"Half the money I spend on advertising is wasted; the trouble is I don't know which half." -John Wanamaker
This applies even more strongly to model choosing. I know for a fact that majority of my work doesn't require a very strong model, but separating the trivial and non-trivial tasks is a famously hard problem (if at all decidable).
I don't see why it should be all that difficult. All you have to do is first find a library that implements a decent solution to the halting problem and you're off to the races.
I know people are working on smart model routers, but if you're going to be doing a task that's constrained enough millions of times, it can be worth it to simply give the AI models the actual tasks they're going to do and see how well they respond.
It's open source and nobody has tested it, so if somebody is willing to give it a try that'd be appreciated.
Exactly. I haven't reached the "let 1000 agents bloom" mode yet, so currently I'm spending real headspace managing agents doing work, and that work is all important, so why "settle" for sub-frontier models for that work? Maybe I'll get there for non-coding work.
HN could be run as a BBS on 70's hardware. Instead of using a CPU with ~10 thousand transistors, you're likely using one with ~10 billion to do basically the same thing, and you don't think twice about it.
> Starting today, GPT‑5.6 Luna, our fastest and most affordable model, will cost 80% less,
I don't have the words.
I genuinely thought we were in a stage where we were plateauing and going in for 5-10% improvements over months. Seeing spikes like this makes me question about where the floor really is.
When model intelligence reliably hits 90%-95% of current day knowledge worker tasks, they are going to burn those weight to silicon and we will see another 10X improvement in price/performance frontier.
The dynamic GPU clusters will be used for the 5% of tasks, and pushing out the frontier. Also there will be a set of knowledge tasks that are not done today (because they are too difficult for most knowledge workers), that will start being done in the future.
Burning the weights into silicon would be many orders of magnitude increase, not just 10x. It's kind of crazy that this hockey stick the AI hype bros talk about seems more and more every day like it might be real
https://taalas.com/ has done it already for a wildly obsolete model. 14000 tokens per second.
https://chatjimmy.ai/ is their interactive. Tiny context, very dumb, but absurdly fast. Imagine this as a tool call for claude code for trivial changes - the tool call from the harness takes longer than the execution.
Holy crap, I was not prepared for how fast it responded. I just wrote "Just wanted to see how fast you are! Can you write me a quick story about a tiger who lives inside a block of cheese the size of a house?"
For what its worth the frontier lab models can surely be a lot faster if they wanted them to be but theyre supply constrained so theyre doing stuff like multi tenancy. Since you cant self host them no one outside the labs really knows speed as a solo tenant
> It seems China is already able to do DUV a lot sooner than others expected.
That's the media and in particular US KOLs of all sorts driving the wrong impression of China and other places. China and many other places for example have fast public transport that the US doesn't and can't even imagine today. They're not behind.
China's DUV still isn't that production grade (mass produce-able) so don't get that hyped up the wrong way (in a different direction).
The whole China-is-behind with tech and in particular semi wasn't that they can't. The truth is they spent decades in internal politics and corruption. That all got solved with the bans, so thank the bans! Jensen even said the bans were bad.
To be fair we don't really know in terms of prices what's real and what's just investor subsidised attempts at market capture at this point. It could well be OpenAI's attempt to drown Anthropic while they've got the halo product if they feel they've got deeper pockets.
I wouldn't be surprised if they still had some margins since cheaper models are much harder to nail the accurate sizes off, and you still pay 2x for 1M context window.
But if this is even at 400B size it's insanity those inference prices, maybe 10-20% margins, if it's higher I would like to know is it their own chips or maybe they have accurately sized the model to fit on exactly a B300?
Could be a lot of magical things we can only speculate, but from here there likely isn't another 60-70% margin, like I have heard people claim, I would definitely be willing to bet on that.
Could still be a healthy 10-30% margin. Especially with Terra.
Yes. For example, third-party inference providers serve DeepSeek V4 Flash just as cheaply as DeepSeek themselves, if not even more so. This is very strong evidence that the low price of the model is not subsidized.
Hard to believe numbers. I don't mean that as a critique, but literally I am so impressed. Even if the model is a few percent lower for performance but is 80+% cheaper than competitors and is a US company hosted on US based hyperscaler clouds this is kind of a no brainer. Hard for most businesses to justify otherwise.
DeepSeek v4 Pro & MiMo v2.5 Pro (Opus 4.6 quality models for code) are insanely cheap for agent-driven work due to their super low cached-input prices ($0.0036/mtok) [0]. For Luna, the cached-input price drop isn't disclosed in TFA, but the pricing page puts it at $0.02/mtok, 5x more expensive.
[0] I am constantly surprised how much work pay-as-you-go with DeepSeek / MiMo will get done. I've barely crossed $2 each in a month of use.
Let's suppose each models was subsidized at 70%, so that we only pay 30% of the cost. They would loose much more money per token on the more powerful models. It's in their interest to encourage the use of the less expensive models. Let's say they increase Luna subsidies at 90%. They would still "save" relative to the use of the more expensive models.
High-performing open weight models being released recently, and your customers looking into working with multiple providers as a result, are a great reason to drop prices on your non-frontier offerings.
Although I'm sure there are some efficiency gains, the technology is too new and labs are scrambling to release too quickly to think that the low-hanging optimization fruit has been picked already.
Vera Rubin will be hitting racks very soon, and this is purported to have a 10x improvement in token throughput per megawatt. Of course, old chips don't get replaced with new chips overnight, but I don't think we're anywhere near the floor yet.
In a data center that is power constrained but not space constrained they could build out new racks and flip the power from the old racks. Wonder if this will lead to moderately used server GPUs on the secondary market someday.
Besides Nvidia Hardware is still sold out and super expensive. Not a single Nvidia consumer GPU got cheaper at all, Nvidia DGX Spark got more expensive too.
It will be swooped of the market the second it hits the market.
But yeah I do'nt want to know what Kimi 3 is pushing buttons inside Anthropic, OpenAI and Google.
Besides any floor: For every year the tokens get faster and cheaper, we will see new things like properly working AI factories which mimic expert teams. A lot more parallism as well.
As if ALL OF THAT doesn't represent inherent and crucial components of "intelligence" itself.
We are not purely rational creatures, thank God. Sometimes those "limiting factors" you listed -- stress, peer pressure, hormones -- are crucial elements of informing the problem solving process and arriving at a decision or a solution that actually works.
All an LLM can do is fulfill a prompt, no matter how misguided, backwards, or incomplete that prompt actually was.
"Go jump off a bridge." Hmm. Dying makes me stressed out. I'm not gonna do that.
By benchmarks, which sadly is a poor measure. Yes Luna is a good model under certain circumstances. Whether it is great for general usage is another story. Sonnet is definitely better when prompts are more vague and it needs to decide things. Luna generally sticks to things very strictly and goes off in bad ways.
Yes but there is a big, big market for subagents to consume lots of tokens cheaply and condense information up to parent agents. Luna would not be my choice for planning. But an explorer to comb through a codebase to find relevant parts? Or for enterprise retrieval, where it needs to search across many different types of data to see where to focus efforts for a smarter model? Or to wake up periodically to evaluate some conditions and determine if a bigger model should be spun up? Definitely.
I've previously found flash (for all the hate it gets) to be good for these kinds of things. Haiku was fine but it's ancient.
> Yes but there is a big, big market for subagents to consume lots of tokens cheaply and condense information up to parent agents.
That's again not some "intelligence factor" here. Different agents work for different use cases. Luna wins some. Terra wins some. Sonnet wins some. Flash was really good at exploring.
So I'm not sure what your point is? There's a big market for everything. Even within the market you describe it's likely not a Luna-size fits all either.
Sonnet and Haiku were already in an awkward spot, likely by design.
Anthropic's big marketing push this year has been entirely focused on getting people to use Opus via a Claude Code subscription, to the point that Sonnet is almost viewed as the poor man's alternative, and from what I've seen, almost nobody uses it.
Actually, here's an interesting project for all the vibe coders looking for their next front page post: scrape a ton of commits from GitHub with Co-Authored-By: Claude and figure out what the percentage split between Opus/Fable/Sonnet is. I'm willing to bet it's less than 10% Sonnet.
>figure out what the percentage split between Opus/Fable/Sonnet is.
This may be misleading, since I suspect many are using a blend through sub-agents. I tend to bias for Fable to orchestrate and Opus for implementation via sub-agents.
Opus 5 is not strong enough as the top-of-stack model, and feels idiotic after a week or two of heavy Fable usage, to the point where I'm paying for Usage Credits to keep using Fable rather than having to slum it with Opus.
They have no (other) equivalent to nano, so it makes sense that it’s much cheaper now. It may have been better, but it was also hell of a lot more expensive.
I typically do lots of mini calls for research (100s of millions or something in that ball park). Newer models made that absolutely impossible, and the fact that the older ones are starting to get deprecated made me switch to e.g. deepseek for some of my runs. We'll see if I move back after this.
I don't see how this follows. The cost of nails has fallen by 95% over the last century. It's because the cost of manufacturing has fallen. Not because they are selling the information of nail consumers.
Tokens are not normal software, because they have marginal cost, and I think people who are used to software economics really struggle with this. With token generation there really can be manufacturing cost efficiencies where one producer is just straight up better at serving product at a lower marginal cost.
That’s ridiculous. Every major AI lab is compute constrained. That’s exactly why nvidia is worth trillions today. If OpenAI had a single extra GPU they’d be using it to run another training cycle or experiment for their next model.
sama literally just said they wish they had bought more. the price drops are almost certainly due to good old fashioned hardware innovation (wafer scale with cerebras) and optimizing hardware development based on model architecture and inference costs. other inference providers will try to do the same if they can.
Making Luna, which was already very cheap and extremely capable, 5x cheaper is crazy. I use Sol at work but Luna at home, and while there's definitely a difference, it doesn't feel like night-and-day. After a year of ever-increasing prices it suddenly feels (between this, Kimi K3, GLM 5.2) that prices are falling again.
It's cheaper currently on many of the inference providers.
Personally, I'm having surprisingly good results with DeepSeek 4 Pro at home, which is very good value for money: it's not as good as Claude / GPT 5.6 (I have Co-pilot license at work), but it's still really useful for code reviews, validating thoughts, and especially designing / writing unit tests for new (and old before refactoring) functionality.
And it's very cheap per task. (Flash is even cheaper, but I've had issues with that on more complex tasks where it starts forgetting things and arguing with itself "but wait, let me read the function again").
Has anyone ever done a comparison between the smaller models like Luna, against the previous GPT 5 frontier models? Have we gotten to the point where the small models are as good as the frontier models of the past, or is there still a way to go?
This feels like the dialup->broadband transition to me.
I was already a huge proponent of Luna for things like deep research. Being able to run 5x more for the same cost is simply bananas. We are already running 10 parallel agents for hypothesis generation. I cannot imagine 50. The statistics become much more interesting & powerful when you can run so many samples of the exact same prompt+model without breaking the bank.
Very interesting. Can you share more about your hypothesis/research pipeline? I have been using Sol for those types of task because I figured you'd need more reasoning for getting good ideas, but maybe quantity > quality at a certain point?
Here is a rough approximation of the pipeline I use:
Phase 1 - Run X copies of Luna in parallel over the user's prompt. The purpose is to generate a diverse set of hypotheses.
Phase 2 - Run Y copies of Terra in parallel to investigate the hypothesis results, with each receiving them in a randomized order.
Phase 3 - Run 1 copy of Sol over investigation reports.
The goal is to ensure that the agent covers more initial starting points before presenting a final conclusion. If you only run a single copy of Sol and it hooks onto something wrong, it might not recover.
Anything related to reading and interpreting the environment seems to always benefit from the addition of more agents to the search party, assuming you have some rational way to synthesize their results.
Taking actions that mutate the environment is a different story. I think this is where you run into diminishing returns very quickly. You generally want one strong agent to act given the results of all the searching that was done. If the plan is clear, you don't need a genius model to execute it.
> The kernel work helped reduce the end-to-end cost of serving the model by 20%, while its experiments increased token-generation efficiency by more than 15%.
If the cost of serving GPT-5.6 just dropped by 20%, does that add up to literally billions of dollars in savings per month?
We know Anthropic spend $1.25 billion renting inference capacity from SpaceX (in two Colossus datacenters) from the SpaceX IPO, but we don't know how much of Anthropic's inference capacity that is (presumably a small fraction, since they were operating on top of AWS and other providers before the SpaceX deal.)
I've not seen any numbers that hint at OpenAI's per-month inference bill, but surely that has to be in the multiple billions of dollars as well.
lots of places, actually. not everyone wants to be attached to the Silicon Valley culture, and that line alone will guarantee practically any workplace. that person is going to find out what work-life balance is :)
With the right kind of credentials, it's not about need, it's about want. Flip the roles and let yourself become an object of desire, an aspirational hire.
In this case, and I don't mean this critically, I guess it would technically be, "Instructed model to find efficiencies... reducing inference cost by 20% saving company x billion dollars per month."
I have no doubt that further work was required to enable this, but it's still very cool to be possible to say that.
Smaller models are great if you are doing targeted changes in existing codebases. Don’t expect to use it for creating complex architecture from scratch or do major refactors. The larger the context, the greater the drop off will be.
You'd be surprised at what Luna can do, especially on xhigh or max. It's capable of working overnight, usually productively, just like Sol.
Haiku 4.5, on the other hand, is comparable to performance to Gemma4 31B (with working tool call formatting) in my experience, and Gemma4 strongly wins on vision and multimodal.
Isn't OpenAI burning billions and have billions more spending commitments? If they managed to downsize so much the cost they should have kept the price the same and become profitable, really weird move, unsure what led to this.
> In a compute-constrained world where model demand is growing faster than capacity
I don't buy it.
There have been recent weeks where some of the mid-level models (Hy3, Laguna M.1) are free (true for parts of June and July, see Hy3 in Cyan) . Even then the total token usage appears to be reaching a steady-state.
^ the first graph is tokens per week across all models
I guess we just can only throw ideas at an LLM at a certain rate.
I still have ideas and now I can have an LLM vibe code what I want, but I'm not going to let an agent just run unattended for longer than a few minutes or a few bucks for hobby projects.
So maybe it is a matter of lowering the cost of an LLM so I can let it churn for hours at a cost of pennies... But I suspect demand for tokens is very price-elastic.
These are free due to some different type of reasons like Nvidida sponsoring free tokens or the model companies.
My company checks the models and pays for Opus through AWS.
You still send the WHOLE context of whatever you want to do to a random endpoint on the internet. If you want to write a good email, you give that context your email address, names, the reason for it etc.
Big companies don't randomly use some random api endpoint to do so.
Anthropics quarerly revenue is still growing very fast. I don't think we have seen even the real potenzial of it yet at all.
Not only are still a lot of countries missing which do not even use anthropic or any other frontier model yet but also all the agentic based solutions enterprise companies are currently building on mass (at least in my industry)
What are your use case for these? I’m manly interested in coding where more capability is better - give me a 10x model at 10x the price and I’ll take it. A worse model at very low cost has no appeal to me. At-least not for coding. Translation maybe? OCR?
This is one of the things OpenAI has been focused on for an year or so that led to the doomed autoswitcher in ChatGPT .com (switching models based on estimated task complexity) that was quickly reverted
Whereas Google with Gemini 3.x, Anthropic with Fable etc are happy to just go for 'big model with dense params'
It's hard to guess from the outside of course but just this kind of talking points focus on GPU efficacy is what we see from OpenAI and Chinese open source labs more often than from Anthropic or Google Deepmind and this benchmark chart seems to concur
Didn't expect that. Luna pricing is crazy now. I don't think there is anything on the market that competes at this price-performance point.
For our production app, OpenAI clearly is the best provider now. Their API is very reliable and has many nice features. The price-performance of the model lineup is incredible. We used open weights model via Fireworks for a long time (e.g. Kimi K2.5). Fireworks is a great provider but we still ran into issues here and there (Same with Anthropic and Google). OpenAI just works, is fast and in my view has a better price-performance ratio across almost all levels of intelligence.
80% less for Luna is absolutely crazy, in my opinion we may reach a point in the next year where powerful models on the API could potentially be cheaper than subscriptions. Compute just keeps decreasing in price.
While I can't deny this is a huge technological result, and it's laudable they reduced the price because of it, 80% is really a lot. I can't help but wonder, is this because of the model's capabilities, or was the initial system just really sloppy? The public will probably never know the details
This is awesome. I’ve recently set up my opencode to use 5.6 terra for my main agent, who delegates work to a 5.6 Luna coder agent. So far it seems to work well, and reduce costs a lot. With this price reduction, it will work a whole lot better. Perhaps I can get my github copilot quota to last the whole month now.
I would pay significantly more to use these models if there was a legal contract that guaranteed they weren't ever terfing them and some way to prove that.
Might just resub. Will experiment with Luna next sessions.
5 h window is not working very well for me. But if I can drop down to Luna at 20-30 % left and comfortably ride out the wave then.. that might just work.
With this new price change, Terra does look pretty Pareto’ed by Luna.
On agentic coding, pairing Sol Medium for architecting with Luna High for coding does kinda make sense. But beware that architecting can be very read-heavy, and Sol is a bit read-pricey compared to Terra.
Experienced similar between 5.4-mini vs 5.6-luna in our own pipelines but after spending some time on prompt optimization and testing out various reasoning effort levels 5.6-luna was well worth it. Did you just replace model selection while keeping everything else in place or spend some time on evaling with newer prompts etc?
No we kept prompts as is, just swapped model. The prompt is already quite optimized for the task.
How would updating it possibly make a more intelligent model spend less tokens than a less intelligent model? Care to elaborate?
Most of the time when upgrading models we have needed to change prompts to get the same performance (let alone better performance). Usually, your prompt is overfit to the specific model doing the specific task. For example often your previous prompt is overspecifying and creating contradictions that a dumber model would just gloss over whereas a smarter model will try even harder to follow.
Seems like they're working to destroy the local LLM argument. Right now Haiku is $1/$5 in/out. You can grind out $12,000 worth of haiku (or arguably, sonnet) class tokens in about 5 months on a Blackwell RTX 6000 96GB especially if using concurrency. BUT, but, if you use a g6e.xlarge on aws it's now more expensive than buying tokens from OpenAI @ $0.20/$1.20. It also destroys "the Mac Mini argument", pushing the ROI to ~4 years.
The local LLM argument never really held water tbh. You can get surprisingly good performance for lightweight tasks locally, but you're just fighting economies of scale if you're going trying to beat a datacenter on cost.
The milking games have started. The billionaires want their money back.
Edit: Yes, 80% minus is still milking. Because you empower these greedy mega-corporations. Just look at the RAM prices increase, then you see that the more money you give these hungry dragons, they more they will eat up. Don't get fooled by their "less cost now" advertisement.
> GPT‑5.6 Luna, our fastest and most affordable model, will cost 80% less
Looks like the Chinese models are really making a dent. Having 3 different price categories with the "most affordable" one still costing more than GLM 5.2 never made sense.
I thought the chinese models were cheaper per token, but about the same or more expensive on tasks because they used more tokens for reasoning. Cutting even further, seems like a really big leap.
It all comes back to electricity cost. China has cheaper electricity so as long as China keeps pace there is no way for American companies to undercut them. Each boolean operation in China is cheaper than the one in America.
> China: Household rates average around $0.08 / kWh (¥0.53/kWh).
vs
> US: Household rates average around $0.16 / kWh, though regional variation is massive—ranging from ~$0.10/kWh in low-cost states (like Washington or Louisiana) to $0.30–$0.45+/kWh in high-cost areas like California or Hawaii.
Model segmentation & distillation like this that asks the consumers to pick exactly which version of the algorithm will solve their problem is evidence for lack of intelligence instead of its presence.
You really really don’t need to pick. Just use Sol on high. That’s my daily driver and I don’t touch the model picker at all.
Now, if cost is your concern, then that’s a problem in all of computing. Hence why I’m sending you short plain text messages using an iPhone with a many-core CPU and gigabytes of RAM.
it is really hard to know upfront if you have fuzzy task. sometimes i would choose a cheaper model and it will spin and spin with bad outputs ending up costing more had i chosen a more capable model.
This applies even more strongly to model choosing. I know for a fact that majority of my work doesn't require a very strong model, but separating the trivial and non-trivial tasks is a famously hard problem (if at all decidable).
https://github.com/JarJarBeatyourattitude/evalt
I know people are working on smart model routers, but if you're going to be doing a task that's constrained enough millions of times, it can be worth it to simply give the AI models the actual tasks they're going to do and see how well they respond.
It's open source and nobody has tested it, so if somebody is willing to give it a try that'd be appreciated.
HN could be run as a BBS on 70's hardware. Instead of using a CPU with ~10 thousand transistors, you're likely using one with ~10 billion to do basically the same thing, and you don't think twice about it.
Famously, this is also a problem for human coders in sprint planning.
/s
I don't have the words.
I genuinely thought we were in a stage where we were plateauing and going in for 5-10% improvements over months. Seeing spikes like this makes me question about where the floor really is.
The dynamic GPU clusters will be used for the 5% of tasks, and pushing out the frontier. Also there will be a set of knowledge tasks that are not done today (because they are too difficult for most knowledge workers), that will start being done in the future.
Google is already working on a similar idea but more "flexible".
https://chatjimmy.ai/ is their interactive. Tiny context, very dumb, but absurdly fast. Imagine this as a tool call for claude code for trivial changes - the tool call from the harness takes longer than the execution.
I pressed Enter, and the response was instant.
> Generated in 0.037s • 14,205 tok/s
This is unbelievable.
Blocking Fable for sure made it very politicl a lot sooner than i expected it to happen.
and because China already has massive problems of getting access, they are pushing it on hardware too like what Huawai did without EUV.
It seems China is already able to do DUV a lot sooner than others expected.
That's the media and in particular US KOLs of all sorts driving the wrong impression of China and other places. China and many other places for example have fast public transport that the US doesn't and can't even imagine today. They're not behind.
China's DUV still isn't that production grade (mass produce-able) so don't get that hyped up the wrong way (in a different direction).
The whole China-is-behind with tech and in particular semi wasn't that they can't. The truth is they spent decades in internal politics and corruption. That all got solved with the bans, so thank the bans! Jensen even said the bans were bad.
But if this is even at 400B size it's insanity those inference prices, maybe 10-20% margins, if it's higher I would like to know is it their own chips or maybe they have accurately sized the model to fit on exactly a B300?
Could be a lot of magical things we can only speculate, but from here there likely isn't another 60-70% margin, like I have heard people claim, I would definitely be willing to bet on that.
Could still be a healthy 10-30% margin. Especially with Terra.
[0] I am constantly surprised how much work pay-as-you-go with DeepSeek / MiMo will get done. I've barely crossed $2 each in a month of use.
Assuming the efficiency gains are real, I feel like something has to give, maybe worse quality due to aggressive quantization/kv cache compression?
Although I'm sure there are some efficiency gains, the technology is too new and labs are scrambling to release too quickly to think that the low-hanging optimization fruit has been picked already.
If you use Codex it's different, the harness has a lot to do with it and there's definitely been changes including recently.
Besides Nvidia Hardware is still sold out and super expensive. Not a single Nvidia consumer GPU got cheaper at all, Nvidia DGX Spark got more expensive too.
It will be swooped of the market the second it hits the market.
But yeah I do'nt want to know what Kimi 3 is pushing buttons inside Anthropic, OpenAI and Google.
Besides any floor: For every year the tokens get faster and cheaper, we will see new things like properly working AI factories which mimic expert teams. A lot more parallism as well.
We are not purely rational creatures, thank God. Sometimes those "limiting factors" you listed -- stress, peer pressure, hormones -- are crucial elements of informing the problem solving process and arriving at a decision or a solution that actually works.
All an LLM can do is fulfill a prompt, no matter how misguided, backwards, or incomplete that prompt actually was.
"Go jump off a bridge." Hmm. Dying makes me stressed out. I'm not gonna do that.
Luna is an extremely strong model.
By benchmarks, which sadly is a poor measure. Yes Luna is a good model under certain circumstances. Whether it is great for general usage is another story. Sonnet is definitely better when prompts are more vague and it needs to decide things. Luna generally sticks to things very strictly and goes off in bad ways.
I've previously found flash (for all the hate it gets) to be good for these kinds of things. Haiku was fine but it's ancient.
That's again not some "intelligence factor" here. Different agents work for different use cases. Luna wins some. Terra wins some. Sonnet wins some. Flash was really good at exploring.
So I'm not sure what your point is? There's a big market for everything. Even within the market you describe it's likely not a Luna-size fits all either.
Anthropic's big marketing push this year has been entirely focused on getting people to use Opus via a Claude Code subscription, to the point that Sonnet is almost viewed as the poor man's alternative, and from what I've seen, almost nobody uses it.
Actually, here's an interesting project for all the vibe coders looking for their next front page post: scrape a ton of commits from GitHub with Co-Authored-By: Claude and figure out what the percentage split between Opus/Fable/Sonnet is. I'm willing to bet it's less than 10% Sonnet.
This may be misleading, since I suspect many are using a blend through sub-agents. I tend to bias for Fable to orchestrate and Opus for implementation via sub-agents.
With an 80% reduction in cost that becomes a ridiculous outlier in efficiency.
You mean they increased the price and then cut it back and now it is amazing?
Luna had a price hike vs mini (its previous replacement). The cut now just puts it back in that ball park.
Not that this isn't good news, but what's impressive?
I typically do lots of mini calls for research (100s of millions or something in that ball park). Newer models made that absolutely impossible, and the fact that the older ones are starting to get deprecated made me switch to e.g. deepseek for some of my runs. We'll see if I move back after this.
Tokens are not normal software, because they have marginal cost, and I think people who are used to software economics really struggle with this. With token generation there really can be manufacturing cost efficiencies where one producer is just straight up better at serving product at a lower marginal cost.
This is very likely priced below recovering the cost of the hardware but still above operating expenses.
I have no idea either way but one thing that detracts from these threads is folks claiming things as a fact without evidence.
https://www.youtube.com/watch?v=XDB5beon4DY&t=4m20s
> it doesn't feel like night-and-day.
I see what you did there. :)
Personally, I'm having surprisingly good results with DeepSeek 4 Pro at home, which is very good value for money: it's not as good as Claude / GPT 5.6 (I have Co-pilot license at work), but it's still really useful for code reviews, validating thoughts, and especially designing / writing unit tests for new (and old before refactoring) functionality.
And it's very cheap per task. (Flash is even cheaper, but I've had issues with that on more complex tasks where it starts forgetting things and arguing with itself "but wait, let me read the function again").
I was already a huge proponent of Luna for things like deep research. Being able to run 5x more for the same cost is simply bananas. We are already running 10 parallel agents for hypothesis generation. I cannot imagine 50. The statistics become much more interesting & powerful when you can run so many samples of the exact same prompt+model without breaking the bank.
Phase 1 - Run X copies of Luna in parallel over the user's prompt. The purpose is to generate a diverse set of hypotheses.
Phase 2 - Run Y copies of Terra in parallel to investigate the hypothesis results, with each receiving them in a randomized order.
Phase 3 - Run 1 copy of Sol over investigation reports.
The goal is to ensure that the agent covers more initial starting points before presenting a final conclusion. If you only run a single copy of Sol and it hooks onto something wrong, it might not recover.
Taking actions that mutate the environment is a different story. I think this is where you run into diminishing returns very quickly. You generally want one strong agent to act given the results of all the searching that was done. If the plan is clear, you don't need a genius model to execute it.
If the cost of serving GPT-5.6 just dropped by 20%, does that add up to literally billions of dollars in savings per month?
We know Anthropic spend $1.25 billion renting inference capacity from SpaceX (in two Colossus datacenters) from the SpaceX IPO, but we don't know how much of Anthropic's inference capacity that is (presumably a small fraction, since they were operating on top of AWS and other providers before the SpaceX deal.)
I've not seen any numbers that hint at OpenAI's per-month inference bill, but surely that has to be in the multiple billions of dollars as well.
So 20% is a really, really big deal.
> reduced inference cost by 20 percent saving company x billion dollars per month
I have no doubt that further work was required to enable this, but it's still very cool to be possible to say that.
I was still using GLM-5.2 in my personal projects, but this just made Luna a very easy choice.
I bet it must be better in some contexts and worse in others.
makes it by far the best choice for most workloads that do not need bleeding edge intelligence (reminder: luna can be comparable to opus 5!)
Haiku 4.5, on the other hand, is comparable to performance to Gemma4 31B (with working tool call formatting) in my experience, and Gemma4 strongly wins on vision and multimodal.
- Haiku: 30 points
- Luna Medium/High/Xhigh/Max: 38/46/49/51 points
That's a massive difference:
- 30 points is Gemma 4 31B territory
- 50 points is GLM-5.2 (744B) territory.
I don't buy it.
There have been recent weeks where some of the mid-level models (Hy3, Laguna M.1) are free (true for parts of June and July, see Hy3 in Cyan) . Even then the total token usage appears to be reaching a steady-state.
https://openrouter.ai/rankings#top-models
^ the first graph is tokens per week across all models
I guess we just can only throw ideas at an LLM at a certain rate.
I still have ideas and now I can have an LLM vibe code what I want, but I'm not going to let an agent just run unattended for longer than a few minutes or a few bucks for hobby projects.
So maybe it is a matter of lowering the cost of an LLM so I can let it churn for hours at a cost of pennies... But I suspect demand for tokens is very price-elastic.
My company checks the models and pays for Opus through AWS.
You still send the WHOLE context of whatever you want to do to a random endpoint on the internet. If you want to write a good email, you give that context your email address, names, the reason for it etc.
Big companies don't randomly use some random api endpoint to do so.
Anthropics quarerly revenue is still growing very fast. I don't think we have seen even the real potenzial of it yet at all.
Not only are still a lot of countries missing which do not even use anthropic or any other frontier model yet but also all the agentic based solutions enterprise companies are currently building on mass (at least in my industry)
Your support bot.
Your research long running bot.
Your SEO Optimizer bot.
Your incident analyser bot.
Your personal assistent bot.
- lower input/output token pricing
- the cached token price is $0.0028/Million tokens, which is like 50-90% of tokens
Whereas Google with Gemini 3.x, Anthropic with Fable etc are happy to just go for 'big model with dense params'
It's hard to guess from the outside of course but just this kind of talking points focus on GPU efficacy is what we see from OpenAI and Chinese open source labs more often than from Anthropic or Google Deepmind and this benchmark chart seems to concur
For our production app, OpenAI clearly is the best provider now. Their API is very reliable and has many nice features. The price-performance of the model lineup is incredible. We used open weights model via Fireworks for a long time (e.g. Kimi K2.5). Fireworks is a great provider but we still ran into issues here and there (Same with Anthropic and Google). OpenAI just works, is fast and in my view has a better price-performance ratio across almost all levels of intelligence.
https://deepswe.datacurve.ai/ - (See the Agent Steps view)
Or is the output speed so much higher that it cancels out?
I don't see a lot of benchmarks that record actual time. But on AA, Sol on Low beats Luna on High for Time Per Task.
Haiku was already in a ditch.
But this is coming straight for the jugular of a ton of models on openrouter.
presumably it's a much bigger model
https://files.catbox.moe/csxl32.png
(2 cents to run AA index, score 40)
Looks like OpenAI broke the pareto frontier on the trust-me-bro benchmarks!
(One has to wonder if they used any of the neat tricks from the DSV4 paper :)
With this new price change, Terra does look pretty Pareto’ed by Luna.
On agentic coding, pairing Sol Medium for architecting with Luna High for coding does kinda make sense. But beware that architecting can be very read-heavy, and Sol is a bit read-pricey compared to Terra.
Edit: Yes, 80% minus is still milking. Because you empower these greedy mega-corporations. Just look at the RAM prices increase, then you see that the more money you give these hungry dragons, they more they will eat up. Don't get fooled by their "less cost now" advertisement.
Looks like the Chinese models are really making a dent. Having 3 different price categories with the "most affordable" one still costing more than GLM 5.2 never made sense.
> China: Household rates average around $0.08 / kWh (¥0.53/kWh).
vs
> US: Household rates average around $0.16 / kWh, though regional variation is massive—ranging from ~$0.10/kWh in low-cost states (like Washington or Louisiana) to $0.30–$0.45+/kWh in high-cost areas like California or Hawaii.
Estimated final electricity price for large industrial customers in energy-intensive industries:
USA 50 USD/MWh
China 68 USD/MWh
https://www.iea.org/reports/electricity-2026/prices
Now, if cost is your concern, then that’s a problem in all of computing. Hence why I’m sending you short plain text messages using an iPhone with a many-core CPU and gigabytes of RAM.