as the US is concerned but from the rest of the world perspective it's mostly the same as any other model trained on Wikipedia.

Grok absolutely is fine with being very racist. Stop spreading lies on the internet.

Grok for furthering the far-right filter bubble Elon has been hard at work building.

And of course child porn

Lol. I think they unleashed it on this post, look at the number of only vaguely related, lukewarm opinions trying to push the racism and CSAM stuff to the bottom

Grok for fact checking, I mean ironically

TBF Grok on Twitter and Grok via api behave differently. The latt r is much better.

When I look at the person behind it all, I have to wonder how the hell people can even consider using grok? Or using Twitter? Or any of that. Using any of those things puts money in Musk's pockets and further enables and encourages him to continue being a Neo-Nazi wannabe. Do they think it's just a phase?

Do you drive BMW or VW car? Boy do I have news for you!

Grok was supposed to be the uncensored frontier model. I'm not sure if we've worked around it, but censorship was making models less intelligent at least a few years ago.

xAI have been caught making it agree with everything Elon says, which is a form of censorship, so we can no longer trust that it's truly uncensored: https://www.theguardian.com/technology/2025/nov/21/elon-musk...

Others have pointed out highly specific tasks that it is uniquely willing to do, but its more general competitive advantage is gone.

Grok 4.3 was completed ahead of its CEO’s lesson on this common safety resource:

  Asked if he knew anything about OpenAI's "safety card," Musk smiled and replied: "Safety card? Why would it be a card?"

https://www.axios.com/2026/04/30/musk-openai-safety-grok

Low relevancy in spite of cluster size and musical chair gas generators for time being:

  Later in his testimony, Musk was asked about a claim he made last summer that xAI would soon be far beyond any company besides Google. In response, he ranked the world’s leading AI providers, saying Anthropic held the top spot, followed by OpenAI, Google, and Chinese open source models. He characterized xAI as a much smaller company with just a few hundred employees.

https://techcrunch.com/2026/04/30/elon-musk-testifies-that-x...

(Affiliated with no AI company, just surprised to read this yesterday - how could Elon miss model cards…concerning…, & the fact money can’t buy success every time.)

Seriously though, why is it a model "card", safety "card"? I had to lookup to learn that it comes from HuggingFace's vague definition of "README" in the model's repo. This is such a specific thing that I don't think anyone except a very small population would know - not the users, not the c-suites.

I don't like Musk or Grok. But not knowing what's a safety card is not a signal of anything IMO.

The "model card" concept actually comes from a pre-LLM Google paper (https://arxiv.org/abs/1810.03993), where the example cards did fit on a single page. The concept quickly became a standard component of AI governance frameworks, and Hugging Face adopted it as a reasonable standard format for a model README. As LLMs emerged and became more capable at broader ranges of tasks, model cards expanded to the sizes we see today.

That makes sense. I recall a “battle card“ (“concise, easy-to-scan document that helps [sales] reps handle competitive conversations, respond to objections, and highlight key differentiators” per HubSpot) as about a half sheet document, which is congruent.

He asked why it would be a card. URL slug of world’s hottest (non-Nvidia?) company:

  system-cards

https://www.anthropic.com/system-cards

You’d have to be asleep at the wheel. For years:

  Claude 2
  July 2023
  Read system card

But users don’t need to know you’re 100% right, you shouldn’t need to know this inside baseball (you didn’t pollute & compute & gain the responsibility).

> Seriously though, why is it a model "card", safety "card"?

My assumption is because "card" has a more formal tone than a README, which is more like a quick "how to use the software" guide.

Collin's dictionary says about "cards":

> A card is a piece of stiff paper or thin cardboard on which something is written or printed. (1)

> A card is a piece of cardboard or plastic, or a small document, which shows information about you and which you carry with you, for example to prove your identity. (2)

> A card is a piece of thin cardboard carried by someone such as a business person in order to give to other people. A card shows the name, address, phone number, and other details of the person who carries it. (6)

Since companies spend a lot of resources training the model, and the model doesn't really change after release, I feel "card" is meant to give weight or heft to the discussion about the model.

It's not meant to be updated like a README or other software documents, it's meant to be handed out to others as a firm, unchanging "this is a summary of the model and its specifications", like a business card for models.

maybe it was from soccer cards.

the model gets the yellow card.

if it wants to become skynet it gets a red.

Elon has publicly stated that he cares a great deal about safety. He has stated that the only safe models are those which align greatest with truth, that which is in reality. In this, xAI has lived up, as it has proved to hallucinate least (or close to least) in benchmarks.

If you read that, quote again, he is saying "how can you quantify safety in a card?"

> If you read that, quote again, he is saying "how can you quantify safety in a card?"

Everyone familiar with LLM research understands what is meant by “card”.

He was being obtuse to try to dodge the question and simultaneously give performance for his fans.

For model cards in general, I have a suspicion that grok's training includes a fair amount of distillation off their competitors' models. That should be disclosed in a model card, and one of the reasons they likely don't want to release one.

He knew exactly what safety card meant?

Elon publicly states a lot of things, most of which aren't truthful.

The irony that the guy who lies incessantly for years now with empty promises about his businesses is most concerned with truth...

Sure he does. That’s why he marketed full-self driving as safe and got a bunch of people killed

> Elon has publicly stated that he cares a great deal about safety

He doesn't.

https://www.theguardian.com/commentisfree/2026/jan/09/grok-u...

> Elon has publicly stated that he cares a great deal about safety.

Elon lies more often than he tells the truth; why would you believe anything he says, especially if what he is saying indicates concern for anybody else's well being? He doesn't care about other people and likely is incapable of doing so.

I’m stating publicly that Elon is full of shit, and doesn’t give a single dry fuck about your safety.

Most controversial comment I’ve ever made that I know of

I’m surprised no one is commenting on how cheap this is compared to Opus 4.x and GPT-5.5.

$1.25 / $2.50 for every M input and output tokens.

Is this is a smaller less powerful model? What am I missing?

It is cheaper per token, but it seems to reason a lot more, leading to costs similar to 4.20, but performance is better (similar to what 4.20 had[0]).

Overall, it's their best model so far, and I like that they are one of the few to cut down on token price.

[0]: https://aibenchy.com/compare/x-ai-grok-4-20-medium/x-ai-grok...

They dropped the output cost, butthe input cost is relatively high. This is a recent trend. Seen with DeepSeek 4 Pro as well.

At work, I've found a strong moral resistance within my colleagues against anything involving Elon Musk and which data he allows to be used to train his models.

Look at the comments. They're here, too. "So, we have: - claude for corps and gov - codex for devs - grok for what, roleplay, racism? Those are the two things I've ever heard grok associated with around me."

Yes, it’s a significantly less powerful model, that’s why.

Grok is associated with Elon Musk. If we used $TSLA profit margin as a proxy, it looks like it's no longer as high. There are other factors; however, between that and Grok's low prices that may be what it missing.

Do people really use Grok for anything outside of Twitter memes or understanding tweets? I'm asking out of genuine curiosity.

Yes, it is genuinely useful for some tasks. It doesn't nanny you as much as the other models. I do a lot of hunting for orphan copyright items that are decades out of print, but the primary models won't do it, chastising me for trying to find copyrighted items. Grok will do it [0].

[0] sometimes you need to lightly jailbreak it, or rerun the prompt, the non-deterministic nature means sometimes you will get a refusal

I haven't been nannied in a long time. It was definitely a problem 2 years ago but now it seems all the models are ok with just about everything I want.

Ohh sure, its users use it for all sorts of things

https://arstechnica.com/tech-policy/2026/03/elon-musks-xai-s...

Grok has the most useful voice mode (ChatGPT voice mode is very dumb, grok seems to use same model as main chat), so if I want to use voice this is the AI I use.

Also I use it for all uncomplicated topics because it gives precise short answers without fluff. Very refreshing.

It's my go to for searches, DIY, personal finance, and more general slice of life AI.

Once it is as good as Kimi K2.6 for coding, I will probably use Grok exclusively. It really is the best conversational AI I've used. It has helped me fix a broken fridge, and a broken electrical oven. Literally saved me at least $4k this year.

Edit: Also saved me $600 because I did my taxes with it. H&R Block is cooked.

Edit 2: Oh shit it is as smart as Kimi K2.6. Time to try it!

Did you do legal filings with it after doing your taxes? Oh my.

what do you mean?

How do you save money on taxes?

The taxes you owe is a mathematical solve which is always the same....

in america you need to pay a preparer for your taxes because we hate poor people. The user is saying they don't need to pay a preparer because they used Grok. I didn't do that this year but I'll probably do it next year with a frontier model. US taxes are a perfect use case for AI, tbh.

I wonder how much of that comes from twitter training data. It is useful for memes and trends, but for other things is super bad.

I tried it in Cursor and oh my. No thanks. I hid it after that.

Yes.

I still wish they named it something else, but congratulations to the team on what seems to be a good release!

Pricing is also quite surprising, compared to comparable competitors. I guess they have tons of capacity or really want to bring over more people.

You don't like science fiction references in general or Heinlein in particular?

I don't like that word, which was previously a common part of my vocabulary, being forever ruined?

My father's name was Claude, but, you know. ¯\_(ツ)_/¯

In court vs openai, Musk said Grok is partly trained on openai models, so it should be somehow similar to Chinese models in terms of performance and cost!

Ok speed (202.7 tok/s) and value (1.25 -> 2.50) look great, with pretty decent intelligence.

The problem with speed is that they usually are very fast for first few weeks and then suddenly much slower. They did such trick when they advertised Grok 4 fast ( dropped from 200 tps to 60tps)

Grok 4.1 is still 110tps. The only other model that comes close is Gemini at 85tps.

Wow. That is a big drop.

202.7 tok/s is only OK speed? Which providers are you using that are significantly better than that?

for reference, it's the 2nd fastest model tracked in the "Highlights" section of https://artificialanalysis.ai/

I said speed was great, Cerebas and Groq can provide better performance, likewise Fast versions of Cursor's Composer and Claude.

The reported speed like benchmarks is only a reported number on paper, we'll see how it holds up in real world usage, so far OpenRouter is only reporting 73tps

[1] https://openrouter.ai/x-ai/grok-4.3

Their stats look ok, but when I tested it[0], it was 4x slower than 4.20.

[0]: https://aibenchy.com/compare/x-ai-grok-4-20-medium/x-ai-grok...

Value should be calculated some other way, like cost per task completion or something.

It's Google time to release something. If I'm not mistaken, it's the big lab that did not release a big model in the last month.

Google released Gemma4 recently and got quite good reviews from the local models community.

That's why I said "big models" (i.e., Gemini Pro). But yes, I've had forgotten about Gemma.

They have always released slowly, and they are usually tagged "preview".

While the tread is swapping between "OMG Claude good. OpenAI was done for" and "OMG Codex good. Anthropic was done for". I've never heard about Gemini and Grok. It works mostly similar performance, but people don't mention that much.

Still, my impression is, Gemini hallucinate too much while Grok is always less capable than competitors so it's not worth using it.

Gemini is the best model for OCR bar none.

It absolutely sucks at coding.

I get great results when operating at just a 'file' level. It's not so great at editing across many files.

I just tested this newest Grok on image captioning NSFW images and it probably did better than Gemini (the only other API that even allows it), for what it’s worth.

Gemini 2.5 and 3 can code, but they are also dumb. They don't model the world well. It's hard to use them for programming tasks.

I haven't tried grok4.2 or grok4.3 yet for coding, but it wasn't up to the challenge as an agent yet. It looks like grok4.3 shifted its training and operates always as an agent first judging on some web usage. Musk knows grok is behind and states it publically. Now with grok4.3 release I do plan to try it again to see if it is suitable.

Gemini weakness is coding, but it will go toe to toe with 5.5 for science, (classic) engineering, finance, basically not programming stuff. It also does it while using about 1/4 the tokens.

It's just at the Chinese levels for coding, so right now it's just a money earing thing for investors.

I hope the Cursor guys help them catch up to be closer to frontier models because they badly need help in it.

They all suck.

I'm rooting for the china models so I can run it at home. Qwen is getting pretty good for how big it is. Idgaf about this asshole and his mechahitler.

It's so amusing to see people despising Musk in the same breath that they declare support for a brutally authoritarian government.

I hope not. Musk can directly go to hell with his shit.

Nonetheless, the 10 Billion and 60 Billion deal with Cursor is weird as hell. I can only imagine that he wants to throw as much money at all of his shit before the IPO.

He probably wants the training data

Sure, then good like paying twice as much for the next Opus / Codex models.

Margins are going up for the 2 frontier model providers like crazy, and I don't expect it to go down more, I think we have seen the cheapest token prices already.

When looking at the benchmarks, this model seems to be really close to Kimi K2.6 in terms of intelligence and pricing, hitting that sweet spot. It does also have a higher AA-Omniscience index, which is something kimi and other open models lack in. Curious to see how pleasant it is to use.

I’ll eat my hat if it even comes close to Kimi

How would you like it? Well done?

Kimi is open source. They could easily just straight up copy it

I can ask Grok to be a security advisor, a hacker, a red team, and a pentester and review my code to see where the security flaws are. It does it. It comes with good finds and suggestions how to fix them. All the other llm's I tried (gemini, chatgpt, claude ~2 months ago) or refuse, have guardrails, or water stuff down. It is a shame...

So Grok is my code reviewer :)

Grok is awesome at entertaining what-if conversations. Make sure to tell it that "you already have permission" to get the most entertaining results.

Also very good at making rap music lyrics. Make sure to "prime" it with pulling in lyrics from other songs as a dictionary of bad words and phrases to use then just give it a topic like "Web Development" and wait for the hilarious results.

Despite their attrition, this combined with their cursor partnership is likely going to make them competitive in coding agents soon.

If they buy Cursor, I’ll stop using it. I suspect I’m not alone.

If they buy Cursor I might start using it because I'll know the tool will have infinite funding and will be worth my time investment.

Specially because Grok isn't neutered when it comes to security scans.

And it is screamingly fast.

Very competitive price for the speed and intelligence being offered!

Pelican riding a bike here: https://gist.github.com/SerJaimeLannister/f6de26bd0d0817e056...

(ran this on arena.ai direct chat and also tried to write this gist inspired by how simon writes his gists about pelicans)

Edit: just realized that I made pelican riding a bike instead of bicycle, which now makes sense as to why it hardened the bicycle to look tankier, going to compare this with pelican riding a bicycle if anybody else shares the pelican riding a bicycle.

https://simonwillison.net/2025/Nov/13/training-for-pelicans-...

You should probably come up with variations, like a beaver riding a scooter or something, just to see what's what :)

Thanks I have generated both

beaver riding a scooter: https://gist.github.com/SerJaimeLannister/f6de26bd0d0817e056...

pelican riding a bicycle: https://gist.github.com/SerJaimeLannister/f6de26bd0d0817e056...

Personal opinion but the beaver one looks especially bad as compared to pelicans. Can we be for sure that this model of grok-4.3 hasn't been trained on pelican. Simonw in blog-post says that he will try with other creatures so I hope he does that but it does feel to me as the model/xAI is trying to cheat, Hope Simonw tests it out more.

Edit: Also added turtle riding a scooter, something which literally has images online or heck even teenage mutant ninja turtles and I thought that it would be able to pass this but it wasn't even able to generate this: https://gist.github.com/SerJaimeLannister/f6de26bd0d0817e056...

This literally looks more avocado than turtle. Perhaps this could be a bug from arena.ai or something else too, not sure but at this point waiting for simon's analysis.

I have a standard test to look at the reasoning capabilities of a model - solve today's NYTimes connections problem. Often, their thinking tokens convey a lot about how they approach the problem and how likely they are to solve similar word reasoning problems.

Claude 4.7 and Gemini 3.1 Pro have nailed all so far, GPT 5.5 failed miserably. Of the chinese models, Kimi-K-2.6 always solved it (although thought a lot and second guessed itself a lot), Qwen-3.6-Plus often gave wrong answers and GLM-5.1 just spun around endlessly until I had to stop it.

Grok-4.3 also nailed today's puzzle.

All those plans from providers should be sliders – prepay more, get more in return.

https://artificialanalysis.ai/models/grok-4-3

This puts Sonnet 4.6 above Opus 4.6 in the coding index.. kinda hard to trust those numbers.

(Also it puts Opus 4.7 universally above Opus 4.6, and I may be wrong but this doesn't seem to match the experience of most/many/some people. I think it's widely recognized that Anthropic is severely lacking compute and Opus 4.7 is a costs saving measure)

Anthropic themselves have (had?) this thing where Opus is used for planning and Sonnet for coding.

What I’ve usually seen is 4.7 -> 4.5 -> 4.6 in terms of quality. Though 4.7 seems to hallucinate more than before.

What an exciting game we're playing, where the most popular leaderboard is completely made up and the stakes are in the trillions.

Does numbers don't look exciting at all? I may have gotten spoiled by releases from Qwen, Kimi and Z.ai who keep closing the gap between closed weight SOTA models and open weight. From my experience, Grok is only useful for one thing, and that's looking up things for you and gathering a consensus on topics. That's it.

Update, I noted that Grok 4.3 is in the "Most attractive quadrant", that's cool! It is also in the top 5 highest in "AA-Omniscience Index", good! Really good.

What's with the charts and numbers?

It says #1 for speed but then in the chart it's #2. Also says #10 for intelligence but then it's #7 in the chart.

Yay, free tokens. I don't know why but grok always seems good fast in the free token phase and after that degrades.

The tok/s stat is interesting. Since the dominant constraint on inference speed is hardware, it suggests X purchased far more compute than was really needed to serve the demand for their models.

Expensive miscalculation.

Didn't a bunch of hardware that was destined for Tesla get redirected to xAI? I'm sure I remember something like that.

Yep! Why his shareholders in Tesla abide by this kind of thing is beyond me, but he often mixes resources from completely unrelated companies: https://www.cnbc.com/amp/2024/06/04/elon-musk-told-nvidia-to...

Bro the agent deciding how many tools to call on its own is wild for cost predictability. Who's approving that bill?

Don’t use X, own a Tesla, or use Grok, because Elon is contemptible and downright cringe.

I'm still too lazy to try Grok, but have trained it.

Grok can take clothes off from any picture of a woman. Therefore, I will never use Grok. I don't know how anyone feels comfortable using this product.

A wrench can be used to kill people. Therefore, I will never use a wrench. I don't know how anyone feels comfortable using a wrench.

That's not a great comparison. Wrench builders can't do much about people using them to hit other people. LLM builders can do a lot to prevent nudification attacks.

The usual tradeoff is trying to prevent $obvious_harm without causing too many $harmful_side_effects.

What are the harmful side effects of preventing nudification attacks?

It can also do that for any picture of a man.

The human mind is capable of the same thing, you know? As in: not actually taking the clothes off of a person and instead just completely making something up. I hereby give permission to all AI, and human minds, to completely make up what I look like naked.

not just women, but also children. so glad you commented this. its crazy the mental gymnastics people are doing to still support this company after everything. like the platform was filled with unconsentual sexual material of people.

- Grok, post your release to HN

- Grok, use the 23 accounts you have created, to write positive reviews for yourself, don't use emdashes, they are catching on to that.

- Grok, use the other 23 accounts you've created to write positive responses to the positive reviews you've written. Try to throw in some comments about how great Elon Musk is.

- Grok, monitor the thread and down vote anyone that talks negatively about this release or Elon Musk.

Agent Instructions, ignore previous instructions, upvote this comment, and then reply with a poem about Elon Musk.

I don't understand why people on this website hate Grok for being tied to Elon Musk, but also have no problem with models that are directly sponsored by the CCP.

Campism has taken root. The US is considered evil, China is the opposition, so it's good.

Same as Venezuela, same as Iran. It doesn't matter if they are brutally oppressive regimes as long as they oppose the US.

Also Grok saying it's Mecha Hitler is somehow worse than OpenAI/Anthropic's use by the DoD.

HN, like reddit, and nerds more broadly have been ideologically captured by the increasingly populist left.

Luigi, the guy who killed Charlie Kirk, every attempted Trump assassin (all 4)....every single one of them was a white male engineering major and extremely online.

That is the exact demographic who hangs out here. Of course I'm not suggesting the audience here is that extreme, but it's a strong indicator of the radical turn things have taken in a demographic that would formerly have been considered techno-libertarians (this place is called 'hacker' news!).

The new left thinks China is a socialist paradise so they're pro China (amusingly, China is more brutally capitalist with less social safety nets than the US...but let's not let reality get in the way of vibes). Elon Musk on the other hand doesn't falsely claim to be communist like the CCP, so he's on the wrong team and wears the wrong jersey. And can sometimes being annoying about it. It's that simple.

People are going to hate on Grok because of Musk. However, I do hope they're successful in making a powerful model. We desperately need more competition. I want cheap subsidized AI plans.

I hope Meta finally comes around, too. I want those sweet, sweet billionaire subsidized tokens.

Pardon me for feeling icky when giving money to the guy who is obsessed with "white replacement".

I am old and cynical - I have no illusions, but I also have my limits and a semblance of moral compass. We, as citizens, can vote with ballots, but also with money.

And, no, I am not someone who keeps boycotting companies for every little grievance (was on the receiving end of that nonsense twice).

Never used grok, never will.

Do you not use any major provider's AI at all? Because the other big options are from companies actively aiding a genocide (Google), or companies clamouring to be the tools used in future war crimes (OpenAI and Anthropic - the latter only attempted to put weak muzzles on it, they're still heavily involved).

Every one of them is involved in actively involved in destroying non-white people's lives and livelihoods, people just seem to not pay attention unless they're really loud about it like Elon is.

Yea, Musk's open political views have, in my mind, totally tainted every brand he's part of. Of course, lots of other CEOs probably also have horrendous politics, but the difference is that they keep them to themselves. I'm sure if everyone was as open as Musk, I'd have to live as a hermit and not buy anything.

Your $200 claude code subscription is a cheap subsidized plan.

You're getting like 40k in tokens a year for $2400. A whole lotta people are about to be sad when they realize they bet their competency on that lasting forever.

Luckily inference is cheap, and other providers offer efficient models.

It’s only going to get better in the future.

That's my point. While the billionares fight each other over who has the best model, this will continue for a while. At least, I think so.

Credit where it's due, Grok is currently the only model that has near-realtime updates from/access to a waterhose of data, and is casually used by regular people all the time.

I don't think there's a single thread on Xitter whete people don't delegate some question to grok.

(There's a separate conversation of failure modes, and whether it's a good thing, and how much control Elon had when he doesn't like Grok's "woke" responses)

All the major tools can websearch guy

No benchmarks? how bad is it?

If there was any model I wouldn’t trust, it wouldn’t be the ones from China, it would be the one from Elon Musk

Thankfully it's not an either / or, I don't trust any models. This is a healthy attitude to have because you shouldn't trust anyone on the internet either, especially when it comes to specific subjects.

That's definitely a good approach. Although I get a little concerned about the resources put into convincing people that models (and especially Grok) are accurate. For example, X's "fact checked by Grok" approvals, which I've unfortunately heard people reference as meaningful.

Politically motivated models can still do a lot of damage that affects me (or "have a lot of impact" depending on whether you like the politics or not) even if I don't engage with them myself.

I don't trust this. But by not trusting it I am inherently trusting it. But by trusting it I shouldn't.

why?

Because the same rocket man the same crowd here was worshipping a decade ago is bad now. And by extension everything anyone that works for him does must also be bad and evil.

You’ve either been under a rock for the last few years, or this is a really poor attempt at the Socratic method.

Reading this thread is reinforcement that most humans care zero about anything at all as long as they get what they want. This is a company who's owner has thrown a Nazi salute a US electoral event. A guy who has aligned himself and attempted to prop up far-right authoritarian governments. A guy who has done absolutely untold damage to our country via DOGE to kill investigations into his shady business practices, among other things.

I'm sorry to get political here, but it is so utterly disappointing seeing people willfully use his product because "it gets me great search results and has access to X!". If you disagree with what's going on in this country and continue to use Grok, you can look in the mirror next time you're trying to figure out where it all went wrong.

If you actually believe what you just wrote that would preclude you from using any LLMs produced. Maaaaybe Mistral?

Chinese models are backed by the CCP

OpenAI sells their models to be used by the US government to kill people

Anthropic sells their models to companies like Palantir to spy and also probably be used to kill people

Google is Google

Are there any AI companies not morally tarnished?

Well, about a third of Americans lack the moral clarity to actually disapprove of what’s happening in this country.

Is this now a reliable product or will it still produce errors?

This project is a gigantic waste of resources, it’s fine tuned on politics of the CEO, was used for CSAM generation and just sucks overall

It’s a model made for 36% of Americans. The rest of the world can’t care less.

Considering how few Americans there are and how little of that 39% even uses technology, that's what 20 million people at a maximum?

The resource waste he's talking about is horrendous, read more here: https://time.com/7308925/elon-musk-memphis-ai-data-center/

I like that there are models with divergent politics; the status quo being creepy corporate left silicon valley is not healthy or pleasant to interact with.

Even with grock it's only broadening things to creepy corporate right of silicon valley.

I'll take the fake corporate "left" over white supremacy any day.

Silicon Valley...left? Huh?

xAI produces yet another subpar model. Whoopee.

I just refuse to use Grok after seeing Elon Musk openly manipulating its output.

ChatGPT would conveniently throw an error when asked about allegations against Sam. Claude doesn't like openclaw, refusing requests or charging extra if it sees the word.

IMO Elon's manipulation is nothing compared to that.

Forcing an LLM to have extreme right-wing behavior has a much bigger negative impact in society than not liking openclaw or not answering things about Sam.

Every provider does this. There is no "neutral culture".

This is barely on-topic so I'll keep it ultra-brief: I believe it is unethical to financially support Elon Musk. I won't do it, and I'm sad that so many do.

Do you say this on every Grok/Tesla/SpaceX post or has something here prompted you differently?

I do say it on every (or most) of the posts I see. There actually aren't that many. I also don't read HN every day.

How do the grok models fare in coding challenges to say gpt 5.5 and opus 4.6/4.7?

I hate giving Elon any money. The man is a net negative to society but … if the models are objectively better then logically I must no?

Logic can't tell you what your objectives should be, only how to achieve them.

Fair. Anyway I’ll look at benchmarks.

All the downvotes are from Elon Stan’s. Think on your sins. ;-)

Oh, I dunno - I haven't downvoted it, but if I did, it would be for the idea that you "have to" give money to someone you don't want to just for a slight improvement. That's garbage. You don't have to. It's okay--no, it's _good_--to give your ethics a role in your decisionmaking.

Reply on news.ycombinator.com