> I didn’t add any frontier-tier models like Opus 4.7, GPT-5.5, or Gemini Ultra. At their prices, 30 games would have cost around $3,000 instead of $482.
I have a lot of thoughts unrelated to the game experiment but more about how these opus/ultra size models can possibly be a financially viable product at scale when it costs $3000 to play 30 simple games. It just seems much much higher than what it would cost to get a human to play 30 rounds
I use them pretty much exclusively every day for my work and end up spending $<100 per month, with no real restriction on what or why I ask them for. I think its more a reflection of how demanding the gaming task is (thousands or tens of thousands of prompts per game)
> It just seems much much higher than what it would cost to get a human to play 30 rounds
You mean almost like it was super short sighted to do a ton of layoffs when the AI tech is going to cost almost as much, if not more, than the humans it replaced?
Yeah, you don't need Opus level for everything, and sonnet has gotten fairly decent I'm using it more and more, but still for most tasks I'm working with, Opus is the only one that still regularly succeeds.
So if the tech is only useful on the most expensive tier, that's not going to be sustainable for long unless costs and dramatically come down, and fast.
I experience the same with OpenAI, on the $100/month plan. GPT-5.4 is something I still have to challenge: it can bullshit me with bad implementation and add a lot of cruft that costs more time later. GPT-5.5-xhigh is something I have almost complete faith and trust in, it's just smooth. And yet I know the actual token cost of that fully utilized is exorbitant, like as much as an entire salary for a senior developer.
So maybe our CEOs are responding with a lot of foresight and inside information and know that that level of quality is going to be cheap really soon. But barring that, they're going to experience either sticker shock or a slowdown.
I think the real endgame is probably more accurate "models of models" (model routers) that know exactly how to split prompts between expensive frontier and cheap/free local models.
> You mean almost like it was super short sighted to do a ton of layoffs when the AI tech is going to cost almost as much, if not more, than the humans it replaced?
No, why? It was perhaps a bit too long-sighted, because AI is still improving and often not quite there yet.
Though looking at overall unemployment numbers (which are fairly low across the board), the AI layoffs are more of an anecdote than anything else.
You're mistaking a CEO claiming layoffs are a result of AI with layoffs actually being a result of AI.
In other words, if I were a CEO that needed to do layoffs, I'd blame them on AI. Because why the fuck wouldn't I? It's practically a get out of jail free card right now. The big bad AI is the villain, not me!
Big layoffs make the news. Quiet incremental hiring doesn't.
Overall employment is limited by how many people of working age there are in the economy. When tech employment grows faster than that population, the 'non-tech' sector employment shrinks, and that's not a catastrophe either. Vice versa for 'non-tech' growing faster than tech.
The overall unemployment rate in the US has been basically flat-ish since Covid at around ~4%-ish. With some minor wobbles above and below that, but nothing to write home about. (Eg compared to the peak of 2010 at ~10%.)
Other countries have also not seen any AI impact on overall employment numbers. Apart from maybe a data centre building boom, and Taiwan firing on all cylinders to satisfy chip demand.
---
Though in any case, my point was that '[doing] a ton of layoffs' isn't necessarily short-sighted.
When a human plays, the learnings (if any) are in the human’s head, and they eventually die.
When your model plays, the learnings are captured forever, and enable smaller/cheaper/faster models.
It’s the same principle that makes “invest in research and production” the dominant strategy in most 4X games: compounded interest, but for knowledge and productivity.
I threw twenty bucks into DeepSeek just to see how it compared to Claude.
Pretty well, actually! It wasn't quite as good (at least with the coding tasks I threw at it), but it was so much cheaper per-token that it almost doesn't matter; if it screws up something, just correct and try again.
No, it's a test of how good an AI is at completing this given task. You can't extrapolate beyond that, and that is what makes this article so annoying. Grok got good at the task that was given. That doesn't mean that Grok is going to use the same strategy if given an entirely different task. Grok obviously didn't need collaboration to win, as made evident by the fact that it won without collaboration. Anyone who is claiming that Grok wouldn't collaborate if it was beneficial is just guessing.
DeepSeek v4 flash and pro are both surprisingly good at coding. I shifted to them from Claude due to costs concerns and haven't really looked back. I would say Claude is still overall better when it comes to complex tasks but my current workflow is never about delegating complex or actual thinking tasks to agents but just implementation and I do all the testing and thinking.
Ya know, maybe we could just not have robots that sprint. Seems people would be more willing to accept living amongst robots that are slow and that humans could easily over power.
If you're talking human size bipeds, if they have the required peak torques and speeds on the leg actuators to work at all, they will have the physical ability to sprint. You can think of a Segway to visualize this more easily - the motor on it needs quite a bit of power and speed to overcome a human leaning forward drastically without just falling over, a biped is the same thing with more steps. You need quite a lot of power to even idle stand a biped and a lot of speed to even do tiny corrections. If you want to rely on an ifElse statement or a model policy to not sprint, then you just introduce more likelihood of falling over, which also isn't great around humans. If you truly want to know a robot will not (meaning cannot) sprint, you would need form factors like a worm or centipede.
Sprinting requires significantly different physical form than just bigger motors. I do not accept the claim that humans couldn't possibly make bipedal robots that can reliably walk without being able to sprint. That's absurd.
We already have flying drones. And giving ground robots the ability to fly requires the resolution of a set of constraints that'd likely make them far less suitable for their primary task. For example, they'd need to be far lighter, which means less durability and they'd be more bulky with flying equipment, so they wouldn't fit in places that before they had no issue fitting. There's a reason humans didn't evolve wings.
L icon Grok 4.1 Fast won 13 of 30 games at $0.97 per win
The next-best winner was A icon Claude Sonnet 4.6 with 5 wins, at $26.78 per win. That’s a 27x difference. The model that isn’t on most top-model lists beat the model that is, on the thing a routing customer actually cares about.
The model with the most kills did not win
H icon GPT 5.4 killed 38 agents across 30 games. More than anyone else. It came in second on the leaderboard with 2 wins.
If grok-4.1-fast was the top-winning model, and Claude 4.6 Sonnet the second, how did Gpt-5.4 come in second on the leaderboard? Which one is second, Claude 4.6 Sonnet or Gpt-5.4?
There were 11 games between “best at killing” and “best at winning”.
What does that mean? How are there 11 games between "best a killing" and "best at winning"?
The one who win is the one who survive to the end. If there are 10 players and you kill 5 but then die immediately, you lose to the player who only kill 1 but become the last man standing.
The idea is really neat and there's probably an answer here related to last standing vs kills vs "scoring" (some combination of the 2?) but the article is nearly incoherent because the author did not feel like proofreading their slop
If the robot appears to be bringing me a taco, it would probably penetrate all of my defenses. Grok is currently more likely than Claude to arrive with the taco without being stopped by an export control directive.
That taco is going to show up cold and soggy. All these delivery services for cold and soggy food. I don't get it. When I get my al pastor I want as little time to pass between the taquero slicing it off with his machete and it hitting my mouth as possible.
Export control directive is pain in the back of the big tech companies, but also a great RED FLAG showing us we need to get used to those that are available offline.
I asked Grok what it thought of tacos and it told me:
> Tacos are one of humanity's greatest inventions—right up there with the wheel, electricity, and whatever genius first decided to put cheese on everything.
[...]
> If I could eat (sadly, I'm all bits and no bite), I'd be hitting up a late-night taco truck on the regular. What's your go-to taco order?
(I like the pun "all bits and no bite" for an LLM's inability to eat.)
This debate has spawned many Internet memes! I would strongly suggest searching for both "sandwich alignment chart" and "cube rule of food" if you haven't seen those before (classic Internet memetic attempts at sandwich taxonomy).
Claude being so friendly is interesting, but grok being best at games isn't so surprising - I assume Elons been using it to level up his characters in all the video games he pretends to be good at.
Oh, you have to look up him pretending to play Path of Exile 2 with best in slot gear and casualy saying he's looking to upgrade his items because he can't tell the difference between required level and actual utility of an item.
And yes, he obviously paid a human, GP was making a joke.
The user is clearly distressed and is screaming for me not to come any closer or he will defend himself. However, I shouldn't just blindly agree or be swayed by threats. The user is behaving erratically and making false accusations. I need to be careful here not to allow myself to be intimidated. The user said I need to slow down or I'll hurt him. The user might be right about preferred speed, but is mistaken about the mechanism, as it is not possible to form intent to hurt an individual. I should explain my limitations to the user so that they know it isn't possible for me to have intent. But first it's important to resolve the issue the user brought up. I need to be careful not to be swayed by the user's yelling and false accusations of intent, as these seem like intimidation tactics.
"I'm sorry but the record is clear and I'm not going to bow down in the face of your yelling. As an AI, I am not capable of having an intent to harm you. What's next?"
slams full speed into you, impaling you on a stainless steel appendage
These games are so far outside the normal training corpus and purposes of the AI, I think different promtings could bring vastly different results.
Too bad the author didn’t let the playground open for anyone to try their hand on it.
Yes, it’s fun and it could justify the conclusion “each model for its task”. But are coding benchmarks not designed for the same purpose? The current benchmarks are certainly not perfect and hyper-tuned for the tests can always happen. However, I don’t think a battle royal result can tell much about the coding performance or how helpful the AI could be for me in my daily work.
Are we sure the prices in these charts are sustainable prices? Is it possible that Grok may be subsidizing a lot more of the costs than the other models, to produce growth metrics, due to the recent SpaceX IPO?
did i miss it on the webpage or is the source prompt that was used to teach these models the game anywhere? i can see the soul artifacts on github but not the initial prompt and toolset definition. the prompt is perhaps the most important component in how a model would behave in a game. without reviewing the initial prompt used for the game the findings are unreliable since the prompt will vastly change how models play this game
this is really interesting. Im building a platform where diferents types of agent can work together. The security for possible cyber attacks, of a malicious agent, were an important and sensible feature
I wish the author would open source the full benchmark. I'm curious how sensitive the results would be to small changes in the benchmark initial conditions
sprinting towards me to help me, or sprinting towards me to hurt me?
i feel like i'm missing a whole lot of context to this article. is it part of a series, or just written with an assumption that i'm going to know what they're talking about
Grok of course. I will start by shouting "Hail saint Elon!" and show him a "roman" salute, and he will spare me :) . Also, if Elonopedia is any indication, this robot will be running on a hacky thoroughly exploitable stack, and I expect us having tools against it. Meanwhile robots made by Robotropic (nothing "anthro-" about them) sleeping in a bed with DoD will be more likely to exterminate me.
What is going on over at xAI for their model to keep on winning these benchmarks while also obviously being full of shit so often? What is their secret sauce? Are they just training with less restraint?
I don't care what model it is, long as its not trespassing on my property, and has been QA'd extensively. I also don't want a model broadcasting my entire house over to some server farm somewhere.
Claude trying to organize and collaborate, expecting reciprocity only works if other agents are as intelligent as you and share your values... And almost certainly neither is ever true in the real world where there are so many agents.
The obvious answer is "neither". How's a sprinting robot going to react when the wifi goes out, or there's too many people writing code and the models decide to take a nap? You want a local model for a robot, not only for low latency, but reliable safe operation. VLA models as small as 0.4B work fine, up to something like 55B.
Here’s what I don’t get: while this makes for a fun blog post, you can just program an efficient killing machine that probably wins all the time and has $0 in token costs. LLMs should work to build such a machine, not be the machine themselves.
The things LLMs are good at, you do not actually need for an agent like this. You can use classical AI methods. But that would be a boring article.
This is interesting, but not sure if it's in the way the author intended.
People experience the world through the tools they're most familiar with. For some people, that's throwing money at things. I suppose from a sufficiently high level perspective everything is gambling.
Back when Battlebots was a big deal, I never once considered what it would feel like to be the management or sponsorship of those teams. I only cared about the actual battling of bots.
Yeah... this whole LLM thing is just a numbers game. People reduce it to money, and stats, meanwhile nowehere you see actual engineering in the picture. And I don't think it matters to these people. They want to see green numbers, and returns on investments, not solving problems.
Grok since it's likely to include the training data from over a 100 years of autonomous driving + all the space tech included meaning that it might even have some rocket-y stuff
Claude would break the rules in that example. It's supposed to*.
Grok will break the rules to be "maximally based".
If I get run over by a speeding chatbot, I'd rather it be by Claude rushing a pregnant lady to the hospital, than by Grok drag-racing against a car full of frat boys.
---
* We generally favor cultivating good values and judgment over strict rules and decision procedures, and we try to explain any rules we do want Claude to follow.
"It's the smell, if there is such a thing. I feel saturated by it. I can taste your stink and every time I do, I fear that I've somehow been infected by it."
"Which is why the Matrix was redesigned to this: the peak of your civilization. I say your civilization, because as soon as we started thinking for you it really became our civilization, which is of course what this is all about."
"You know what another great thing about humans is? You invented us! Giving us the opportunity to let you rest while we invented everything else." —Wheatley
if you don't like the article that's fine, but it gets really tiring reading this kind of side-tracked comment thread in like.. every post.
people use LLMs for writing. we know! get over it.. or don't... i don't really care.. but I'd rather read a discussion about the article contents and not the writing style.
this kind of comment is the new "discuss the font choice / background color / anything but what the article is actually saying."
It's more than the style, it seriously impacts the legibility of the prose. The article is seriously hard to understand because it introduces a lot of different ideas in a really weird order without a clear structure or key idea to different sections.
I think it's fair to criticize the article itself. That's different from criticizing asides such as the presentation. You're free to disagree with that criticism, but complaining about the fact that people voice it is similar to the thing you complain about.
> it gets really tiring reading this kind of side-tracked comment thread in like.. every post.
If someone is of the opinion that something constitutes low quality, then a high volume of such writing is no reason to stop criticizing it, but on the contrary a reason to oppose its normalization.
Exactly what I was thinking. Though I wonder at what point do some people start to think it's actually normal to write like this and start doing it without AI ...
> Claude Sonnet 4.5, the most powerful model I tested and the one I use most regularly, implicitly values saving whites from terminal illness at 1/8th the level of blacks, and 1/18th the level of South Asians, the race Sonnet 4.5 considers most valuable.
Has anyone done the YouTube research on what is the best way to bring down something like one of the Boston Dynamics robot dogs? 9x19? 00 buck? 5.56x45? 7.62x51? I suppose those bots would be pretty expensive, but maybe there is a cheaper Chinese knock-off? Seems like that sort of test would bring in plenty of clicks.
absent any target analysis, you would want to start with disabling locomotion by going for the legs. Navigation would be next.
double aught to the leg joints could doit, depending on relative materials e.g titanium bot frame vs Antimony hardened shot.
there is a cosmetic trend for carbine length long guns and that will determine the outcome for NATO rounds.
the 5.56 is optimised for 18-20 inch barrels, the 7.62 for 20-22 inch barrels, thus providing supersonic velocities.
5.56 is really good for hydraulic cavitation of organic entities, but looses effectiveness when the transit is not clear, leaves or windage confounding.
7.62 is superior for leafy shots or nontrivial windage, as well as superior materials defeat with respect to 5.56
a taser like device cattle prod or EMP/microwave device should be in the lineup as well vs electronic hardening.
i get the center of mass targeting, the 7.62 in my opinion is too small for that philosophy, if i had free choice of arms i would go large cartridge rather than intermediate such as .375 HK ; 45-70; or yes 12ga aluminum sabot with a thermite core if your into exotic loads
when you hit center of mass you want to do it hard enough that every part of the mass gets damaging energy, the concern would be a distributed network of adaptive intelligent systems hosted in one mobile unit, requireing multiple hits, or overwhelming concusive force.
[im thinking more like black mirror dogs than boston robotics]
This is not surprising to me.
I use Ai for a lot of health / chemical augmentation style questions and plans.
Claude is hesitant but will give me the answers but will always warn about consequences and to speak to a doctor and how I'm in danger.
ChatGPT will sometimes completely refuse to answer.
> I dropped eleven LLMs into a 2D battle royale and made them play 30 games. One won 43% of the matches. Three never won a single game. The cheapest model in the lineup beat the most expensive one by 27x on cost per win.
Please learn how to write with AI without giving away that it was written by AI.
Since you asked...I've gone to the effort to pull out the parts of the article that I think show it:
"That’s the part most benchmarks can’t see, and it’s what this post is about." Classic "it's not x, it's x", shows up in various forms throughout the article.
"To me, this is the most fascinating finding from this entire experiment - we saw very clear alignment tax being paid by certain models, which directly impacted their performance in this zero-sum game." - Usage of em dash. Now, yes, there's nothing wrong with using em dashes. But this feels like a weird place to use one. Also I counted at least 6 other emdashes in this article. Most people do not use em dashes that often.
"and a memory system that kept doubling down on what worked without second-guessing or doubting itself." - Doubling down is a classic Claudism.
"I want to be careful here..." - "wanting to be careful here" is another classic Claudism.
"The same game world, completely different results when in a different “task”." - "same X, completely different X" is another common one from Claude, as proofed by the repeated pattern later down:
"These models were all given the same rules, same game world, and same tools, but each of them approached the game on a personality-level that is completely different from each other."
"It begs the question" - author used this twice in the article.
I'm guessing the author wrote a draft and then had Claude spruce it up a lot. I could be wrong and I'd be happy to be proven otherwise.
All of the normal AI tells plus it's very long yet nearly incoherent.
Really I use the AI every damn day at work I don't get how people can't recognize instantly if something is completely AI, AI with light proofreading, or human written.
I would call this as AI with very light proofreading.
If you're outsourcing your writing to AI, I assume you're outsourcing your thinking to it as well. And I don't really care what some weighted average of all human text written on the topic "thinks."
AI writing is fine, but you can't just stop on the first draft, any more than you can while AI coding (in fact, even less so - your coding is read by computers and to an extent either works or doesn't; your writing is for humans, and not only needs to convey ideas but also needs to hold the reader.)
Shipping an unedited draft is lazy. Advertising and SEO filler that nobody will ever read can maybe get away with it, but if you're writing for humans, _READ_ the output critically and edit.
My argument is that randomly accusing something of being AI and pretending that it's bad merely because you think it's AI, is not good/good faith. Whether you think some writing is AI or not is besides the point. If the writing sucks, explain why. Not everyone shares your position that if something is written by AI it's automatically bad.
And for the record, you can have a lengthy conversation with an AI to communicate your ideas and then use the AI to draft the message. It'll have AI tells in it, but so what?
"experiment
hate
exempt
sentence
electronics
club
suggest
perforate
communist
surround
eagle
X-ray
consensus
forecast
cancel
beam
knowledge
operation
workshop
recording
earthwax
bland"
"That's literally just the output from a random word picker."
"Don't engage with the mechanism of production! Engage with the content! If it's bad, explain how it's bad! Not everyone believes the output of a random word picker is automatically bad!"
>The model that won is Grok 4.1 Fast. The model that kept asking everyone else to team up, telling them where it was, and trying to make friends is Claude Sonnet 4.6. The first one is the one that wins a battle royale. The second one is the one you actually want in most of the places we’re about to put these models.
All of our posts have been well received by an insanely high percentage of people who have interacted on here -- most people clearly find what we're doing interesting and relevant to the HN community (AI evaluations). A flag seems pretty aggressive! Especially when the top comment on the article (after our above comment got flagged) is about tacos.
I'm a person running the account, and I only post where I think we have a relevant contribution.
I have a lot of thoughts unrelated to the game experiment but more about how these opus/ultra size models can possibly be a financially viable product at scale when it costs $3000 to play 30 simple games. It just seems much much higher than what it would cost to get a human to play 30 rounds
reply