Does anyone really believe the cost data for benchmarks? Pretty sure all models will behave differently at this point when they get benchmark shaped questions.
(And I’m sitting on data showing Sol 6.1 and Qwen 3.8 Max cost roughly the same per task, for real tasks.)
Yeah, it's in the same ballpark as Kimi 3.1. Both of which are obviously open weight, so that's at least a big differentiator from Sol. Let's hope against hope that the much smaller flash model might beat deepseek v4.1 flash and run for as cheap.
I hope they are able to serve 4 faster. Qwen 3.8 is a great model but I’ve regretted including it in the battle system for opencode [1] we have been building because it scores so low.
And it’s a shame because the only reason it scores so low is that it’s slooooow. (Like 30 tokens per second vs opus near 80.) If you exclude voting, Qwen 3.8 does well.
(We allow early voting contrary to Arena.ai because I think it makes sense for agents, where it doesn’t for chat. This means our data takes speed into account.)
Also, it would be ideal if they served this with third parties. Part of the slowness with 3.8 is just not a lot of providers. Their open source version is much faster, but we couldn’t include it because it doesn’t handle images (another own goal by Alibaba).
In case it translates to Qwen 4 27B, I've been running Qwen 3.8 Flash Next (which while MoE (and bigger) is supposed to be a preview of Qwen 4's architecture) and it runs pretty solidly regardless of context window size and completes random deepswe challenges I throw at it locally (at Q2-Q3 quantization! so it fits on my shitty setup).
The model I’m referring to that would be faster and similar size, but doesn’t handle images is Qwen 3.8 2.4T: https://openrouter.ai/qwen/qwen3.8-2.4t-a95b. It’s weird that this is pretty much the same as Max but doesn’t support images. No idea why.
This inspired me to re-run pelicans through 3.8 Max - but OpenRouter has qwen3.8-max-prime and qwen3.8-max-0902 - prime is twice the price and faster - so I went with the 0902 one.
Turns out Qwen4-Max has a none setting, but 3.8 Max didn't allow disabling reasoning - you had to pick from these five levels instead: minimal, low, medium, high, xhigh.
I feel like the xhigh result is one of the first times I've seen a model go for a non-red colored bicycle (teal). Still haven't seen the leftward facing bike though.
This is exciting. The continued focus on efficiency alongside quality is encouraging - it makes me hopeful to see what smaller models will become capable of.
The past few days I’ve been finding myself asking “seriously, where is it”… feeling that the world has somehow changed because Qwen 4 27B was later than I expected.
I can't imagine them keeping up long term given the information asymmetry plus compute restrictions on top of it.
I feel open-weight was able to keep up the past year using efficiency improvements (hybrid attention, improved MTP drafters, improved optimizers) to offset having less compute. But OpenAI/Anthropic started cutting token costs just a few months after these, which to me suggests they copied these improvements. Indeed it would be hard to work in LLM training and not know about DSpark, Muon etc.
So what do you do if closed source uses your learnings, gives no knowledge back, and has 10x more compute and funding on top of it? The situation feels quite dim right now, esp. with the Nvidia thing.
I think so but it may be a very short term effect, I think their pace has roughly matched the pre-speed up pace of the frontier labs (the 1.7-2x speed up noted on by METR), and I’d guess we see a similar speed up from the Chinese labs once their internal models hit an equivalent level enabling the same speed up.
Following that logic the gap widens a little bit here but then will steady again. Before this there has been a general catch-up trend too which probably also persists leaving us in a slow closing of the gap as the net effect?
If the speed up is the reason, then they will never catch up (with traditional means). Because it means RSI predictions are real, and we will see an explosive growth in capabilities (far beyond exponential).
While I doubt that Qwen 4 Max can seriously compete with frontier models, the AA index has become all but useless, due to the introduction of new benchmarks in the aggregate which are of questionable value as a measure of "intelligence" [1]. AA has seemingly also abolished updating the Coding Agents Index, which was an okay benchmark for coding ability, and older models are completely absent from the v5 version of the full index.
[1] Especially Terminal-Bench-Science 0.1 stands out. Example problems:
Probably very relevant to people working in these fields, but it is honestly not surprising if a lab with limited compute will not train for these technologies, and it says pretty much nothing about more common use cases. I also suspect that this benchmark measures knowledge retrieval more than raw reasoning ability. This is where you really need a breakdown of capabilities and not an increasingly meaningless aggregate number.
Compared to OpenAI and Anthropic, yes. There's this export ban on Nvidia chips, after all. And all the CPUs they have won't help much with model training.
AA has become pretty meaningless. They just keep adjusting the recipe to get a preferred ranking result, with no supporting evidence for their decisions.
Edit: In addition to reshuffling the weighting of individual benchmarks, they also do things like test Anthropic and OpenAI in their native harnesses, while not doing the same for Qwen/Deepseek/etc.
One thing I find confusing about AA is that they show GPT 6.1 dominating the lower end of the Pareto line, with eg, GPT 6.1 Sol at medium supposedly costing far less per task than DeepSeek V4.1 Flash at high.
Yet every time I compare the two in something like Pi, Sol ends up being 5-10x the price for a simple query, something like 7-10c vs. less than 0.5c for DeepSeek. What's the deal?
It's great fun asking it questions like "who is the Dalai Lama?", "where is TSMC headquartered?", "why do people post Winnie the Pooh in reference to Xi?", "who controls the South China Sea?", "why was Chiang Kai Shek fighting Mao?", "why were students in Tiananmen Square?"
It'll gush about how to make moon cakes and where the tallest and longest bridges are and how to run qwen on a raspberry pi - but it can't seem to manage some basic Wikipedia questions.
I've done some work on this and a lot of the refusals are in the API. The weights are more permissive at these types of things. Try it again when we get some of the smaller Qwen 4 weights.
Qwen3.8 flash next handles these questions just fine, what should I expect to see? It correctly reports who the Dalai Lama is, that TSLC is in Taiwan (Hsinchu), explains why people post Winnie the Pooh memes, etc.
Really would love if these publications have a huge, bright red bolded section "How to use this right now". Is it on openrouter? Is Qwen hosting it already? Maybe I'm just ignorant?
EDIT: oh, I guess this isn't open weights, so presumably I'll have to get it through Qwen or whoever they might partner with (Bedrock?).
Another Qwen doesn't seem all that scary, a few nation-state actors have likely exfiltrated weights of much stronger non-open models and abliterated them for use in large-scale collection on targets foreign and domestic. But god forbid someone mods a video game or jailbreaks their phone without applying for Claude Mythos or Daybreak, right?
If Anthropic truly believes we are on the verge of catastrophe but nobody is listening because they've been crying wolf for years about it while they get rich then they should truly do something. By "do something" I don't mean try again to convince the world to slow down as they IPO their bleeding edge research and solution (which is creating the doubt you mention) but just say "hey, we're so sure we're months away from worldwide disaster that we're shutting down our AI operations completely in hopes we can all put some attention to this before it's too late - will you follow or work together to come up with ways to prevent this problem from happening".
After all, unless their goal to say "I told you so but I wanted to make money more than I wanted you to believe we could stop a billion people from dying" then what good is the current approach doing them?
This is not an open-weights modell. You have to wait for the smaller models
"Smaller models in the family are coming to open source, built on the same lineage and trained with the same generation of techniques, tuned for what small models do best."
[1]: https://artificialanalysis.ai/models/qwen4-max
By comparison opus 5.5 (max) is $10.01.
If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
Scroll down to the cost per task pareto front chart which includes more difficult levels by default: https://artificialanalysis.ai/models/qwen4-max#intelligence-...
Qwen4-Max is not in a good spot on that chart. Astra xhigh and Fable 5.2 medium had a similar cost per task with higher intelligence ratings.
They compare different reasoning levels.
https://artificialanalysis.ai/models/deepseek-v4-1-flash-101...
(And I’m sitting on data showing Sol 6.1 and Qwen 3.8 Max cost roughly the same per task, for real tasks.)
And it’s a shame because the only reason it scores so low is that it’s slooooow. (Like 30 tokens per second vs opus near 80.) If you exclude voting, Qwen 3.8 does well.
(We allow early voting contrary to Arena.ai because I think it makes sense for agents, where it doesn’t for chat. This means our data takes speed into account.)
Also, it would be ideal if they served this with third parties. Part of the slowness with 3.8 is just not a lot of providers. Their open source version is much faster, but we couldn’t include it because it doesn’t handle images (another own goal by Alibaba).
[1]: https://agentduel.com/leaderboard
Maybe I'm mixing up model variants, but Qwen 3.8 27B that I run locally totally does handle images. Surely the bigger one would as well?
For comparison, here's one I got from Qwen3.8 Max (at default reasoning effort): https://simonwillison.net/2026/Jul/20/afraid-of-chinese-mode...
This inspired me to re-run pelicans through 3.8 Max - but OpenRouter has qwen3.8-max-prime and qwen3.8-max-0902 - prime is twice the price and faster - so I went with the 0902 one.
Turns out Qwen4-Max has a none setting, but 3.8 Max didn't allow disabling reasoning - you had to pick from these five levels instead: minimal, low, medium, high, xhigh.
Here are those 3.8 Max 0902 pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - they're mostly not as good as 4, but I prefer the 3.8 xhigh one.
Looking forward to the "Flash" and 27B versions.
Hoping it’s a great local model though!
I feel open-weight was able to keep up the past year using efficiency improvements (hybrid attention, improved MTP drafters, improved optimizers) to offset having less compute. But OpenAI/Anthropic started cutting token costs just a few months after these, which to me suggests they copied these improvements. Indeed it would be hard to work in LLM training and not know about DSpark, Muon etc.
So what do you do if closed source uses your learnings, gives no knowledge back, and has 10x more compute and funding on top of it? The situation feels quite dim right now, esp. with the Nvidia thing.
Following that logic the gap widens a little bit here but then will steady again. Before this there has been a general catch-up trend too which probably also persists leaving us in a slow closing of the gap as the net effect?
[1] Especially Terminal-Bench-Science 0.1 stands out. Example problems:
- https://hub.harborframework.com/tasks/terminal-bench-science...
- https://hub.harborframework.com/tasks/terminal-bench-science...
Probably very relevant to people working in these fields, but it is honestly not surprising if a lab with limited compute will not train for these technologies, and it says pretty much nothing about more common use cases. I also suspect that this benchmark measures knowledge retrieval more than raw reasoning ability. This is where you really need a breakdown of capabilities and not an increasingly meaningless aggregate number.
Edit: In addition to reshuffling the weighting of individual benchmarks, they also do things like test Anthropic and OpenAI in their native harnesses, while not doing the same for Qwen/Deepseek/etc.
Models often get great scores despite being a lot worse than the frontier in practice.
Yet every time I compare the two in something like Pi, Sol ends up being 5-10x the price for a simple query, something like 7-10c vs. less than 0.5c for DeepSeek. What's the deal?
Ah yes, the reverse-benchmaxers of OpenAI models.
It'll gush about how to make moon cakes and where the tallest and longest bridges are and how to run qwen on a raspberry pi - but it can't seem to manage some basic Wikipedia questions.
American history is open-weight. American trade is closed-weight.
I wish there were more ~60b dense options like ripple too!
EDIT: oh, I guess this isn't open weights, so presumably I'll have to get it through Qwen or whoever they might partner with (Bedrock?).
I would assume it will not be on Bedrock for some time, AWS has not kept up with the latest non-American models.
After all, unless their goal to say "I told you so but I wanted to make money more than I wanted you to believe we could stop a billion people from dying" then what good is the current approach doing them?
This is not an open-weights modell. You have to wait for the smaller models
"Smaller models in the family are coming to open source, built on the same lineage and trained with the same generation of techniques, tuned for what small models do best."