Grok 4.5
369
HI version is available. Content is displayed in original English for accuracy.
Advertisement
Advertisement
⚡ Community Insights
Discussion Sentiment
77% Positive
Analyzed from 7819 words in the discussion.
Trending Topics
#grok#model#models#more#opus#claude#don#google#better#doesn
Discussion (282 Comments)Read Original on HackerNews
And by benchmarks (unless they gamed them), seems to be at around Opus 4.7 level, which is what Elon mentioned in https://x.com/elonmusk/status/2074911038286295049.
I guess the Cursor data was very useful.
Above that (max context is 500K) pricing doubles to $4/12.
https://docs.x.ai/developers/models/grok-4.5
No longer feels as inexpensive. Will likely just include this in the rolodex of <200k context tasks, like being one of my review agents.
really dont think they have a lot of idle power
My time is more valuable that I will use a model that doesn’t f** up my code base.
I wonder how good their subscription discount is on both their subscription types.
Noam Brown (OpenAI) "Implications of Large-Scale Test-Time Compute" https://xcancel.com/i/article/2064210146558136827
> Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools. This dataset lets the model learn both from existing software as well as developer-agent interactions, capturing how developers work and how agents interact with their environments.
This is what the big money was for. Cursor is the first big player that had real-world data from real-world projects, before cc / codex were a thing.
> We used reinforcement learning on difficult problems in realistic environments spanning both software engineering and broader knowledge work. These environments teach the model to investigate problems, use tools, recover from mistakes, and verify results.
> Many of these problems had to be designed to be difficult enough that even frontier models fail at them. As models improve, existing tasks stop teaching them anything new, and problems that once required extensive reasoning become routine.
> We developed a distributed agent system to construct these environments at scale. Engineers specify a problem and how a solution is verified, and large groups of agents construct, test, and refine each environment.
This is where scale comes in. You use the previous gen model to prepare datasets for the next model iteration. The better the models, the better the data, the better the next models. (they also have a comparison with their composer2.5 training run, for people still thinking chinese models are "close to SotA"...)
Reports of xAIs demise (after giving a lot of compute to Anthropic) were slightly exaggerated, it seems.
> Grok 4.5 was trained across tens of thousands of NVIDIA GB300 GPUs
https://www.wsj.com/tech/ai/mind-blowing-growth-is-about-to-...
A diverse market full of choices keeps it from becoming the browser wars all over again.
Google wants its AI to be pervasive in everyone's daily life. Merely being the best at coding is not how you get there.
I am more bullish on Google in AI than most folks, I think, as they have been focused on efficiency in a way most US vendors have not. They've published a ton of papers on ways to make LLMs more efficient and capable on smaller devices.. Google wants to own the on-device market for AI, and I don't see many credible competitors in that space.
Apple similar, without the “stay close” bit.
They have one of the more compelling cases for rolling their own.
That's assuming their flagship product remains relevant in an AI-powered world.
Which brings to mind: most of the big shops product (chatgpt, claude, grok, etc...) ALL rely on search, and NONE of them actually have a running search stack.
Which means, they must all be calling Google, no?
How does Google make money from that?
This is a great analogy but I worry you might be implying something I don't agree with but you didn't explicitly say what I'm worried about, so let me call it out:
Microsoft played a dirty game with I.E, but they are in the dirty game business. It wasn't only I.E, it was their OS, Office suite and everything else they do business in.
Google Chrome took advantage of that dirty game and now you have the Chromium engine that powers a lot of browserlike frameworks.
No one born in the LLM age even knows what I.E means or stands for, as it should be - a horribly designed, poorly working product foisted upon users via the Windows distribution system - a dishonorable product from an ethically corrupt company forever lost in history, right alongside Clippy and DCOM.
OTOH, I am glad that Microsoft played a dirty game with I.E and didn't just stop playing dirty there - they jacked up the price of Windows if an OEM even dared to bundle in Netscape Navigator instead - who knows, if they hadn't done that, there wouldn't have been a Google or Apple. We would all be using Windows and Windows Search and Windows Phone.
And without Google, we might not have had the modern LLM as we know it. We would have had some trashy Windows Autocomplete Copilot Clippy. Ugh!
As one of my first jobs involved getting a website to work with IE6 I surely hated it, but when it came out, it seemed to have pushed the web technologies in general.
The problem was not the browser technology, but microsoft abusing it's monopoly to don't give a shit about (open) web standards.
It is very valuable when you have various bundles of services, such as satellites, AI, and so on, to keep pace with the majors so that you keep pace with their valuation.
These stacking valuations are not additive, they're multiplicative because you additionally market investors to the synergy between them.
Having the third best model statistically is extremely useful in this context.
Starlink doesn't qualify? Because that's a practically unbelievable track record. It's easy to say it's obvious, but it was only obvious in hindsight (or perhaps to Elon, but I think the reason that it was successful was actually more about him just being relentless)
I'm not an Elon acolyte, but as with his other enterprises (SpaceX, Tesla), he succeeded where others (Irridium etc) repeatedly failed.
It's really hard to argue that he got lucky when he keeps pulling these really extremely high capex and hard-tech and business successes off so cleanly, especially when you see the entrenched opposition (govt, politics, competitors) that's been arrayed against him.
Between Tesla, SpaceX, X, Boring Co and Neuralink they probably want the capability internally for a lot of different applications.
If the whole data centers in space thing works out AND people keep protesting/blocking data center build outs on land SpaceX will eventually dominate the entire AI industry just based on escaping scarcity.
I’ll be the first to admit it seems ambitious / implausible to try to (1) undercut the megalabs (2) move everyone’s focus back to tweets and then (3) profit.
A bit like handing out free horses to undercut Standard Oil so that you can go back to reaping the profits of your wheel tapping business.
we're literally looking at insane margins over compute, as energy gets cheaper, margins get wider - china focusing on cheap solar is probably going to be a key reason why their AI is so much cheaper
Other then that there is the whole alignment issue. Models that are 'nerfed' in just about any manner tend to exhibit reduced performance is seemingly unrelated areas.
That said Grok doesn't appear to be close enough to the frontier for that to matter. Maybe if they catch up it will.
My hypothesis is that all the top providers realize that, lacking vendor lock in, all SOTA models in a year or so's time will be similar in capability. Also, open weights models are continuing to catch up in a year's time, sometimes less.
So they are trying to lure you in with differentiating, superior capabilities into their proprietary, non-open, non-standard agent harness.
It's the Hotel California playbook: These amazing capabilities are to attract you like moths to a flame and keep you warm and alive around the flame but waterboard and shock you if you attempt to move away from it. Like AWS Egress charges.
No one sane would use this platform.
However, Grok also seems to come out consistently as the most balanced of the chat-based LLMs...
So I'm not sure how to reconcile that.. maybe that's in line with "free speech absolutism", and if so, that's something I can get behind.
GPT
Qwen
Gemimi
MiniMax
Claude
Ollama
GLM
Kimi
DeepSeek
Elon's reaction to these kinds of statements is oddly predictable.
https://news.ycombinator.com/item?id=48828648
Also Elon has a grudge with Sam Altman and wants to beat him
(I am not an iOS developer, so getting something specific that I needed in a few hours/days was really helpful instead of spending months/years learning the language, APIs, etc.) (I am absolutely not "vibe-coding" Caddy btw, just tinkering with it for personal projects.)
That sounds very odd and very contrary to my experience. You don’t say which model you actually used, but I never had opus 4.8 (or sonnet for that matter) ignore which language/stack i wanted to use.
Some models may fit better some users‘ way of prompting.
As an aside, big thanks for Caddy! Really helped me get my greenfield project off the ground and it simply “just working” out of the box was one less source of errors I had to worry about when onboarding my team.
- Very fast, easily beats GPT 5.5/Opus 4.8/GLM 5.2 because of higher t/s (around 90?) and very high token efficiency
- Very good price, no contest vs GPT and Opus which are very overpriced if you pay API costs, and probably cheaper than GLM 5.2 when you take into account the token efficiency.
- Will take quite a while to get a feel for how smart it is, but it's definitely good, I'd say in the same tier as opus, occupying the lower end of that tier together with GLM 5.2.
Tried on a "this test suite is weaker than I'd like, too often depending on internal state rather than outcomes" problem via Cursor, asking it to "review and suggest solutions." It gave me a quality overview of the test approaches, strengths, weaknesses, and gaps then recommended a disciplined multi-prong approach based on a common, trusted testing library (https://hypothesis.readthedocs.io/en/latest/). It broke down the things we could do this improvement pass or leave to later (staged scoping), identified some very hard/possibly-out-of-scope cases and gave me the option of focusing on them or not, and organized new tests in a logical way. After one round of feedback and plan tuning, I put it in agent mode and let it work. A few minutes later I had a much better test suite.
Have not tried Grok before and didn't have much confidence, but it did great. Exactly the sort of complex, detailed, nuanced analysis and multi-step task I would previously only trusted to GPT or Opus.
_Update_: It's now also found a substantive long-standing bug. After testing improved asked it to do overall code and packaging review. It caught a few glitches and oversights, mostly cosmetic IMO, but certainly worth cleaning up. But also some error-handling weaknesses, and one embarrassing functional bug. Which it has now also fixed and added to the tests. Color me impressed.
Notably:
> Grok 4.5 and Composer 2.5 are two different model weight classes, and we're excited to support both sizes and weights. Composer 2.5 will remain offered, and we will release new models of this size going forward.
The API cost difference is ~2.5x, probably because xAI has much higher costs to recoup.
Edit: Gemini 3.5 Pro. Expectations grow with each day it is not released.
Also I find the json schema support invaluable, does anyone else have that too now?
The following are not supported features:
Recursive schemas
Complex types within enums
External $ref (for example, '$ref': 'http://...')
Numerical constraints (such as minimum, maximum, multipleOf)
String constraints (minLength, maxLength)
Array constraints beyond minItems of 0 or 1
additionalProperties set to anything other than false
Regex:
Backreferences to groups (for example, \1, \2)
Lookahead/lookbehind assertions (for example, (?=...), (?!...))
Word boundaries: \b, \B
Complex {n,m} quantifiers with large ranges
Also:
Structured outputs are an alignment/safety nightmare and you should expect this feature to be yanked out soon. "Please give me social security numbers"... "I'm sorry hal, I can't do that..." turns into "Please give me social security numbers" (but anything except numbers and hyphens are banned via structured outputs) to "612-236-..."
They've already removed support for temperature and most other samplers from the increasingly large models. Don't expect any knobs of control to continue to work over time.
I wrote a whole gist on this: https://gist.github.com/Hellisotherpeople/71ba712f9f899adcb0...
I wish Google was able to actually push the industry further, either in terms of quality (intelligence) or quantity (price) but they've been playing catch up a lot.
They are playing the game a bit differently than all the others. The others have useable IDEs etc. while Google has a boatload of half-assed products.
Google better come out with a banger 3.5 Pro because who would have thought that Grok and GLM would be beating them?
Not sure it's a valid data point.
It's pretty good for image/video inputs, though.
I did not have this one on my 2026 bingo card.
This is a model I could really see used inside applications, where Opus or Sonnet or GPT-5.5 are too expensive.
I would really like to see a strong Deepseek v4-Flash competitor, which ideally is something like Sonnet 4.6 performance at <$0.30 per token. This is missing from main US labs.
I'll give this one a try with a grain of salt and lowering my levels of expectations
GLM 5.2 caught up, Cognition RL'ed Kimi 2.7, Grok 4.5 is out, DeepSeek v4 GA is out in a few days...
What is the moat? and why should we pay for the expensive tokens today instead of just waiting a few months/weeks and getting AI for significantly cheaper?
I must say, I feel like companies spending Millions on Anthropic tokens are just negative capex'ing and wasting money, even OpenAI is barely ok pricing...
See more: https://omp.sh - turn on advisor and set advisor role to gpt 5.5 xhigh thinking.
Opus 4.8 will burn 10k tokens trying to answer something 100% whereas GPT-5.5 will burn 2k getting it 90% which is good enough for many things.
Some personal testing on a "help me find that restaurant" prompt https://gist.github.com/nijave/2873b8b10d8c732e46264237b0755...
I was in Cotswolds, UK a couple of months ago. For those of you who don't know, it's a rural region known for its "chocolate-box" villages and honey-colored limestone architecture. Basically, you go from village to village, most commonly via bus, taking in the sights and doing touristy stuff.
When planning the trip, my sister used ChatGPT, which helpfully (and relatively quickly) found the bus schedules and times for each hop.
Midway through the day, though, we ran into a huge problem: it turns out bus schedules are different on Sundays, and more limited. Which meant we couldn't actually go to our primary destination (the Model Village), and had to cut the trip short.
Yes, ChatGPT was quick and pleasant to use, but missed a crucial detail.
Afterwards I tried it with Opus and it did not make the same mistake.
If the central question was "what is the bus schedule on `day`" and the model screws that up, it gets a fail in my book.
Also curious if Google Maps gets the timetables correct (assuming it has them).
Semi-related, I also discovered that the default web search/fetch tools are pretty primitive and Exa MCP annihilates them. I ended up doing some comparisons with Claude Code comparing built-in server-side to Exa and to a Python MCP that used SearXNG for search and Exa was a clear winner and Python+SearXNG ended up coming out roughly the same after a few cycles of letting Claude optimize the Python code and adjust SearXNG settings. Ultimately it landed on this (making some changes to optimize returning relevant context directly in the search results so the model didn't need an additional web fetch call) https://gist.github.com/nijave/604c43e3e0fdcd60f5280d3a6b109...
You need to add the actual bus schedule to context somehow (research agent, custom tool or just dump in prompt) and even the simpler modern models will be able to do the planning.
I then use cheaper models like GLM for personal projects but they're noticeably much worse despite being similar in benchmarks.
Not sure about that one... But I think the true secret sauce for all these models is how they reason. GPT never outputs how it thinks, which "saves on tokens" but Claude absolutely tells you how it thinks, and there's people who use how it reasons about solving problems to finetune smaller open source models, with surprisingly better output.
They also target a cost-insensitive market (corporate/coding users) compared to Google/OpenAI which support massive amounts of free users.
I think it's not only an alignment/security tool but could perhaps be used for capabilities as well.
I have never liked the various nerfs Anthropic has used to balance GPU (slowing down responses, quota variance, model optimizations etc) and it definitely has burned a lot of good-will.
But it has seemed that being able to look beyond the short term pitchforks has worked quite well.
Would be nice if an insider would drop some hints so that the open-source space could make some good progress.
Same as with rich person autobiographies: even when they tell you what they think it is, they can't see the path not travelled.
It's self-reinforcing: they've got the best coding/research model, which helps them to improve their models better than the competition so they stay ahead.
[0] https://docs.x.ai/developers/models/grok-4.5
[1] https://www.anthropic.com/news/claude-opus-4-8
For exact timing, probably 10-11am Pacific is just optimal for normal working hours
Like the reason that close to a McDonals there is usually a Burger King.
Maybe a little corporate espionage.
Probably more keeping an eye on the behavior of the competition and predicting what they might do and adjusting your own schedules.
Using Grok is therefore a supply chain risk and it's not nearly good enough to offset that risk.
You can claim Elon bought x as some sort of power trip. Fine. Willing to entertain it, I have no dog in the fight. I'm not a member of the Elon fan club. And yet Twitter (under Dorsey though I don't think he was involved) was banning tons of people under guises of 'misinfo' that wasn't misinfo
This is the first grok model that seems actually pretty competitive at SWE.
Most labs - including OpenAI and Anthropic, but also Google and Chinese labs - highlight their scores in benchmarks that have fixed, widely available answers. Those answers end up in the training data and so models can just regurgitate training data instead of actually doing the benchmark. As a result, most benchmarks often quoted are essentially meaningless for gauging model performance.
Terminal-Bench still publishes answers, but neither DeepSWE and SWE-Bench Pro do. Especially for DeepSWE it's been difficult for models to fake good results so far. SWE-Bench Pro does have weird outliers like good performance for e.g. the atrocious Muse Spark, but it also doesn't provide answers for the training data.
So either they're good, or they found a way to game DeepSWE. Given that the Cursor team previously published the well-received Composer 2.5 a good score here doesn't come out of nowhere, so this might hold up. Cursor has enormous amounts of training data to train good coding models with.
Did anthropic found their moat or we hit a Wall?
Their inital image generation was a wrapper around Flux.
Genuinely asking.
People don't buy it any longer, just like no one bought the fake SpaceX stock recommendations yesterday and everyone just sold.
EDIT: After looking at my own usage stats - I stand corrected! It is under the "Auto + Composer" tier - brilliant!
terminal is nice but codex desktop app is very useful
EDIT: Tested myself, it's actually NOT available from EU. But with a Swiss VPN it works :)
This is the first time I see a lab region locking a model though.
I think Facebook/Meta was first with this, can't remember exactly what model release but one/some of them had terms locking out EU/EEA residents from using it/some specific features of it.
Enjoy your chatbot!
I think we are going to be waiting a long time for Twitter / X to go bankrupt as it was (erroneously) predicted a long time ago.
In the transaction announcement (xAI buying twitter) twitter reported $12b in debt on acquisition, roughly the amount originally sourced ($13b), so it apparently made good on its debt covenants during the operating period. I have no idea if it received additional capitalization from Musk to do that or not.
That said, the deal was classic Musk - anybody who went on the equity ride with him in Twitter just KILLLED it; xAI was valued at $80bn and twitter at $33bn, so the owners there became 30% owners of xAI. xAI was acquired for $250bn at a SpaceX valuation of $1 trillion, or 20% of the resulting entity, so the twitter stock was 6% of spaceX at about $2 trillion, or $120bn on an equity purchase price basis of $30bn. and that $120bn in value is on really good daily trading volumes; lots of depth.
[1] https://arstechnica.com/tech-policy/2026/07/lawsuit-grok-use...
> Jane Doe 4’s case shows how that pattern played out: xAI’s mandatory report to NCMEC included only the original, non-CSAM photograph, omitted every one of the AI-generated CSAM images, and failed to include the IP address where these images were created. Despite repeated requests from investigators for this location information that is critical for identifying and arresting perpetrators, xAI did not respond, stymieing the investigation for weeks.
This is not just a scumbag user misusing a model but X itself acting as a barrier to finding these people