I've been saying efficiency is the next "frontier" in AI, at least for LLMs that people use daily. Companies have started to really balk at token costs from the major providers, and there's some evidence that the cheaper Chinese models are chipping away at Anthropic/OpenAI dominance from below (as cheaper Chinese products have done in other industries for many years). I continue to think that the vendor that figures out how to efficiently serve a Good Enough model at a much lower price will be the one to win in the end.
DeepSeek is remarkably efficient at caching and their cached token rates are crazy cheap; using it with Reasonix is free real estate, like 97% cached tokens, ends up costing like 30 cents an hour to use DeepSeek V4 Pro. I hadn't dug into MiMo's caching behavior as I haven't used it as heavily as DeepSeek, but this indicates it's close to DeepSeek.
At this point I don't see a reason to use Sonnet, Haiku, or the smaller GPT models, because their API rates are much higher than the best models from MiMo and DeepSeek.
We're still figuring out the upper bounds of capability and I am still finding next generation models are unlocking things I couldn't readily accomplish before and I'm willing to pay more for them (at least, I'll pay the $100 or $200 subscription rates for them, I couldn't justify the token expense for most of my dev work), but we're already at a point where someone building standard CRUD web apps doesn't need the top models and probably doesn't benefit much from using them.
I assume they're jealous of the Fable/Mythos hype. People talk about Fable like it's a whole new thing, rather than another incremental improvement over the existing best models (which has happened several times and continues to happen).
Yeah, I pretty much had to switch to using GPT rather than Opus completely for all my security benchmarking and harness development. I was annoyed enough to blog about it: https://swelljoe.com/post/why-i-had-to-switch-to-gpt/
Gemma 4 is the first really small model that feels smart, to me. I mean, Qwen 3.6 is arguably better at some coding tasks. But, Gemma 4 has shockingly good reasoning for a small model. Even the tiny 12B, at 7GB on disk in the 4-bit QAT quant, feels like a really big model of a couple years ago. It's a good tool user, can search the web (when given the appropriate skill or MCP), has good vision capabilities, and pretty good prose.
I've only used the Hy3 preview, so I don't want to judge too harshly, yet. But, I wasn't very impressed with it a couple of months ago.
Nearly every model can be found on OpenRouter and used with a single key. Meta Spark is not among them, but Grok and almost every other model is. That's how I try models I don't already have an account for.
MiMo v2.5 Pro is very spiky, in my experience. Sometimes excellent, sometimes mediocre. Weirdly high non-deterministic behavior. Run the same task three times, get three different results. I mean, they're all rolling dice for the next word, but MiMo seems to run hot on the randomness dimension in my benchmarks.
But, it performs very well for its size. I just looked it up, and it's much smaller than I thought it was when I was testing it. 310B A15B is tiny for how well it performs. I guess that explains why it's so cheap.
The Hy3 preview has been a mediocre performer in my benchmarks of security auditing with models, and yes, it is outperformed by Gemma 4 (31b soundly beats it, the MoE does slightly better, even at 4-bit quantization when using the QAT version). Qwen 3.6 27b also beats it.
I'll try it again now that it's out of preview and has been updated with more post-training. It presumably can't be worse, so maybe it's better enough to compete with a 31b model.
Yeah, I guess I could also write the code myself, if all the models refuse. But, it seems worrying to have a handful of the largest corporations and a few governments having access to the best models while the rest of us are using hobbled ones.
So far, we're not in that boat. I've had two models refuse to participate, but most just do what they're told.
I mean, I'm sure there are people within Google who are behaving as though they can keep the dream of the 00s alive in Mountain View, but there's also a whole bunch of people doing work at the frontier in AI. Google has a large lead in hardware, they have the smartest very small models, they have among the most efficient large models (I'd wager their margins on Gemini 3.5 Flash inference are absurd). They have among the best image, video, and audio models, going in every direction (generating, editing, understanding).
Viewed from a consumer lens, the AI the average person interacts with daily, Google seems like the clear leader, especially after locking in Apple as a customer for iPhones.
I've found in my current work on a security auditing harness and benchmarks, both Fable and Opus are useless. I recently switched to using GPT for Nelson and the security benchmarks I've been doing because Opus started refusing to do the work. I guess I probably could also use GLM or DeepSeek or MiMo, and I'll probably do some experiments to see the shape of all of their guardrails in this area soon, now that I see it's more than one model that behaves this way (Gemini in Antigravity also refuses any security auditing task, even as simple as "find security bugs").
Google is playing a different game. I don't really know what game they're playing, but they're not trying to beat Claude Code. They have coding capabilities and Antigravity, but I'd be surprised if it's much more than an afterthought. They're focusing on efficiency, models at the edge, human interaction, image and video, etc. in ways Anthropic, in particular, is not.
Google wants its AI to be pervasive in everyone's daily life. Merely being the best at coding is not how you get there.
I am more bullish on Google in AI than most folks, I think, as they have been focused on efficiency in a way most US vendors have not. They've published a ton of papers on ways to make LLMs more efficient and capable on smaller devices..
Google wants to own the on-device market for AI, and I don't see many credible competitors in that space.
Yeah, the stuff they preinstall is crazy. It's a ~16GB OS image, which is a huge pain in the ass (in addition to all the other things that are a pain in the ass about Nvidia's OS situation); I can't ship their 16GB pile of crap plus our custom 3GB of stuff on my robots, there's no reasonable way to do OTA upgrades with a pile of shit like that. It's just an unpleasant experience all around. What I always want is a nice clean basic install, and then my custom packages of everything I need goes in. Especially for embedded cases, like the Jetson line is allegedly intended for, their whole approach is just stupid, and they don't provide any documentation for how to not use their stupid thing.
Still, I'd rather have the blazing prefill speed of the Nvidia over my Strix Halo, but I wasn't willing to spend nearly twice as much for it (when I bought, the Strix Halo was $2100 and the GX10 was going for $3700, I think). Now that the difference between the two is much smaller, there's no reason to get the AMD.
This particular situation just feels like Fable is able to figure out that it's in a simulation, so it's just playing the game it's been put in.
That's not say I think a machine should ever be in a situation where it is allowed to make ethical decisions with real world results, I don't. At least, not given the current basis of the technology. I mean, generally, I don't see how LLMs can ever be capable of making ethical decisions or trustworthy in that role, no matter how good they get at what they're good at. There would need to be a fundamental change in how AI works for me to change my opinion on this, I think. They are ephemeral, they can never experience consequences, they can never want or need anything, thus there is no mechanism for them to take responsibility for decisions.
Anyway, I think the Andon Labs stuff is kind of a stunt, mostly, and I wish somebody would give me a few million bucks to dick around with the little thinky guys in my computer letting them do silly things.
This is couched in prepper nonsense, but it's got LLM, WikiPedia, maps, etc. A bunch of genuinely useful stuff to keep on a USB stick or whatever: https://www.projectnomad.us/
But, the current model you really want for an emergency kit is Gemma 4 12B QAT 4-bit. At ~7GB on disk, it's small enough to run on a tablet or any modern computer, slowly if you don't have a GPU or modern Apple silicon, but exceedingly smart for its size, excellent vision capabilities, good tool user, surprisingly good reasoning.
I work on Virtualmin. You can find it at http://www.virtualmin.com
I blog at https://swelljoe.com
I toot at https://mas.to/@swelljoe