Closed AI models dominate the price-performance frontier
village.dev
village.dev
Closed models might be better generalists because they don't have that much constraint on size and come from well funded labs that can afford to better train for generalization. If a model's actual goal is to be open then smaller is better.
By the time, we want to code, the scope has been limited pretty greatly to the point that my rule to the Ai is “don’t make any decisions outside of the plan, if something in the plan turns out to be untrue or cannot be implemented after 2 differing attempts stop and ask for clarification”.
Sub frontier level Ai can circle endlessly thanks to a bad assumption, but a human can cut through that by coming at the problem sideways.
Sure some world knowledge like things are affected by gravity, how people function, how a society function I think is needed for a coding AI but not trivia.
Words that are or were fairly common in tech circles — "bailiwick", "gatekeeper", "daemon", "boundary", "infer", "method".
It's true to say that most of these words have some specific meanings in tech, but you don't necessarily want the LLM to have only those narrow meanings when you are talking to it. A broader grasp of the word means the LLM won't go down a specific pathway before understanding you by analogy. That broader grasp comes from non-programming understanding.
The Opus frontier model will spin up a cheaper Sonnet subagent to explore the code base. Similarly we can have a Pareto optimal subagent write most code and verify it.
Words only have their meanings through general context.
edit: and of course from first party labs as well, like xiaomi, z.ai, kimi etc
Yep, on subscription plans DS4.1-Flash-1017 feels more or less infinite. I still reach for Astra at the high end of tasks but DS is great for cheap "repeatedly crawl the entire codebase and find query optimization" type tasks. I also use a lot of K3.1 Pro lately although admittedly it's a lot more expensive.
Also personally I find that the closed tiny models (Haiku, Luna) feel a lot less capable in practice than their benchmark scores indicate relative to DS.
That being said the closed labs own the high ground of intelligence, nothing really compares to Astra in my opinion in terms of sheer "it will achieve any task you give it," Kool-Aid-Man running through walls style capability. It's quite expensive but it's also very very capable.
First, OpenCode and Command Code offer you basically 6x-10x usage for some models (not all, though, look at the exact rates per model). That's a very similar model.
Second, while Anthropic subscriptions on paper offer the most generous usage per dollar spent [1], this comes with a few caveats. First, they are the most expensive of all models in cost/task. Second, if you use Claude Code non-interactively (claude -p "..." or SDK/ACP mode), they apparently charge you 70% more usage; you have to use their apps directly and interactively to get the full benefit of a subscription. Also, Anthropic has the lowest cache TTL at 5 minutes, which can really hurt for interactive use. Contrast that with OpenAI's default of 30 minutes (and according to a recent post by Tibo, apparently 12 hours for subscriptions, though I haven't verified it myself) or the fact that DeepSeek caches prompts for three days (which is a major reason for their extremely high cache hit rate and how cheap they are in practice), and it doesn't look like THAT great of a deal in comparison anymore. It is of course still good value for money, just not as great as it may look at first glance.
[1] At least before the recent usage bump to OpenAI's Pro plans, though that doesn't seem to include the Plus plan if that is what you are on.
If someone is in a position to use the personal subscriptions, the $20/month plans are about a dollar per work day. I'd rather have the $20/month plan for Opus or Sol 6.1 than deal with all of the extra churn and babysitting from Luna or Deepseek.
I've had really great results with cheap flash agents using a system that has a parent agent that spawns 2 sub agents. They remain active throughout the whole session and don't lose context.
The 2 sub agents get the exact same prompt and tools as each other and they independently answer but before responding back up to the parent they exchange their findings which must be backed up by a citation and then debate their findings and citations repeatedly until they can agree on correctness of all points. They are also heavily encouraged to never assume and always verify with the prompt and tools and guards where they can be realistically enforced.
It churns through tokens but tokens are crazy cheap with some of the plans out there. Could Opus one shot most of those same queries in like 1/4 time? Yep, but that's not the point for me.
DeepSeek Flash for parent then DeepSeek Flash and GLM 5.3 flash as the debating subagents gives best overall results in my experience.
For context I do equal parts sysadmin/devops and having them write software
As for the session maintenance, I have it configured to trigger a compaction when it ends a 'debate' with >25% of context used but before the compaction it puts what it thinks are key points into an MD file. The approach has worked so far in terms of stopping the agent from getting silly with too much context and retaining information. These flash agents get too flakey for my liking past 25%
DSH has been fantastic, the customisation is great, like I mentioned before. What has got me to stick with it though is the way it presents information like diffs, subagents, tool calls, thinking and sessions. It's nothing revolutionary but everything is where I want it to be and provides just enough information
As stocks fall and the investment money dries up, it'll be easier and cheaper to buy hardware to run local models
Let’s say I want to run qwen 27B or flash next on a machine I own, off solar panels that are also sunk cost.
How did they price the marginal cost of a token? Depreciation on the box + reverse billing opportunity cost with the power company?
This is totally suitable (speed wise) for coding.
This is not some top of the line hardware (might be now, but it should be run of the mill for people that built their system <1.5 years ago).
Seems very slow, but it's actually fine.
It is also evidence of commodicisation / brutal competition. I wonder if they cut costs because there is stiff competition from open weight models (and arguably we are in the blitzscaling phase where they try to corner the market, not the enshitification stage).
Where it differs from the larger models is that it has less knowledge; respectably as it performs for its size, 552B parameters are still a lot less than the multiple trillions that frontier models have. Sol/Astra/Opus/Fable may know the solution to the problem directly or to a very adjacent problem, while DeepSeek has to work its way through it from first principles and then sometimes fails (because LLM reasoning is really a heuristic-guided search). But this is also a failure mode that I have seen with Luna.
[1] As evidenced e.g. by ErdosBench. https://www.ulam.ai/erdosbench
[2] Using this as an example over software engineering tasks here, because it tests reasoning capability more than knowledge retrieval.
I'm using Claude all day at work and deepseek privately for very similar tasks. The price/performance difference is extreme.
I don't think I am going to try to beat them at this game anymore. I've got a machine in my closet that is both collecting dust and accruing value. A very maddening state of affairs.
At this point, trying to beat a major provider is like running your whole house off grid when the alternative is 12c/kWh that is already wired to the premises. You might win if you make it your only job for years on end and have a very specific lifestyle that is tolerant to weird tail risks (assuming you aren't willing to go back to the grid in an emergency).
So what's the point? After the codex <> web quota merge next month it'll be useless ...
What's more important is that open models can be inspected, they can be customized, and you don't have to worry that a company will decide that you're not allowed to use them for stuff like reverse engineering things or do security work.
That's why open models, and especially ones you can run locally will always be important. When you're renting closed models, you're basically a serf who has no say in how the model can be used.
I'm running full swarming on a 90$ subscription lol
Using DeepSeek I've made the Wayland compositor I use every day. It's finished now in the sense that it doesn't crash or have any bugs, nor can I think of any features to add or anything to change.
When somebody else is paying, I'll absolutely just throw Opus 5.5 at it. Who cares. It gives the best chances that the output is what I want, at no extra cost to me. But for my personal projects, using DeepSeek in Pi with a workflow extension is competent.
Sorry for throwing lots of text at you. All of this to say there's more to it than just which model is more intelligent. There's also speed, cost, and "good enough". Perhaps the problems I'm working on right now are just not complicated enough for "Opus is better" to be a real selling point.
Open weight models are generally more humble when working at fronteir, seek human input more reliably on thi gs they have no buisness deciding on, and can be trusted to not fuck up codebase in some way. Also, the way claude talks by default and tends towards even with all the "use this standard or that" tricks, grinds my gears in a way that drive me postal.
So, all in all, the shit prose, the constant oversight robbing me of attention better spent elsewhere, and high cost, and seemingly correct and difficult to catch errors on first pass make avoidance/non-reliance a nobrainer.
When dumb models are right, they're right, because likely its a solved problem elsewhere, and when they're wrong its quite transparent. Claude and fable are extremely good at spinning resilliant bullshit that can take quite some mental labor to verify and disprove, usual problem is flawed or deeply i correct assumptions on top of which then relativly sound reasoning and evidence is layered, or otgerwise weird contradictory decisons made "because i thought it would be better this way" that waste 48 hours of gpu time on an invalid model because claude fucked the one small part entrusted to it, which was extremely well defined and reasoned/backed by literature and it just made an executive decision to deviate without consulting or warning about it.
so fuck claude.
At least Gemini 4 Argon has made writing the front ends I hate and don't care about a breeze.
I have no idea how much that's worth to other people, but it's more than worth the account or three I use.
But I think it becomes a little more relevant when making a price-performance comparison such as the article makes between “open” models and “closed” models.
Because the reader is invited by the term “price-performance” to think about efficiency, when in reality the results are being skewed by the fact that the user is not the only one paying for the service: investors are chipping in fairly directly for every token, at inference time.
Users are going to continue to do their own thing for their own reasons, and they should; that’s the market at work. But this is not an efficient market where meaningful comparisons can be made; all of the major closed providers are loss-leading.
For us who spends more than $150K tokens a month with less than $5K spending in subs, it's really hard to justify the switch and it's even senseless (for now).
I had a check on artificialanalysis.ai and it shows a 16% reduction in token and 200% increase in time for Opus 5.5 Max compared to DeepSeek V4.1 Flash Max. I don't think the data is very accurate for my real life workload, but for any workload it represents I would think Opus having little advantage.
Open-weights excel primarily at coding.
Let's scale this out 3 years into the future: - Opus 6.7 costs an arm and a leg to run for 2 prompts - Astra 7.89 literally requires you to buy-now-pay-later because it's so expensive - As a bootstrapped dev, your only option is serving local Qwen-4.04-27B that runs on your old RTX (that you bought in 2024)
Sure, it's worse performing. But if you can't afford the best ones in the first place, what's so bad about having a "pretty decent" model that you run essentially free (only paying electricity) if it does what you need it to?
... and if we're all in the permanent underclass anyway, what does it matter that your local model is dumber than the SOTAs? At least you have one! Some people who can't afford GPUs still have to type emails by hand.
(jk, this is actually a cool dataset though ^)
It’s pretty obvious that local AI has a higher TCO if you look only at money. But there are other things to be considered.