- 42comments
- 234comments
- 152comments
- 21comments
- 12comments
- 119comments
- 3comments
- 458comments
- 69comments
- 149comments
- —discuss
- 17comments
- 15comments
- 13comments
- 296comments
- 81comments
- 2comments
- 376comments
- 484comments
- 5comments
- 36comments
- 382comments
- 54comments
- 142comments
- 360comments
- 30comments
- 194comments
- 438comments
- 35comments
- 43comments
"When evaluating the Intelligence Index, it generated 140M tokens, which is somewhat verbose in comparison to the median of 140M."
Nowadays these error can be a good thing :)
Human error means this wasn't just stopped together by some bot.
My bet is that it's a bot error, but of a rule based one.
Yes, it seems they have a template that they fill with numbers. Similar issues spotted on Grok's performance page: https://news.ycombinator.com/item?id=49789558
I have seen them say "and reasonably priced when comparing to other models of similar price" on several models. I always get a kick out of it.
It feels suspicious that MiMo-V2.6 Pro gets 46 in de index while DeepSeek-V4.1 (https://artificialanalysis.ai/models/deepseek-v4-1-flash) gets 39. According to the appendix at the bottom of https://mimo.xiaomi.com/mimo-v2-6 the deepseek model sometimes surpasses mimo and it's not so far behind in capabilities. A week ago opus 5 appeared 1 points ahead of fable 5 despite fable being a much smarter model (this has been corrected already)
Why?
W what?
The main AA benchmark keeps changing, and had to be radically changed when Astra came out and showed zero improvement over GPT 5.6 Sol in their benchmark. Opus 5 is still 1 point ahead of Fable 5.0 on the index, if you manually add Fable 5.0 back into the list, so it hasn't actually been "corrected". It's only Fable 5.1 that is shown as ahead of Opus 5.
The AA benchmark is a weighted average of other benchmarks and some internal ones. I think the difficult part is finding benchmarks that reflect your own use of the models.
The way Artificial Analysis keeps changing their weights feels kind of like deciding who the winner should be and making the weights reflect that. They’ve been changing their weights to add more weight to improved long-running agentic capabilities, but doing so means they’re reducing the relative importance of world knowledge and of writing ability.
I’ll grant that maybe world knowledge isn’t that important for these models. But writing ability is important for human understanding, and I think the weird turns of phrase and word choices reflect the labs’ underweighting of the importance of human understanding.
I mean artificial is in their name....
DS 4.1 is good but it’s clearly not as “smart” as non-flash models- it just doesn’t have the training data. Without a solid plan, it goes off the rails pretty regularly.
All of the chinese labs have been overfitting on benchmark data to game the results for a while now - MiMo and Deepseek are not anywhere near frontier and mostly compete with models like Luna - which they are still worse than.
There isn't much compelling reason to use these unless you are just averse to giving money to openai/altman. A $20 codex sub gives you ~$150 of luna use per weekly limit, while there isn't any good subsidized options for chinese models at all (and the few who were subsidizing, like opencode, rugpulled by reducing monthly limit to $60 to $15 with no notice to users).
It is an impressive model. Agreed on most that is written on this page, with the exception of it being fast. I ran it on my own LLM benchmark suite[1] and it is faster than DeepSeek but still much slower than leading models. But it's pricing is where it really shines.
KillSwitch-Bench 1.0
1 - https://bench.killswitch-lang.org/
Speed seems to vary a lot with demand. Last night it was reaching 80+ tok/s
Interesting - have you done MiMo-v2.6-flash?
Not yet. But can do so if there is interest.
That looks like you aren’t using the ultra speed endpoint.
I’m using openrouter which I think is a fair representation of what the typical user will experience.
Why does Luna have a score of 0? I will say that in my limited experience, I use Luna and DSv4 Flash (haven't tried 4.1 yet) and Luna is wayyyy faster. They both output at the same speed but DS has an endless thought process
Why sol is not in the comparison?
the graph has a dropdown for selecting models
Not on mobile unfortunately
Not the graph at the top, the one further down.
OpenAI usage limits have been severely cut, and intelligence appears to be markedly declining, so I'm going to start trying these Chinese models seriously now. I don't mind if it takes longer. I just need the intelligence to predictably work the same way from day to day.
Same- I pay $200/mo for Codex but whereas I used to get a week's work out of a weekly limit, now I get roughly 1~2 days.
I've stopped using Astra entirely and remain on Sol orchestrating Luna Xhigh, but it's still not nearly a week's usage for a week's allotment.
And even then, whenever a new model is about to come out, it feels like the model I'm using is being dumbed down substantially.
I have no evidence for this and can have no evidence for this, but I can vote with my wallet regardless.
I strongly agree. Check out the Codex subreddit. Many empirical examples of Astra silently downgrading the models. One found Astra was silently using Luna Max (but still billing for Astra).
Even when I try to stick with Sol X/High, my limits are at best half of what they were before Astra launched, and the intelligence has declined markedly.
I cancelled my $100 plan. This is absolutely absurd and frankly unusable now.
It feels bizarre reading about the amounts spent on it here and paying like 10 eurobucks a week for DS
These tools were pretty great if you could afford them, but now they are expensive and shit, and that combination doesn't work.
For me, DS Pro is still behind Sol. But I do think many people hitting the Codex limits in 1 or 2 days are doing something wrong.
Entirely depends on the work asked and the repo involved.
When I’m doing work on a repo where I’m implementing a standard and the agents have to read the standard to keep from hallucinating my usage skyrockets.
Hell this changes depending on which language I’m working with.
My spending with Deepseek was more than a Codex subscription, though. How are you keeping it to $10/month? Super light usage?
see that's interesting, because I'm prompting all day and I usually end up with 15-25% by the end of the week. I'm using Astra xhigh exclusively
Weird. That's definitely not my experience. On the $100 plan and I can burn my entire week's budget in a few hours easily with Astra. It's borderline unusable.
Are you using codex or another harness? The problem could be the harness or using plan mode.
I wonder what dangoodmanUT is using! This is the time to compare!
People often use a bunch of subagents, poor context management (though Codex's tight context limits and constant compaction tend to mitigate this), a bunch of projects at once, etc., as well as not using workflows that do heavy planning once up front and then consult it rather than thinking endlessly about what to do during the implementation part.
It's also the case that working on massive codebases is just a different beast. If they've been slopmining a monorepo for months with 200x, then their codebase is probably Lovecraftian at that point and requiring extensive effort to iterate on.
For me Sol is less efficient than Astra - Sol makes many avoidable mistakes and has issues with context compaction - sometimes it goes haywire after a couple compactions.
Agreed on the “being dumbed down” observation. It appears they’re most powerful at release time and then are gradually “optimized” so every new model feels more powerful. But there’s no evidence on routing to a deployment with other weights. It would be plausible to do so though at least at peak times.
What really is fun is when the opposite of the public outcry about purported distillation happens - when Astra in Codex suddenly responds in Chinese. Now that is fun. Wonder what it’s routing to.
Serious question: does anyone have evidence of this?
It’s something that’s constantly asserted, and has been since 2023. Every time someone posts a site that tries to track this though, I look at it and it’s just a flat line.
https://marginlab.ai/trackers/codex/
By and large they don’t. I have seen this drop a few times, eg before fable came out opus dropped a lot probably due to less compute available.
My guess is it’s a combination of getting used to the new cliff models fall off on and forgetting that model performance drops significantly when context fills up.
So new model comes out, people try it and it’s amazing on a task or two. Then they start using it, context window fills up and it gets a lot worse.
Sol 5.6 xhigh had been a very reliable workhorse for coding for me via the 200 bucks sub.
But this week they seem to have tweaked the system to a point at which all models (Astra, Sol, Luna) hit rate limits all_the_time without me being anywhere close to the weekly limit.
Early results with MiMo 2.6pro are quite encouraging for anything that's non-UI work so likely switching spend for the time being
I have done the same. I hesitated for way too long. I shouldn’t have.
I get way more usage for way less money without any quality or performance degradation. My $200 Codex Pro plan allowance is depleted in 2-3 days. Sometimes Tibo announces a usage reset. But GPT-5.6 models are really not good for coding. Sol has been making increasingly more mistakes in the past two weeks even in the reviewer and advisor roles. Astra is usable for coding but slow and very expensive. In the past two days I’ve used up over 70% on simple copy editing, with dedicated short specs and short sessions. Really little one can do to make it more efficient. Similar work took 20% at most just a month ago. I’m looking to use Astra for milestone reviews/advisory. Perhaps a $100 Pro downgrade will be enough. But my main work is now on open-weight models. And you don’t need to depend on someone to send you a reset. And it’s cheaper by the end of the month too.
With Claude the limits are not even fun anymore - my weekly $100 Max plan quota is gone in one day on merely review invocations, no coding. And my $200 Pro quota is gone in two with some coding. Sonnet 5 is not usable for coding. And Opus 5 tends to always make a couple avoidable mistakes on every task. Fable 5.1 is ok but tends to ignore skills and to work around explicit instructions. Completely canceled all my Claude subscriptions.
With Qwen 3.8, DeepSeek 4.1 Flash, GLM 5.3 I’ve been getting Opus 5-level performance, with less blah blah and no overengineered churn. Public benchmarks are really not telling the real story. The models are more dependable and more predictable. They have their own failure modes. Sometimes DeepSeek 4.1 Flash is quite stubborn but it fails in a good way. Bad for full autonomy - I need to intervene, but it sticks to the rails and instructions - other than Fable and Opus that try to outsmart you and your harness.
Grok is interesting but has been a bit underwhelming on Grok plans - my SuperGrok allowance is depleted in a single session overnight. SuperGrok+ gives more but it’s still about the same as with OpenAI, Claude is way less now.
Since the allowance volume has been shrinking with the major model providers, to me, open-weight alternatives are really necessary now to at least maintain the momentum and budget.
But at work it’s really an uphill challenge - it’s become impossible to convince the tech leadership once they got hooked on Anthropic. They No facts will help. Some people underestimate how expensive Claude really is after getting used to the subscription plans with allowance resets. OpenAI models are expensive too.
Per Xiaomi, MiMo v2.6 training run cost $3.47m. A far cry from the estimated costs ($100m+) for the Big 5 (MSL, xAI, GDM, OAI, Ant). I wouldn't be surprised if salaries and R&D costs have similar drastic disparities.
For a model that matches Muse Spark 1.3 in benchmarks, MiMo v2.6 Pro is incredibly cheap, given its cache rates will remain $0.0036 per million.
I sorta got the impression that the $3.47 million only covered post-training , given that few of the graphs start at zero. Is a barely-trained model going to score 48 on DeepSWE v1.1 ?
https://mimo.xiaomi.com/rl/
That is the RL training cost only. Their announcement blog mentions this: https://mimo.xiaomi.com/mimo-v2-6#scaling-rl-fully-open-sour...
My understanding of tech salaries in China is that they are pretty decent, but not as high as in SF; closer to typical European salaries.
Mostly due to lower cost of living; Shenzhen is way cheaper than SV
I seriously doubt salaries are included. It must be just the electricity and GPU costs.
In these metrics, yes. In the reported training budgets of anthropic/openai, who knows?
where's the flash model? it's out already isn't it
Was going to say the same thing. The announcement blog[1] has the flash benchmarks. AFAICT Artificial Analysis has not added it yet though.
Interested to see if it also beats DeepSeek V4.1 Flash.
[1]: https://mimo.xiaomi.com/mimo-v2-6
Mimo2.5 is really good, but tended to loop too much for my taste. Locally, Pro2.5 wasn't much better. I would reach for it for one shots, hopefully they sorted it out with v2.6, it's a model that's slept on by many. I found that most people that used it did so because it was free. It's a top model worth exploring if you have never given it a go.
Why is it labelled as #1/114 for artificial intelligence (at the top of the page) but if you scroll down and look at the graphs it's obviously not?
Hover the #1/114 and it shows only open weights, that's probably why
Feels a little pointless and disingenuous to present it this way (as default) unless you specifically filter it as such