It seems to be extremely economical - 4x better reasoning efficiency compared to Opus while being priced at $2/$6. For comparison, GPT 5.4 is $2.5/$15, GPT 5.5/5.6 are $5/$30, Opus 4.8 is $5/$25, Fable is $10/$50.
Also, the cache hit pricing is 25% of the input pricing ($2 vs $0.50). Long agentic workflows are dominated by cached input. The US frontier labs typically have this at 10% of the input price, and DeepSeek/Xiaomi etc take it to the extreme 1% range (which is why those are cheap to run in real world agentic loops with dozens of toolcalls per run)
Just to add context (no pun intended), OpenAI also charge differently based on context usage with GPT 5.5 being $5/30 below 200K, and $10/45 above.
Anthropic have a fixed price regardless of context usage.
These per-token pricing schemes aren't directly comparable though since these models all use different numbers of tokens, even for input (Anthropic's recent tokenizer change generates 30% more tokens for exact same input), as well as for reasoning, and context/token usage also varies wildly by harness with Claude Code using 3x the context/tokens of Pi.
I guess it does increase their cost, or rather your share of their hardware depreciation.
AI serving cost is apparently mostly hardware depreciation rather than operating cost (electricity etc), and if your large context request is occupying VRAM for some fraction of a second then you are paying for the depreciation that occurs in that time!
e. H100 costs $20-40K to buy, with a lifetime of maybe 3 years, and will only consume maybe $2K in electricity if run 24x7 for those 3 years.
Think of the entirety of the context (the full thread of conversation, all tool call output, etc.) as one message that's been submitted to the LLM all at once just so it can generate the next token (which is just the next word or even syllable.) Once that token is generated it is appended to the context, and the entire context is once again used to generate the next token. Keep doing that until the entire response is generated. The larger the context, the more stuff the model has to pay attention to as it generates each next token.
To use an analogy: imagine your friend is the author of an unfinished book. They die with 19 chapters written, and on their death bed ask you to write the 20th chapter. Assuming you're up to the task, you can only do this well if you take the time to absorb the entirety of what's been written so far.
This is how LLMs work. Context caching is an optimization on top of this, but it has its limits.
Yeah but depends how you use it - with superpowers and it’s prevalence of splitting things into smaller focused subagents - this could seriously reduce costs …
I wish my company gave me more options than just using Claude to test these things out
The recent Databricks comparison has GLM 5.2 performing identically to Opus 4.8 on high effort, and some early Twitter reports (e.g. from the OpenCode developers) strongly favor GPT 5.6 Sol over Fable.
As always it depends on what you are using them for, and how you are using them.
I have a theory that xAI has one of the largest clusters but with far less traffic + tokens to process bc its less popular than its competition, and xAI can pass the savings on to the end user.
They already invested in the massive datacentres of GPUs sitting idle. They have fewer users so they can deliver more inference per user - more thinking, larger models.
There are various difficulties with renting GPUs, especially if your setup is very custom.
The competitor would have to port their training systems to your specific network architecture, system design, rdma Vs ethernet vs infiniband Vs nvlink etc.
Getting it running might not be too hard, but getting it running efficiently and making good use of all those flops will require considerable human effort and wall time.
Add that to the fact most frontier labs seem to have a single huge training run - and to my knowledge nobody has figured out how to distribute that training run between data centers effectively.
I'm not saying this as a hypothetical -- I believe xAI is already renting their GPU farm out to frontier labs. Whatever logistical challenges exist, enough of them seem to have been overcome.
So where are these mythical savings coming from? You're saying they have spent more per user therefore can charge each user less or something? I'm not following.
The (optimistic?) take is that xAI is genuinely better at building datacenters at scale than anyone else, and the freedom to use Nat Gas as the primary energy source allows them to have lower marginal costs.
The (pessimistic?) take is that they have loads of idle GPUs and want to get some revenue out of them rather than none. Compare this to OpenAI/Anthropic where every token used by a consumer has to compete with enterprise spenders, and there’s not enough to go around for everyone.
It's also sensible for them to provide a cheap, intelligent model to users if they have capacity, then once they built a user base, tighten the screws. All the other AI providers have done that.
SpaceX, like Tesla, seems to have the same "portrayals over profits" mindset investors. So it doesn't even really matter whether or not xAI is making any money.
Profitability is never a constraint for Elon companies. He has always been able to be able to extract money from the middle east, government, banks, retail investors (or these same parties through his other companies) whenever they need more.
His net worth is orders of magnitude bigger than the cumulative profits his companies have ever produced (even if you only count the profitable quarters)
It's really easy to do this actually. You just create cars that drive themselves and rockets that land themselves and people start throwing money at you.
From your local Tesla dealer. Technically, you do still have to "supervise" it, which basically means making sure the little camera that watches your eyes doesn't catch you with your eyes off the road too much.
My sister-in-law's mother drove one from Florida to the northeast without touching the steering wheel or pedal/brake, right down to the parking at each end.
I've been a humongous fsd sceptic for a while, but had to lay that aside after I went for a test drive (test ride?) in one of these things.
What that technicality means is that you are liable when the car kills someone, not Tesla. Level 3 self-driving is a completely broken idea. People cannot closely supervise a process that never requires their input - when you suddenly are needed you will not be prepared to respond quickly enough. Either you are driving the car, or you aren't. If you are liable, you are driving it.
Companies are only going to do Level 5 self-driving unless they can own the vehicles.
You will never get a company to accept responsibility for crashes with hardware owned external to the company. I mean there's all sorts of things you could maliciously do to break hardware that you personally own. This is especially the case with something like Tesla where the people who absolutely hate Elon already go to extreme lengths to create news stories to attack him with.
There is just no advantage to the SAE level nonsense. It doesn't even make sense from a technical perspective as vehicles do not function based on "ODDs" (Operational Design Domain) that the SAE levels rely on for their definitions. The SAE levels were created by people who don't understand automated driving vehicles.
It doesn't matter what you call it. FSD is a system that can mostly drive itself, but it has not been certified to drive autonomously, and it cannot safely do so. Humans cannot supervise such a system. Has Tesla resumed reporting disengagement data publicly or to ANY government agency? Why not?
> You will never get a company to accept responsibility for crashes with hardware owned external to the company.
This is a completely bullshit argument. For one thing, Mercedes-Benz has already done that. For another, they already have all kinds of liability regarding the proper functioning of the vehicles they manufacture that their customers own, and they manage it very comprehensively. If there is a failure due to a manufacturing defect, do you think they have no ability to determine or litigate the owner's maintenance or improper modifications that may have contributed?
The reason Tesla can't assume these liabilities is because the technology does not support them. It is not safe, and they have no shortage of Elmo stans who will use it anyway and suffer the consequences of their God Emperor's hubris.
Wouldn't be bad if the product hadn't been paid for in advance. If it can't be delivered in working form for this car, that's not the end of the world. Tesla can just refund the undelivered product, with interest.
Now if they could have an "equivalent" to Claude's $100 plan with similar compute limits. I have the $40 a month version of Grok and I get a max of like 8 hours of "non-stop" Grok Build coding, per month.
You get "Grok Build" (the CLI) that uses the Cursor and/or Grok 4.5 models when you buy SuperGrok, which is like $300/year. I don't know if there is a feature-to-feature comparison by anyone on this, but you can get access to these models with unmetered tokens with SuperGrok.
give it a structured plan and it it does really well compared to similar priced models. I'd never use it for anything that required heavy reasoning and it's not built for that.
I think we need to be explicit about the domains we're applying Composer 2.5 to in these discussions.
I mentioned here (https://news.ycombinator.com/item?id=48766275) how poorly it handles my specific use cases. My coworkers in DevOps and frontend UI swear by its cost-effectiveness, whereas I strongly prefer the reasoning capabilities of Opus 4.8 and Fable 5.
Composer 2.5 seems to be SOTA for Helm charts and React/Vue, but, for my usecases it absolutely struggles spectacularly when tasked with rigid body dynamics or kinematic logic.
Composer 2.5 is so underrated IMO. I built a really feature rich application, insanely complicated, close to 200k LOC since it came out and for the most part it ran like a champ. Only used CLaude a couple times to get it unstuck. 8 hours a day and I'm paying about 30 a month.
> I built a really feature rich application, insanely complicated, close to 200k LOC
If you listed it, how many features/LOC or vice-versa? Really hard to know if 200K LOC is good or bad, at the surface it sounds like too much, but I don't know what the application was either.
It’s a fantastic signal processing / engineering app. There are 5 major players and this app isn’t quite as good but it’s in the ballpark. I’d day when I release early next nonth this will be the biggest fully featured vibe coded app I’m aware of.
I currently just use a Mac Mini M4 with cursor. I had 30 years programming before this (in my mid 40s now) so I am not a great person to give vibe coding advice because I have so much actual coding m. I’m pretty sure I couldn’t have done it without knowing how to code because there were some times even the best models would get stuck and I’d have to figure it out myself.
If I had to give advice I’d say to just do it. Work on some project that interests you and go for it
Not my experience at all. I've been using Cursor hardcore for about two weeks now and Composer 2.5 and it's wonderful. Now with Grok 4.5 I'm quite excited about the possibilities.
Cursors whole moat is their harness. Other people have benched opus and GPT models through Claude code, codex, and cursor, and cursor came out on top everytime because of their harness.
How does it compare to Chinese APIs? It doesn't seem like xAI is meaningfully more competent or any single bit more honest than Chinese labs anyway, so you might as well send tasks straight to China unless theirs is substantially cheaper.
Grok 4.5 is a huge step up from their next best model and now around the same performance as GLM 5.2, but it's not exactly at the frontier of the cost efficiency curve in our coding evaluations. That curve is defined by the 2 lighter GPT 5.6 models.
However, the fact that they finally have a strong post-training and RL setup bodes well for future releases. They certainly are not compute-constrained anymore.
Sonnet 5 is a huge token hog, though, it uses far more reasoning tokens than Opus models while being priced at $2/$10 with promo, and $3/$15 (usual Sonnet price) afterwards.
I'll probably get hate for it, but I was not impressed by Fable, I felt like it was just Opus with more tokens for thinking. I feel like the second I turned on Fable I drained my usage more quickly, despite them billing it as though it were Opus level of usage. The value is just not there for me. I wish they could make Haiku remain low-cost and drastically more capable to the point you could use only Haiku.
I'm not sure if you are aware, but you have to approach prompting Fable slightly differently from a model like Opus.
It's important to include the reason aka the why of your task [1] in your prompt. You'll get more mileage if you verbalize your thought process when prompting Fable. Anthropic say you should think of Fable as a "thought partner".
Some things require skill to use most effectively. It's fair enough to consider this a failure if the thing in question is "making a phone call", but when it's something like "getting an AI system to do a good job for you" this is not a reasonable thing to make fun of it for.
It's like...
"I wrote a program, and it segfaulted instead of printing out a list of prime numbers." "Yeah, look, you've got an off-by-one error here." "You mean I'm holding it wrong?"
"I'm trying to play the violin and it's making horrible noises." "You want to change your grip on the bow like this, and be more careful in where you put your fingers on the strings to get the right notes, and there's a whole art to how you adjust the speed and pressure and so forth to make it sound good." "You mean, I'm holding it wrong?"
"I'm managing a team, and one of the people on the team doesn't always do the things I tell her to." "Maybe you should sit down with her and see whether somehow your explanations of what you want aren't getting across, or whether she feels like you aren't treating her with the respect and dignity she deserves, or whether she's bored with the work, or etc. etc. etc." "You mean, I'm holding it wrong?"
Yes. In the second case you're literally holding it wrong. Some things don't work as well when you hold them wrong and it's worth some effort to learn to hold them right.
I hold no particular brief for Anthropic. I don't know whether Fable is really much better than Opus or whether the alleged improvements are all just pareidolia or something. But "getting the most out of this immensely complicated thing that's in some ways kinda like another human being can be tricky" doesn't seem to me like an implausible proposition, and if it's really doing something akin to human-like work[1] then it's not unreasonable if you have to approach working with it in something a bit like the ways you approach working with other people.
[1] If it isn't really doing something akin to human-like work, then why are you bothering with it at all?
Did you explicitly tell it to use Sonnet or Opus subagents and stick at or below high effort? Asking because such practices make a huge difference in the quality of output and the amount of tokens burned. I used one of my accounts to explore ultramax and it was just a token hog that might be worse than Opus.
Yeah, that's what bit me. Even Anthropic's own documentation seems to indicate that Fable is not all that great as your go to model for tasks. What it seems to excel at is a sort of leadership role because it proactively keeps all the subagents in check.
If you're not explicit in the prompt or haven't configured your environment then the default behavior is to use subagents that match the host.
You can write a skill (many have) to lead it in how to use different models and efforts for subtasks. For searching the code base, for example, I have it use Haiku, which is fast.
I felt the same tbh; I notice more the regressions in the weeks before a new release than any potential improvement the new model might have actually brought.
It may also depend on the workload. At work everything is very domain specific with barely (if any) public training data; both need thorough review and careful hand holding, meanwhile at home Fable is scared of libtorch and falls back to Opus even if it's not touching the ML parts.
But very expensive compared to Deepseek v4 Pro, which performs similarly.
Grok is stuck in a difficult place - not the best model at anything, and not the cheapest either. It's hard to make a case for using it on any dimension, even before you factor in the history (I'm not sure suggesting the company uses the model that refers to itself as "MechaHitler" is the way to a promotion).
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.