Open Teams originally wanted to rent out open source developers to sponsors with Oliphant controlling everything. Now they pivoted to installing local LLMs (on what hardware exactly?).
What will happen is that this will be the third consultancy with a lofty narrative after Enthought and Anaconda that Oliphant established. It is always bait-and-switch.
> The first issue I have with it is that it uses a logarithmic scale on the cost axis. Using a log scale is the only way to make you spot the difference between a model that costs $0.015 per task and one that costs $0.032, while the same plot contains a model that costs $3.69 — almost 250 times as expensive. However, the net result is that the viewers can no longer appreciate the immensity of the price difference between the cheap models and the heavy ones; nor can they realize how inconsequential the price differences are between the cheap models.
This is an asinine complaint, and nobody can seriously tell me that the last plot on their page [0] is more readable than the AA one [1]. If I'm using a model at the lower range of the cost scale for whatever list of tasks, and i switch to another model at the lower end of the cost scale, my spending might double anyways! This should be reflected in the plot, and linear scale doesn't do it justice.
It's also much easier to see the mentioned pareto frontier in the log plot than in the linear one.
I can see why they disagree with the pricing determination for open/local models, but I don't think there is one clear right way to do it. So how do they do it instead?
>Hardware is priced at zero, on the basis that both an RTX 3090 PC and a 64GB Strix Halo are desirable gaming/work machines anyways.
...oh
Would have been nice to mention explicitly how the pareto frontier changes with those new calculations.
It’s funny they state log plot is “the only way” to keep the cheap area readable, say they hate it, and then immediately have to zoom into their non-log plot cheap area because it’s unreadable.
I think the key there is “hardware you already own.”
If you own a graphics card you bought for gaming or a laptop you bought for doing schoolwork there is $0 in cost of local AI tokens, because 100% of the cost was assigned to doing other things.
God, I'd love to have more local VRAM but the card costs are out of control.
Based on these costs I'd almost expect the amount of gaming graphics memory to go down over the next few years putting more stress on running those local models.
This looks great! I also think speed should be part of the metric (i.e. how long does the model take to actually solve a task). For me, I prefer to run expensive models such as Sol on light reasoning, which usually gives me good answers with quick responses.
For my style of coding (quick back-and-forths and corrections) it makes a big difference if a model comes back in 1-2 minutes compared to 5-10, and I am happy to pay a bit extra for that.
I agree on both: indeed a nice article, and I'd like to see a chart with speed as x axis.
As someone who only needs AI for a couple of tasks per day, I don't really care how much it costs, especially when subscriptions are subsidized. I want to filter by speed (eg, max task time < 1 min) and then choose the intelligence I need for the task. This will surface models like Gemma4 31B (xhigh) running on Cerebras and GPT Sol (med) fast mode. Using these models feels great and are affordable for infrequent tasks.
The more dimensions you take into consideration, the larger the proportion of the Pareto frontier becomes across all distributions, and making it harder to choose.
Yeah, this doesn't include older models, some of which were already saturating many common tasks a year ago. I'll often add them to the AA graph for reference.
to be fair, you don't need to start the y-axis at zero [0] but for some of the graphs where the lowest value is close to 0 the best practice is to do so
there's a fun Excel artifact where it auto-selects the 'relevant' range with no adjustment for how proportionally close to 0 the values are - a professional researcher publishing to a journal should know better (and should be ridiculed for not incorporating best practices) but for a personal blog by an SWE this really isn't the worst sin
I disagree. The y-axis is some arbitrary intelligence score that we use as a proxy for performance on whatever our specific task happens to be. So it doesn't matter if a model is a 0, 1, or 20 along this axis, they are all useless for the tasks I want.
And as the complexity of your score increases, the cutoff goes up. We can quibble about where your personal cutoff is, but it aint 0.
It gives you a wrong perspective, especially if you are distracted, on model capabilities: Fable 5.1 is not 30% better than Sol, but is the very first impression you get when you look at the first graph.
If I'm not wrong OAI tried a similar trick when GPT5 was announced ... they have been criticized a lot.
> So we can say the same about the authors "AA’s plot is misleading" claim, he is "not schooled in reading graphs"?
Oh absolutely - as other commenters have pointed out.
> When building a chart is good practice to provide log scale switch and zoom&pan capabilities, so the reader can decide how to look at it.
For the majority of the time charts have existed, your "good practice" would have been impossible. Charts have historically been static images (e.g. published in a journal). So there have been conventions on how to depict them - and at times it is very appropriate to start from something other than 0.
Here's an article from the UK's Office For National Statistics:
I agree on the past, when you have limited resource and have to print something on paper that you cannot recall to fix you have to carefully choose the layout.
But that era is gone since decades, nowadays, given how easy it is, it's a shame to not provide log scale switch and zoom&pan capabilities.
The discussion isn't about what one person should do, but on what is appropriate in general for depicting such graphs. What he did do is in line with current and long standing recommendations.
Well done. The inability to switch between log and linear always bothered me.
Another thing is if you're using the subscriptions with OpenAI or Anthropic you get an order of magnitude discount relative to the per-token price. So you need to move their models ~10x to the left on the plots to get a fair comparison.
Is there a case in which a heavy agentic coding user of mid or mid++ tier (remotely hosted) models is better off using PAYG/API pricing than just getting a subscription? (Assuming no easy access to high end local hardware and I've deliberately left the top tier/cutting edge models out becau).
There is an immense difference in cost between the state-of-the-art models from Anthropic and OpenAI and the much cheaper Chinese models ... How much extra intelligence emptying the wallet purchases obeys the law of diminishing returns: while a top-tier engineer or scientist is probably going to be able to appreciate how much better Fable 5.1 [is] ... most people will have a hard time doing so.
Nebari is officially listed as a JATIC product as part of the next-gen toolchain supporting DoD AI development.
Are we officially ~one degree of Kevin Bacon from the DoD endorsing running Chinese OSS models because they're self-hosted and we're all too dumb to tell the difference?
The complaint about not being able to switch between linear and log is valid, which is what I did for making a 3D speed/cost/quality frontier application for a recent meetup talk: https://www.williamangel.net/apps/model_performance.html
Because speed is important, as the reasoning and hardware determine both cost and speed. it's a three dimensional tradeoff.
> This is fine in most cases, but for open-weights models it can be a lot more expensive than what the exact same model can be rented for from third-party API providers.
Filter by quantization, and most providers will have the same price. There is some "base" price even for open-weight models. Anything cheaper means some tricks on the provider's side.
I've been very impressed with GLM 5.3 Flash's performance even without considering cost, but once you factor that in, it's incomparable. Not surprised to see its position on the chart.
asf1289 | 15 hours ago
What will happen is that this will be the third consultancy with a lofty narrative after Enthought and Anaconda that Oliphant established. It is always bait-and-switch.
fxwin | 14 hours ago
> The first issue I have with it is that it uses a logarithmic scale on the cost axis. Using a log scale is the only way to make you spot the difference between a model that costs $0.015 per task and one that costs $0.032, while the same plot contains a model that costs $3.69 — almost 250 times as expensive. However, the net result is that the viewers can no longer appreciate the immensity of the price difference between the cheap models and the heavy ones; nor can they realize how inconsequential the price differences are between the cheap models.
This is an asinine complaint, and nobody can seriously tell me that the last plot on their page [0] is more readable than the AA one [1]. If I'm using a model at the lower range of the cost scale for whatever list of tasks, and i switch to another model at the lower end of the cost scale, my spending might double anyways! This should be reflected in the plot, and linear scale doesn't do it justice.
It's also much easier to see the mentioned pareto frontier in the log plot than in the linear one.
I can see why they disagree with the pricing determination for open/local models, but I don't think there is one clear right way to do it. So how do they do it instead?
>Hardware is priced at zero, on the basis that both an RTX 3090 PC and a 64GB Strix Halo are desirable gaming/work machines anyways.
...oh
Would have been nice to mention explicitly how the pareto frontier changes with those new calculations.
[0] https://openteams.com/wp-content/uploads/2026/09/all_models-... [1] https://artificialanalysis.ai/#intelligence-comparison-tabs
destrolas | 13 hours ago
sinuhe69 | 14 hours ago
giancarlostoro | 14 hours ago
sinuhe69 | 11 hours ago
Primer81 | 14 hours ago
sarjann | 14 hours ago
Grombobulous | 13 hours ago
If you own a graphics card you bought for gaming or a laptop you bought for doing schoolwork there is $0 in cost of local AI tokens, because 100% of the cost was assigned to doing other things.
datadrivenangel | 11 hours ago
pixl97 | 11 hours ago
Based on these costs I'd almost expect the amount of gaming graphics memory to go down over the next few years putting more stress on running those local models.
oliwary | 14 hours ago
For my style of coding (quick back-and-forths and corrections) it makes a big difference if a model comes back in 1-2 minutes compared to 5-10, and I am happy to pay a bit extra for that.
quinncom | 13 hours ago
As someone who only needs AI for a couple of tasks per day, I don't really care how much it costs, especially when subscriptions are subsidized. I want to filter by speed (eg, max task time < 1 min) and then choose the intelligence I need for the task. This will surface models like Gemma4 31B (xhigh) running on Cerebras and GPT Sol (med) fast mode. Using these models feels great and are affordable for infrequent tasks.
Sha1rholder | 9 hours ago
vb-8448 | 13 hours ago
andai | 13 hours ago
paimapi | 13 hours ago
there's a fun Excel artifact where it auto-selects the 'relevant' range with no adjustment for how proportionally close to 0 the values are - a professional researcher publishing to a journal should know better (and should be ridiculed for not incorporating best practices) but for a personal blog by an SWE this really isn't the worst sin
[0] https://digitalblog.ons.gov.uk/2016/06/27/does-the-axis-have...
vb-8448 | 13 hours ago
Just look at the first chart: the distance between Fable 5.1 and Sol is <5%, but it looks like 25 or 30%.
_aavaa_ | 11 hours ago
And as the complexity of your score increases, the cutoff goes up. We can quibble about where your personal cutoff is, but it aint 0.
vb-8448 | 9 hours ago
But even if it was between 0 and Inf+, it still gives you a wrong perspective, especially if you are not paying attention, on model capabilities.
datadrivenangel | 11 hours ago
BeetleB | 10 hours ago
What is bad is starting at 0, showing an indicator of a gap, and suddenly starting at 30 or whatever after the gap.
vb-8448 | 9 hours ago
If I'm not wrong OAI tried a similar trick when GPT5 was announced ... they have been criticized a lot.
BeetleB | 9 hours ago
Only if you aren't schooled in reading graphs. It's a given that you always have to look at the axes when interpreting a graph.
How exactly would you zoom into a section of a graph and just show that section?
Sharlin | 9 hours ago
vb-8448 | 9 hours ago
So we can say the same about the authors "AA’s plot is misleading" claim, he is "not schooled in reading graphs"?
> How exactly would you zoom into a section of a graph and just show that section?
When building a chart is good practice to provide log scale switch and zoom&pan capabilities, so the reader can decide how to look at it.
BeetleB | 9 hours ago
Oh absolutely - as other commenters have pointed out.
> When building a chart is good practice to provide log scale switch and zoom&pan capabilities, so the reader can decide how to look at it.
For the majority of the time charts have existed, your "good practice" would have been impossible. Charts have historically been static images (e.g. published in a journal). So there have been conventions on how to depict them - and at times it is very appropriate to start from something other than 0.
Here's an article from the UK's Office For National Statistics:
https://digitalblog.ons.gov.uk/2016/06/27/does-the-axis-have...
vb-8448 | 8 hours ago
But that era is gone since decades, nowadays, given how easy it is, it's a shame to not provide log scale switch and zoom&pan capabilities.
BeetleB | 8 hours ago
vb-8448 | 7 hours ago
BeetleB | 6 hours ago
spider-mario | 6 hours ago
andai | 13 hours ago
Another thing is if you're using the subscriptions with OpenAI or Anthropic you get an order of magnitude discount relative to the per-token price. So you need to move their models ~10x to the left on the plots to get a fair comparison.
shelled | 13 hours ago
themgt | 13 hours ago
Nebari is officially listed as a JATIC product as part of the next-gen toolchain supporting DoD AI development.
Are we officially ~one degree of Kevin Bacon from the DoD endorsing running Chinese OSS models because they're self-hosted and we're all too dumb to tell the difference?
https://openteams.com/open-source-isnt-the-real-risk-in-nati...
datadrivenangel | 11 hours ago
Because speed is important, as the reasoning and hardware determine both cost and speed. it's a three dimensional tradeoff.
esafak | 11 hours ago
akazantsev | 7 hours ago
Filter by quantization, and most providers will have the same price. There is some "base" price even for open-weight models. Anything cheaper means some tricks on the provider's side.
SturgeonsLaw | 17 minutes ago