One of the strangest things in the AI industry is 'tokenomics'. It's not very clear why using GPT-4 in early 2023 was so expensive and then six months later 20 bucks could get you a fair amount of GPT-4 inference. This pattern has continued across various labs/providers for years--there is a continuous see-saw of pricing that doesn't seem related to anything.
So what open weight models do is at least provide a baseline of inference cost to add some sanity to the price markers. And of course predictability too--if you really want Kimi K2 instead of K3 you can still use it.
So the competitive pressure and predictability offered by open models is helpful for users
> It's not very clear why using GPT-4 in early 2023 was so expensive and then six months later 20 bucks could get you a fair amount of GPT-4 inference.
> if you really want Kimi K2 instead of K3 you can still use it.
I think this is a very important aspect, especially after the huge GPT-4o backlash when GTP-5 came out. Each model has certain quirks, and areas where the previous model might be better for some use cases than the latest and greatest, and the labs so far seem to have no desire to offer some kind of "LTS" release.
Did they actually ever cut the price on gpt 4? The oldest versions of it in the api still seem stupidly expensive? There were definitely price cuts as they introduced the turbo models and stuff, and new versions of each model might have gotten pricey cuts, but just because they're both called "gpt-4 something" doesn't mean they're the same under the hood or that they didn't change a bunch of stuff under the hood to make it cheaper to serve
Drug companies have a portfolio of compounds they research. Most don’t pay off, so R&D costs make their way into the pricing of those superstar and other drugs that do work. Also, timelines are pretty long.
In that light, the entire market itself looks essentially artificial, given drug manufacturing couldn’t exist without government guaranteed property rights, which are themselves a kind of monopoly on use.
But yes, the reason brand name drugs are drugs are more expensive than generics is due to intellectual property, both the patent and the trademark.
1. the field was nascent and new efficiencies were discovered
2. supply and demand
3. its in the company's incentive to make their models more efficient to increase overall usage so that while the margin remains the same, the total revenue + profit increases
Yeah I've been thinking about this and the analogy I came up with is that tokens are basically equivalent to an in-game currency in -free to play" mobile games. You're trading actual money for some notional "curency" or coins that can only be used for one thing but, unlike mobile game coins, you don't know how many coins something costs before you use them. It's kinda weird.
Because as per usual it's silicon valley misunderstanding economics. AI is HPC. And how the HPC market worked before:
If you're the best performing "computing cluster" (ie. whatever you call the entity that can complete a massive calculation), you get a blank check from Congress.
Why? Because you need those calculations to "pump" nuclear weapons. They are needed to calculate both the geometry to make fusion bombs possible at all and to calculate the effect of a given geometry. They are the reason US/Russia/China have the biggest and strongest weapons known to humanity. And of course, they were replicated worldwide for this reason. I mean not that anyone will admit this but we don't have the best possible solution, and we don't know either the upper or lower limits for fusion devices (plus the lower limit would be very useful for energy generation, which for the US would effectively mean almost literally unlimited large marine ships that never need refueling. And yes, the solution to that problem is almost literally a 3d shape. Not just that, but mostly)
Now it appears it does not work the same when you democratize computation. Humans want a particular amount of computation and are willing to pay a given price for that. But the accountants still saw the blank check from before and ... do what accountants do. Economics don't change because you make things bigger and accessible, do they? Oh ... wait a second ...
As someone put it recently though, we now have data. 2.3% of humans in the US are willing to pay $20 per month for the support of a model like GPT-5.5/Claude code. If that's true (and after years of having this model, why wouldn't it be?) ... it means AI startups are doomed (because it's not even 10% of what they need it to be to make economic sense).
I disagree with some of the framing, but that's what good discussion is about -- figuring out where we agree and disagree.
But consumer uptake strikes me as the worst way to judge whether the big AI shops will make it. That's not where most of the leveraged user return or deployable capital is.
Sadly until china scales production of hardware it really isn’t economical to run this stuff yourself. It is good it exists though to put pressure against the labs.
Honestly imo this is just proof apple will win in the end. Eventually a phone will be able to run a model good enough to do most things and it then is game over.
It's probably vain to hope, but the US could also scale it's own production of hardware. I don't know of many efforts to do that, but it's a wasted opportunity if it doesn't.
I don’t know if others would find this useful, but previous did have custom harnesses etc.. but tools have improved so much that I drastically simplified.
That said, even the foundational models fail at the hard parts of my code so I use it opportunistically.
I have reduced down to just using zed, will three locally hosted models.
Qwen 3.6 27b on 1x3090 llama.cpp with 128k context ~50tps
Qwen 3.6 35B-A3B on 1x titan v + 2x1080ti llama.cpp with full context ~30tps
GPT-OSS 120b on pure cpu (slow)
I just use zeds parallel agents, task switching, stopping and fixing the code when a model gets stuck.
This still lets me stay engaged, and to modify code to be maintainable etc…
It gets me 80% there and I use to keep a subscription but often times just using googles AI mode is just as good.
That said I have 30 years of experience and insist on knowing how my code works, so this gets me 80% of the short term benefits while not depending on a 3rd party to keep my code moving forward.
Your mileage will vary and 2*5060ti 16gb cards would get around 100/tps with Qwen 3.6 35B-A3B on cards that are widely available.
To be honest the more modern cloud models are using draft tokens etc… that while they are superior for common coding tasks are degrading with more domain specific tasks.
That is just the cost of the draft model being ~10-20% of the foundation models size, and even the biggest Blackwell GPU is limited to ~250/tps so MoE or draft models are required for scaling performance at the foundational level IMHO.
The hard part is my use case are the OOD or small examples in corpus level, the above hurts there.
A Lamborghini may be nice, but I personally need a minivan more.
I'm using Kimi K3 + OpenCode. I pay their API pricing, costs about $5 / hour (and chews through ~10 million tokens / hour) during continuous use when I have one or two sessions running and doing their thing.
Can't comment on how it compares to plans (I really don't like the limitations and general shenanigans I see around plans, so I've never tried them).
It is notably slower than Fable / Opus / Gemini, but also vastly cheaper than their API pricing.
While I am grateful for open weights models I never found much use of them in the past, barring those I could run myself. This changed with deepseek 4 - it is staggeringly cheap, even if the performance definitely isn't near sota and it's not particularly fast either.
When I expect to need a lot of tokens and the task isn't too difficult I use sota to plan and create a thorough set of instructions and let deepseek chip away at it. With thorough instructions the quality tends to be satisfactory, and you pay something silly like $15 for 600m tokens.
GLM 5.2 seems like a decent price/perf and Kimi 3 has some real nice performance for an open weights model, but gpt 5.6 is unexpectedly affordable (especially if you don't automatically use Sol at max) so I don't think either is worth it atm. The exception is when you're working on something that US models get cold feet about, which seems like a constantly growing list. For me Fable is already too much of a headache in this regard, but chatgpt is still okay-ish. Hopefully it'll last. If not, there's Kimi.
tldr SOTA for most things because gpt 5.6 is token efficient. If I expect to burn a lot of tokens I use deepseek 4.
I am using GLM-5.2 via Ollama Cloud in the $20/month plan. With the same plan I can get many different API keys that I use to run my OpenWebUI server, my opencode and pi dev sessions. I am usually running 2 to 4 sessions concurrently, and I never hit quota limits. At work I get Claude, and I was getting reports that I was spending $75 per hour of work on Opus.
> The government should use procurement to create demand for portable, interoperable systems rather than permanent dependence on one API vendor.
Now, here is an idea that I have not heard before... and I think there is some merit to this. This is also the sort of thing that a state (looking at you CA, CO, IL, NY) could do, instead of just the federal government.
I was in this state a few weeks ago. I spent a bit of time familiarizing myself then wrote up my learnings as a series of exercises.
If you think of docker as "kinda like vms except not really" and k8s as "kinda like deploying and composing docker containers but not really", this may be for you:
It's for running a massive number of docker containers and automatically managing them and scaling them up and down on demand. It is also so famously brutally complex that basically you need a dedicated Kube expert to handle it.
The problem is that companies tend to exaggerate their own scale and think they need k8s and dedicated infra people when they could get by with a handful of beefy VMs or dedicated servers.
Everywhere, huh? Still plenty just running on VMs or serverless that don't need full blown container orchestration.
I'll agree that (nearly) everyone should know when Kubernetes is useful, but let's not pretend its the default method for everything. Even then, choosing to deploy on K8s falls on the sysadmins/DevOps I wouldn't expect the devs know or do much more than provide the Dockerfile.
I ported my company over to k8s to solve a concurrency and scaling issues.
What 15 concepts? You're making me worry that I missed something. It was straight forward: pods, nodes, hw type, lifecycle, deployment. They run almost the same docker as the old ec2s used.
What did I miss? Is there something important I need to read?
Enormous amounts of money is being invested in the development of AI models. Investors expect returns on their investment or they will not continue investing. Open weights make it harder for investors to get their money back, so it harms the industry.
Once the weights are out, it makes no sense in banning them in the US while the rest of the world takes advantage of it. But that doesn't mean developers of frontier models shouldn't take steps to prevent their weights from being stolen.
The weights themselves aren't stolen. The claim is that Chinese companies are using VPNs and proxies to buy massive amounts of Claude Pro and Codex subscription accounts, and then selling usage on those subscriptions as cheap white-label LLM API usage.
While selling that LLM API usage, they then capture all the prompts, outputs, and intermediate thinking the LLM does, and then sell those logs to the companies making open-weight models. The open-weight model developers then train on those logs to 'distill' a model.
Imagine making this argument with a straight face in any other industry:
"The claim is that Japanese car companies are buying Ford vehicles, and then leasing them to American consumers at cut-rate prices. In return for the cheap cars, the customers are letting the Japanese observe their driving behavior, studying how they use their F-150 and then the Japanese car companies are applying that data to design new vehicles that will directly replace Ford!"
I wasn't arguing against the practice, I was clarifying that the weights of these models are not being stolen, and explaining what these companies do to create their models.
In fact more countries should have government funded models. There are some obvious issues in China completely dominating open weights space. Kimi had funding of just $2B and could literally create national security threat. A lot of countries could fund something in the range of few billion for something so important. At the very least US and EU could fund few companies.
The sentiment in this article is nice. But open source software is a weak analogy for frontier models. Principally because software requires zero capital investment (actually zero) while frontier models demand billions. Open models can only survive in the long run if they can (eventually) generate significant cash flows or if they are paid for by governments. Now China essentially has a monopoly on open weight models. And so supporting open source models means either supporting long term economic capture by China or supporting Chinese government control of your intelligence. Both of these outcomes are unequivocally bad from an American perspective. If you live in the valley and benefit from the US venture ecosystem you should be highly skeptical of open weight models. Banning them may very well be the best course of action.
"Open source" open models can also survive if they are seen as a necessary cost of doing business. Same logic as tech companies working together on Linux. "We need an operating system. But an operating system is expensive to make on our own and does not really provide an edge. So let's just work on and use the open source one."
Hilariously bad take. Open weight models can be retrained of fine-tuned, that's the entire point. The idea that the "Chinese government controls your intelligence" is laughable in the case of open weight models. Once the weights are released you can do whatever you want with them. The idea that there's economic capture by the Chinese for products they're literally giving away is stupid to the point of inanity.
I can only assume this account is pure shilling for the closed-source AI labs.
So what open weight models do is at least provide a baseline of inference cost to add some sanity to the price markers. And of course predictability too--if you really want Kimi K2 instead of K3 you can still use it.
So the competitive pressure and predictability offered by open models is helpful for users
FlashAttention was a hell of a drug.
I think this is a very important aspect, especially after the huge GPT-4o backlash when GTP-5 came out. Each model has certain quirks, and areas where the previous model might be better for some use cases than the latest and greatest, and the labs so far seem to have no desire to offer some kind of "LTS" release.
Artificial inflation to recoup R&D.
But yes, the reason brand name drugs are drugs are more expensive than generics is due to intellectual property, both the patent and the trademark.
1. the field was nascent and new efficiencies were discovered
2. supply and demand
3. its in the company's incentive to make their models more efficient to increase overall usage so that while the margin remains the same, the total revenue + profit increases
I genuinely don't know what puzzles everyone?
If you're the best performing "computing cluster" (ie. whatever you call the entity that can complete a massive calculation), you get a blank check from Congress.
Why? Because you need those calculations to "pump" nuclear weapons. They are needed to calculate both the geometry to make fusion bombs possible at all and to calculate the effect of a given geometry. They are the reason US/Russia/China have the biggest and strongest weapons known to humanity. And of course, they were replicated worldwide for this reason. I mean not that anyone will admit this but we don't have the best possible solution, and we don't know either the upper or lower limits for fusion devices (plus the lower limit would be very useful for energy generation, which for the US would effectively mean almost literally unlimited large marine ships that never need refueling. And yes, the solution to that problem is almost literally a 3d shape. Not just that, but mostly)
Now it appears it does not work the same when you democratize computation. Humans want a particular amount of computation and are willing to pay a given price for that. But the accountants still saw the blank check from before and ... do what accountants do. Economics don't change because you make things bigger and accessible, do they? Oh ... wait a second ...
As someone put it recently though, we now have data. 2.3% of humans in the US are willing to pay $20 per month for the support of a model like GPT-5.5/Claude code. If that's true (and after years of having this model, why wouldn't it be?) ... it means AI startups are doomed (because it's not even 10% of what they need it to be to make economic sense).
But consumer uptake strikes me as the worst way to judge whether the big AI shops will make it. That's not where most of the leveraged user return or deployable capital is.
Honestly imo this is just proof apple will win in the end. Eventually a phone will be able to run a model good enough to do most things and it then is game over.
Quantised models running overnight go most of the way for non-coding tasks.
For the Hyperscalers...and Oracle...cant wait for the day...
What is your stack (harness, model) and how much do you pay per month?
How would you compare your experience to a typical subsidized plan like Claude Code + Pro plan?
I’m asking because i keep hearing that open weight models are cheap and efficient - is that really the case in practice?
That said, even the foundational models fail at the hard parts of my code so I use it opportunistically.
I have reduced down to just using zed, will three locally hosted models.
Qwen 3.6 27b on 1x3090 llama.cpp with 128k context ~50tps
Qwen 3.6 35B-A3B on 1x titan v + 2x1080ti llama.cpp with full context ~30tps
GPT-OSS 120b on pure cpu (slow)
I just use zeds parallel agents, task switching, stopping and fixing the code when a model gets stuck.
This still lets me stay engaged, and to modify code to be maintainable etc…
It gets me 80% there and I use to keep a subscription but often times just using googles AI mode is just as good.
That said I have 30 years of experience and insist on knowing how my code works, so this gets me 80% of the short term benefits while not depending on a 3rd party to keep my code moving forward.
Your mileage will vary and 2*5060ti 16gb cards would get around 100/tps with Qwen 3.6 35B-A3B on cards that are widely available.
To be honest the more modern cloud models are using draft tokens etc… that while they are superior for common coding tasks are degrading with more domain specific tasks.
That is just the cost of the draft model being ~10-20% of the foundation models size, and even the biggest Blackwell GPU is limited to ~250/tps so MoE or draft models are required for scaling performance at the foundational level IMHO.
The hard part is my use case are the OOD or small examples in corpus level, the above hurts there.
A Lamborghini may be nice, but I personally need a minivan more.
Can't comment on how it compares to plans (I really don't like the limitations and general shenanigans I see around plans, so I've never tried them).
It is notably slower than Fable / Opus / Gemini, but also vastly cheaper than their API pricing.
Deepseek is at least on par with Sonnet (ghcopilot at work)- I don’t use opus, too spendy and I don’t need that level of ability.
The cost is for me was $5/6 months of use. Not a big user I guess! It’s good though, fast enough and incredibly inexpensive.
Been testing Qwen 3.6 28B on a 5090, and it’s also quite good for “free”.
I mostly do small self serving embedded projects based on esp32, so not very complex.
When I expect to need a lot of tokens and the task isn't too difficult I use sota to plan and create a thorough set of instructions and let deepseek chip away at it. With thorough instructions the quality tends to be satisfactory, and you pay something silly like $15 for 600m tokens.
GLM 5.2 seems like a decent price/perf and Kimi 3 has some real nice performance for an open weights model, but gpt 5.6 is unexpectedly affordable (especially if you don't automatically use Sol at max) so I don't think either is worth it atm. The exception is when you're working on something that US models get cold feet about, which seems like a constantly growing list. For me Fable is already too much of a headache in this regard, but chatgpt is still okay-ish. Hopefully it'll last. If not, there's Kimi.
tldr SOTA for most things because gpt 5.6 is token efficient. If I expect to burn a lot of tokens I use deepseek 4.
It's not as good as the frontier models I use at work, but it's plenty capable for the types of tasks I am using it for.
Now, here is an idea that I have not heard before... and I think there is some merit to this. This is also the sort of thing that a state (looking at you CA, CO, IL, NY) could do, instead of just the federal government.
If you think of docker as "kinda like vms except not really" and k8s as "kinda like deploying and composing docker containers but not really", this may be for you:
https://ojensen.net/infra/understanding-k8s-1
It's actually really neat, i wish i had bothered to learn it years ago.
Its 2026 and its the de facto method of deploying software basically everywhere.
You gotta really work for it to not know what its for by now.
I'll agree that (nearly) everyone should know when Kubernetes is useful, but let's not pretend its the default method for everything. Even then, choosing to deploy on K8s falls on the sysadmins/DevOps I wouldn't expect the devs know or do much more than provide the Dockerfile.
That is some really impressive bubble you are living within. Basically everywhere - nowhere near that, no.
[edit]: Maybe containers, but software in general is so much more broad than containers.
We walked in, and it was fine. Because it was all kubernetes and laid out like every other app for the most part.
The kube hate is just sad at this point. You need to know like 15 concepts that are all applied in the same way. It mostly just works.
What 15 concepts? You're making me worry that I missed something. It was straight forward: pods, nodes, hw type, lifecycle, deployment. They run almost the same docker as the old ec2s used.
What did I miss? Is there something important I need to read?
I’ve been running my own personal k8s cluster on digital ocean for the last 5+ years now and it’s dead simple.
Takes me 30 minutes to create a new namespace a deploy an app, love the bonus of having complete flexibility on my stack too — PVC + SQLite ftw
The value built on that stability was probably worth acquiring, and it’s infra will decay.
Once the weights are out, it makes no sense in banning them in the US while the rest of the world takes advantage of it. But that doesn't mean developers of frontier models shouldn't take steps to prevent their weights from being stolen.
While selling that LLM API usage, they then capture all the prompts, outputs, and intermediate thinking the LLM does, and then sell those logs to the companies making open-weight models. The open-weight model developers then train on those logs to 'distill' a model.
We used to call that competition.
Imagine making this argument with a straight face in any other industry:
"The claim is that Japanese car companies are buying Ford vehicles, and then leasing them to American consumers at cut-rate prices. In return for the cheap cars, the customers are letting the Japanese observe their driving behavior, studying how they use their F-150 and then the Japanese car companies are applying that data to design new vehicles that will directly replace Ford!"
[1] https://adaptive.live
I can only assume this account is pure shilling for the closed-source AI labs.
Open Weights is free as in beer