The weights have yet to be released. The previous MiniMax M3 had (I think) ~428B parameters. Since this one is called Flash and they advertise it as inference being 3 times faster, there's the possibility that this will be a ~150B size model which would (quantized) run on DGX Spark or 128GB Macs or the new AMD AI systems.
The FAQ on their website says Minimax 3.1 flash is 428B total, 23B active
Which is roughly similar to deepseek 4.1 flash.
I find it interesting that we’re calling these “flash” models now, even though they aren’t “small or light” and in many cases aren’t particularly that fast.
I suppose it’s more so that the standard for “regular models” is becoming 1T+
Flash probably means single expert fits into some cost efficient GPU memory, and non flash models require more complex and cost expensive architectures.
Another day where the Pareto frontier is moved up and to the left!
I haven't tried this model yet but Deepseek and the GLM 5.3 series do feel like "enough" for a lot of tasks, it's pretty amazing to see that we're still getting better and cheaper models all the time.
I only wished they would focus more on wall clock time. For a hobby project I do not mind that a cheaper model takes twice as long but at work it's hard to justify using a "cheaper" model if the time per task is so much higher than the frontier models.
The amount of time that it took to complete a benchmark is almost never shown in these kind of releases.
Would place DS 4.1 1017 roughly equivalent to GLM 5.3 which means it's targeted to price rather than direct capability. Probably why they listed DS adjacent in the comparison table when according to their results more expensive models are beaten by it.
Apparently the domain is shadowbanned (https://news.ycombinator.com/from?site=minimax.io) which is why people kept submitting it (and normally they cannot be vouched), and presumably a moderator gave this one their blessing.
The solution is for @dang to unshadowban it because Minimax is now a newsworthy company.
I went to check and Opus 5.5 wasn't benchmarked on DeepSWE1.1, I can't keep track of which are in vogue at the moment but maybe it's thought to be saturated at this point, as they only have older GPT models as well
Minimax has been a strong player in our tests, if they managed to solve the inference speed issue(be scared, be very scared). A decent speed is at least 13 tps in consumer hardware. Let see if it passes Hugston bench this time, we really hope so.
It did okay in Pac-Man. Not great, but a usable example. Unfortunately, it did take half an hour. I'm trying to figure out what the pricing would be if it wasn't free. But zero dollars is a great entry price.
The deepswe and terminalbench4.0 scores don't seem to jive with what many opencode users' experience has been of this model while it was being tested as "space bunny". I saw a lot of anecdotes of people trying it and going back to qwen 3.8-flash-next or deepseek v4 flash because "space bunny" made some rather egregious errors.
> [Space Bunny] was an early version of MiniMax M3.1 Flash.
They mention that space bunny was an earlier checkpoints, so it may have improved (or may it could be benchmaxxed)
I find it interesting that so many models seem to be at almost exactly 74% specifically which makes me suspicious of the benchmark. I want to say 5.6-sol, Astra, deepseek flash 4.1, opus 5 (and 5.5?), Gemini 3.8-flash and a few others all ended up there?
There was talk of deepswe tests being flawed. I have yet to find a better gauge of useful work for my case though, so I still keep referencing it and sampling it. On the lookout for improved benchmarks.
I do agree that it’s generally matched my impressions of models for general SWE at least until we had Opus 5 and 5.6-Sol beating Fable 5. Which coincidentally is when they hit 74%.
Models like Deepseek 4.1 flash, Muse 1.3, and MiniMax 3.1 flash are certainly capable and very cost effective, but I’m not sure anyone would say they’re better than Fable 5/5.1 in capability
The cost vs performance is insane for all these new models. I’ve been using Deepseek Flash 1017 and it does whole projects and asks for $0.10 on US provider with ZDR on Openrouter. I convinced a few friends to drop their Claude $20 subscription because DS Flash 1017 brings so much better value. They can still use Opus 5.5 for planning though each call will cost north of $2. Excited to try this model out too once there are US providers.
I suspect that we may soon be seeing a modern version of VHS vs Betamax. Betamax was better quality, but VHS beat it in virtually every other metric that mattered to most consumers.
(yes, I'm aware the analogy isn't perfect for many reasons)
The comparisons with models like GPT-5.6 or Opus are pointless because they vary substantially. They are often nerfed shortly after release and then performance is not consistent. One day model will be doing relatively fine, another it will be dumb as markov generator. In that regard self hosted models are better, because you know what is running your requests.
I mean the efforts you linked are showing there aren't actual nerfs happening, making the alternative explanation of people just 'growing into' the capabilities of models more likely to me
> For agentic tasks, the model’s overall usage cost is 70% lower than its predecessor.
For this release, the cost is $0.15 input MToK/$0.60 output MTok/$0.015 cache read MTok, In this case the predecessor is MiniMax M3 (not Flash), which was $0.30 input MToK/$1.20 output MTok/$0.06 cache read MTok. So a 50% cut to input costs and a 50% cut to cache read costs (20% of input -> 10% of input) makes the ~70% cost reduction plausible.
A 70% cost reduction is very funny in light of reports saying LLMs are becoming too expensive. That is likely due to the need to compete with DeepSeek V4.1 ($0.15 input MToK/$0.60 output MTok/$0.003 cache read MTok), which this new model still does not beat because DeepSeek's cache costs are still confusingly low, but it's pretty close.
If anything I am confused by how high cache costs still are across many models
The really expensive part of inference - the compute and memory traffic required to perform prompt processing and decode - is paid for roughly once per token in a multi-turn session when prompt caching is working properly.
The difference to the provider's costs between:
- Submitting 100 back-to-back requests, each adding 1k tokens to the context
- Submitting 1 100k token context request
Is negligible. Yet the consumer is going to pay a lot more in the former case for all those repeatedly submitted input tokens that you get charged the cached rate for (but which required no additional compute on the provider side).
DeepSeek V4.1 has an insanely optimized KV-cache that takes around 900MB per million tokens, making it feasible to move to disk between requests. Makes sense they're cheap
The amazing part of DeepSeek's caching is that the TTL is (although not guaranteed) like 8-24 hours. Makes a big difference to how much time I can spend reviewing code, when most everyone else offers maybe 1 hour (for subscribers - often minutes for API users).
I mean, memory prices are high. They ain't getting cheaper. And KVCache is held in memory. The more people are using cache, the more memory you need. It's not free to cache it.
Plus there's complicated cluster-wide KVCache distribution schemes, built around RDMA networking, etc etc.
By far the most expensive hardware cost now of inference is the memory cost. KVCache is still using it. Though obv GPU compute is using more energy
It's not just like... "stick it in Redis, all done!"
> Rather than cost to consumers going down over time, it will likely go up. Not because it can't get cheaper/more efficient - but because it increases companies' profit.
As shown in the OpenRouter links you provided, there are over 30 distinct providers of these open-weights models. They can't raise prices because it's a near-perfect marketplace and if they do, someone will just lower them to compete as long as marginal revenue >= marginal cost. There's no moat or other economic factors to allow for price control.
Think about mobile users. If you compare the raw cost for MVNOs vs MNOs, they have basically the same cost. There is no special moat around text messages, voice calls, data. By law, MNOs have to give MVNOs access to resell their network access; it's literally the same network. Yet MVNOs are 2x-6x cheaper than MNOs. How is that possible? Trickery. MNOs hide fees, bundle "features", lock people into exorbitant prices, and increase costs over time. Nobody today sees the price per minute of calls, text, data. If you look really, really closely, you can compare data caps, international vs domestic. But almost nobody does.
Now back to LLMs: Normal users do not pay API cost, they pay subscription cost. They're going to be the vast majority of users. No normal user is going to look at API provider costs. They're going to pick one of a handful of subscription providers. Those subscription providers will hide the real costs behind the scenes, making deals with inference providers, who will in turn hide those prices for competitive advantage. They will raise prices, and it'll be buried in the subscription fees. This has been happening for a decade with mobile subscriptions, and it's already happening now with LLM subscriptions. (There is no price-list for current LLM subscriptions; it's a total mess that's constantly changing)
Enterprise and government are large-scale users, and they'll pay exactly the same way they pay for compute right now (LLM cost is literally compute cost). Some will self-host, some will pay API pricing, some will rent, some will use "team" or "business" subscriptions. This is also all happening right now. AWS is making bank from Bedrock purely because people in their ecosystem would rather spend money through a known vendor with marketplace credits, and that's just one of a hundred different examples.
To put it in simpler terms: AI is just servers. Normal people do not pay for servers, they pay a service provider for services. Businesses pay for servers, but in many different ways, and definitely not the cheapest options.
MiniMax models have been pretty disappointing. M2.5 was good, M2.7 was meh and came with a terrible license, 3 meh. Maybe this will be better, but I have my doubts. When I say good vs meh vs bad, this is all relative of course. If M3.1 came out a few months ago it would be groundbreaking. Things just move too quick.
I go off of marketing obviously, but how is every new announcement of any vendor always on top of all others and then cheaper, while next day the next best and even cheaper model pops up? Is there anyone who independently verifies this? Its kind of pointless to comment on this announcement, much like how its pointless to comment on all the other announcements...
It's always cheaper (*than the next more expensive model) and better (*than the next worse model).
And often they also don't mention how verbose it is, that the same text splits into more tokens than in competing models, how slow it is, that the pricing is temporary, that the cost per task actually isn't that good despite being both cheap and capable.
Because they selectively choose which benchmarks to highlight and which competing models to compare against; they don't always compare against the SOTA or the Pareto frontier of performance/price, a lot of times you'll see them compare against the last generation of competing models, or the like, so that they look better if you don't carefully consider what they're comparing against.
For example, here you see they are comparing against Opus 5 and GPT-5.6 Sol, when Opus 5.5, Sonnet 5.5, GPT-6 Astra, and GPT-6.1 Sol would be the latest SOTA models from those labs to compare against.
Also because a lab's own benchmarks tend to run about 6.4% better than third party runs of the same benchmark. Source: https://whatmodel.sh/claims This can happen due to a number of factors, and doesn't necessarily indicate anything nefarious. It could just be that they run multiple times while fixing issues that aren't related to the model, and have a selection bias towards stopping when the score is on the higher side. But since you see that sometimes their own scores are lower than the independent runs, some of the variation is explained just by variance in the model solving problems, since LLMs are nondeterministic there can be variation from run to run and you don't see labs always cheating, just potentially a bit of selection bias about when they stop tweaking their benchmark setup and publish the results.
And yes, there are independent third party runs of benchmarks for just this reason. The most popular with broad coverage of a range of benchmarks is https://artificialanalysis.ai/ but there are a number of others as well.
Being pedantic - LLMs are deterministic software, that is chaotic and typically deployed with a PRNG seed. An LLM ran using the same seed will produce the same output consistently (theoretically - in practice this isn’t always the case because of changing factors wrt the execution environment given how sensitive/chaotic models are).
in practice, as actually used, LLMs are nondeterministic.
It is possible, if you control the execution environment sufficiently, and are careful not to use certain kinds of parallelism which can lead to nondeterminism, and you set the seed deterministically and ensure that the prompts are byte identical, you can operate them deterministically.
But no one actually does all of that. Almost everyone uses cloud hosted APIs where the provider could change any number of variables at any time. Everyone wants faster cheaper inference so uses inference engines which may trade off a certain amount of determinism for more speed. Different versions of harnesses and chat templates and benchmark software may introduce small differences in text.
And even though the cloud providers have in the past offered seed and system fingerprints to try to give people a way to control or detect non-determinism (see https://developers.openai.com/cookbook/examples/reproducible...), that has all been deprecated and removed in the newer APIs.
So yeah, while in theory it's possible that if you control the whole stack you can get deterministic results, that's not how anyone is using it, and there is no way for third party benchmarks to ensure deterministic results of hosted models, the latest APIs don't even support it at all.
This whole "well, actually, LLMs are deterministic if you hold them right" is a red herring. In any actual real-world use, they are basically always non-deterministic.
And of course, there are reasons that providers have dropped these attempts at providing deterministic results; because the results are chaotic, perturbed by small changes in input, it's almost never useful to try to make them act deterministic or detect when non-determinism has happend. Instead, it's better to think about them non-deterministically, and look at the distribution of results, rather than just considering a single pass or fail on a single test to be meaningful.
Fair. I wasn’t advocating for “holding them right”, I feel it’s an important nuance to keep in mind when conceptualizing them, i.e. it’s procedural, determistic, yet both inherently and in practice extremely chaotic and therefore not practically deterministic software.
With the old GPT-3 api in the olden days, you could set the seed and get deterministic outputs! Interesting results could be drawn from it, as for instance if you use the prompts "a blue bird"/"a blue bird in the sky" the blue bird would have the same look in both with similar seeds.
I've got a corollary question, which is "how big is the market for these never ending model announcements?" ... I guess its mostly people coding with agents. Most general AI users are just asking ChatGpt to word an email for them or whatever. How many AI users are on here looking for the next model announcement and working out if the price per token is better? Are we talking like 50,000 developers? More? Less?
Openrouter does >600 trillion tokens a month and the top models shift on a daily basis. A lot of people are paying attention to new releases and even pre-releases (which tend to be free)
There are hundreds of benchmarks (all noisy), dozens of model families, and dozens of modes and variants within each. Announcements are crafted by publicity teams and involve cherry picking a chart that looks compelling.
As a prospective user, none of of the benchmarks details will matter to you, but the family of models and settings they peg themselves against can give you a hint of where you might want to evaluate them withing your own workflows.
As an enthusiast or lay researcher, the more interesting thing is to look at their techniques and any innovations they're exploring in training, inference, etc. Those are usually going to be details tucked below the fold and won't be summarized in a benchmark chart.
As anybody else, why do you even care about this in the first place?
This doesn't just apply to AI. You should know better than to take marketing announcements at face value, nor trust that most "influencers" or "journalists" or "commenter" are repeating and discussing those announcements are doing a lot of work as they buzz over what's said.
Don't be a victim to PR spin, but also don't discount potentially interesting work just because it got channeled through a marketing team. Just look for the information you need amidst the noise or move on if there's none there.
I run a benchmark and independently check all new models. They are very spiky.
If you are doing coding only then official benchmarks are mostly right, but as many people are starting to do for instance marketing work with LLMs now there’s very little correlation between coding and creativity for instance.
I have this idea that the big labs internally have much more capable models which are prohibitively expensive to run though, and the competition is mainly around getting the price down by releasing models that are quantized/otherwise made more efficient at the cost of a lot of performance.
On Openrouter at least, all Mino models are painfully slow. TTFT is many seconds and the models spend a lot of time on its CoT, making effective output per second many times slower than DS Flash. On Flash 4.1, I constantly get ~180 tok/s and the real work got done fast, too. So even when the Mimo 2.6 are smart models, I use them much less recently. Not to mention their costs are also higher than DS.
For me, the DeepSeek 4.1 models are currently the ultimate cost-effective model the rest have to match with, especially if you are not on a subscription plan with heavy discount.
GPT Sol 6.1 is however, for me the most efficient model, both in terms of time to finish and token usage.
This model seems to be pretty benchmaxed. In actual use, it doesn't survive the comparison to models in the benchmark.
This model was released as a "stealth model" in OpenCode Go for a while as "Space Bunny" (they mention this in the blog). So people have been using without knowing what the model was exactly, but it failed to impress anyone as far as I can tell. This is despite the model being served for free (with some asterisks). GLM 5.3 Flash and DeepSeek V4 Flash, even before last week's update, seemed to be much preferred. Stealth releases happen often with OpenCode subscriptions, and GLM 5.3 Flash (stealth name "Ox Alpha") impressed a lot for example.
The perverse incentives around benchmarks are now so bad that benchmark scores are effectively meaningless. Models can score highly and still be crap to use. It doesn't matter whether it's Minimax or Google, or anyone really.
This is my experience with all MiniMax models: They're good on paper but fail to live up to expectations in practice. IMO, at no point has any MiniMax model been the superior choice over others, all things considered.
It's Goodhart's Law all over again for LLMs. I would rather trust my gut when trying a new model. Often I throw absurd prompts at it first before deciding if it's worth my time working around its antics.
I am not sure. The people who actually care enough to look at benchmarks will run evals, and find out quickly. If you post silly benchmark scores too often, you will simply be ignored. You are paying with your reputation.
I don’t get the endgame. If your model underperforms in real life usage, the backlash comes quickly. I guess I’m asking what exactly the perverse incentive is? A day of press at most?
I know I mention it way too often right now, but I am just so impressed with Qwen Flash Next right now. I have been using it locally and it's just such a great little model to work with. It makes mistakes and has its quirks, but in general, it just gets things done and its default manner is just right. It is not overly proactive or making messes for me, even compared to DeepSeek Flash 4.1. I really like DS Flash too, it can just be /really/ annoying in some roles and to chat with some of the time.
I can run Qwen Flash Next on hopes, dreams, DDR4, and ancient data center GPUs at 200 t/s :)
Yes, I feel Qwen3.8-FN is the first model that seriously can replace some lower tier subscription usage without requiring much more than a high end gaming PC. Qwen3.6-35B was impressive but not quite good enough, and 27B takes a lot more time.
Lately I prefer GLM5.3-Flash since I feel it is noticeably smarter and I have the hardware for it, but FN's performance has me looking forward to Qwen4.
Same here. I run Qwen3.8 Flash Next on my local machine and combine it with Opus 5.5 for plan making. The Qwen model feels like it's around Opus 4.x level performance most of the time.
But once in a while it gets into long reasoning loops, which is annoying. It seems less with the Swift model but I haven't benchmarked the difference between that one and the original.
For what it's worth, it does perform on par with DeepSeek V4 Pro on my adversarial programming language benchmark[1]. I wouldn't think they would benchmax on that.
Any chance you could try qwen3.8 flash next? That is currently the best local model for actual coding tasks in my + many coworkers experience. It will run reasonably fast if you have 64+ gb ram and use strata or a strata fork (vram not super important with this method).
I've been building a cybersecurity platform (called ZeroQuarry) and regularly benchmark new models against a pretty wide set of security both code review tasks and live application pen testing tasks, and this one far underperforms relative to GLM 5.3 from every test I've done for real-world cybersecurity tasks
Yeah, ollama has it up for free at the moment. I'm fairly nonplussed about it.
Got it to do a code review on a couple of codebases where I knew roughly what the problems were and it underperformed even with a prod in the correct direction. Its code output is average, and it's more expensive than glm-5.3-flash, I don't see myself using it when it comes off free.
I used it as Space Bunny, was very good at documentation, structuring and organizing the things, but was not as concise and direct as DeepSeek 4.1 Flash in coding, we could not focus on code part as a human and AI team from how many approves are needed for each step discussion.
The truncated English is interesting. I noticed something similar from the OpenAI blackhat talk a couple months ago. Is this pretty standard at this point?
I'm finding is very hard to figure out when it's better to go for a cheaper dumber model vs an better expensive one when the expensives ones get stuff done quicker.
I'm finding these LLM announcements to be pretty meaningless. They all look the same, with different numbers and graphs, even though the graphs are almost all identical. The offered model is compared with models that it can beat, with a couple of entries that "show" they haven't engineered the whole benchmark presentation to mislead.
I've become incredibly skeptical of new model announcements until I can find evidence of someone else using it for something effective. I'd love to see an announcement that wasn't structured around attracting investors with too much money and too little domain knowledge.
To me the most interesting chart on artificial analysis is intelligence index vs cost per task. It's like the third chart down and they graph a pareto line showing where the frontier is and every model on the pareto line has been interesting to me. Luna, Mimo-v2.6-pro, Sol. Hell, Argon.
This MiniMax model is clustered with Haiku 5.5 so (combined with the criticism on here that it was lacklustre while free on opencode go) I don't find it interesting, but every announcement has a chance at pushing that frontier.
I was benchmarking affordable LLMs for japanese->english translation recently and the winner was Kimi K3 - there are a ton of interesting models in the wake of the flagships.
Skepticism is due! Good! The numbers don't mean much! Yes!
I personally enjoy seeing how different models view problems. Take big plans from big models, then see how other agents reinterpret a stance. See which directions they go. I love the diversity of views they produce.
It's not really relevant what the numbers are. I really really enjoyed talking with MiniMax M3 a lot. And M2.7. it had a good world view, wrote the situation up incredibly fluently. It thought out loud well. These were pleasant uses for me. The realities, the world's they could see felt very grounded. Good stuff, to me.
Qwen 3.8 flash next was an important announcement because it easily runs on dual RTX 6000 Pro, which made it available to a subset of fairly high-end local inference tinkerers. For those running local models, checking announcements for vram requirements is the first step in knowing if it’s a candidate to push more intelligence into your local rig. At least, that’s what I do when I read announcements.
"I'm finding these LLM announcements to be pretty meaningless. They all look the same, with different numbers and graphs, even though the graphs are almost all identical."
How dare you, sir/ma'am! the LLM spent a whole 0.9 tokens creating the numbers and graphs!!
I have both platform.deepseek.com (hooray off-peak pricing) and the $100/m Claude subscription and boy do I find myself mostly using DS-flash pretty much every time.
There's something to be said for the speed with which it gets word done - I take fewer "it's compiling" breaks, and lose my train of thought less frequently. What's crazy is that the speedy work it does is pretty much always correct - which is a reason why I don't feel the need to spin up claude-code and "ask the expert".
This speed & good-work combo is the DS signature IMO
I use DS-4.1-Flash a lot but find it often gets caught up in long thinking runs. The "But wait... Actually... Hmm..." type of thing. Even though the tok/s is good, it stills feels slow because of this. Same with 1017.
I use Opus 5.5 at work and it does this much less (I know, different classes). Maybe it is to be expected of the flash models.
Which claude model are you comparing it to? If you're using fable max obv it's going to be slower, but like sonnet high is both faster and more intelligent
126 comments
[ 0.32 ms ] story [ 15.7 ms ] threadWhich is roughly similar to deepseek 4.1 flash.
I find it interesting that we’re calling these “flash” models now, even though they aren’t “small or light” and in many cases aren’t particularly that fast.
I suppose it’s more so that the standard for “regular models” is becoming 1T+
https://agent.minimax.io/tools/m3-1-flash-preview
At 428B it can still fit on 2x 128GB machines at 4bpw, which is nice.
I haven't tried this model yet but Deepseek and the GLM 5.3 series do feel like "enough" for a lot of tasks, it's pretty amazing to see that we're still getting better and cheaper models all the time.
I only wished they would focus more on wall clock time. For a hobby project I do not mind that a cheaper model takes twice as long but at work it's hard to justify using a "cheaper" model if the time per task is so much higher than the frontier models.
The amount of time that it took to complete a benchmark is almost never shown in these kind of releases.
https://news.ycombinator.com/item?id=50151561 https://news.ycombinator.com/item?id=50151576 https://news.ycombinator.com/item?id=50153196 https://news.ycombinator.com/item?id=50153612 https://news.ycombinator.com/item?id=50153725
The solution is for @dang to unshadowban it because Minimax is now a newsworthy company.
https://artificialanalysis.ai/models/claude-opus-5-5#intelli...
https://deepswe.datacurve.ai/
Unless they come in at half the price (which I think will be impossible) on inference ... they have wasted their training compute imho.
https://jonclegg.github.io/pacman-bakeoff/#minimax-m3.1-flas...
They mention that space bunny was an earlier checkpoints, so it may have improved (or may it could be benchmaxxed)
I find it interesting that so many models seem to be at almost exactly 74% specifically which makes me suspicious of the benchmark. I want to say 5.6-sol, Astra, deepseek flash 4.1, opus 5 (and 5.5?), Gemini 3.8-flash and a few others all ended up there?
There was talk of deepswe tests being flawed. I have yet to find a better gauge of useful work for my case though, so I still keep referencing it and sampling it. On the lookout for improved benchmarks.
"We found issues in at least 23 out of 113 tasks (over 20.3%), leading us to designate the benchmark as Flawed." - https://epoch.ai/benchmarks/deepswe/review
Models like Deepseek 4.1 flash, Muse 1.3, and MiniMax 3.1 flash are certainly capable and very cost effective, but I’m not sure anyone would say they’re better than Fable 5/5.1 in capability
(yes, I'm aware the analogy isn't perfect for many reasons)
For this release, the cost is $0.15 input MToK/$0.60 output MTok/$0.015 cache read MTok, In this case the predecessor is MiniMax M3 (not Flash), which was $0.30 input MToK/$1.20 output MTok/$0.06 cache read MTok. So a 50% cut to input costs and a 50% cut to cache read costs (20% of input -> 10% of input) makes the ~70% cost reduction plausible.
A 70% cost reduction is very funny in light of reports saying LLMs are becoming too expensive. That is likely due to the need to compete with DeepSeek V4.1 ($0.15 input MToK/$0.60 output MTok/$0.003 cache read MTok), which this new model still does not beat because DeepSeek's cache costs are still confusingly low, but it's pretty close.
The really expensive part of inference - the compute and memory traffic required to perform prompt processing and decode - is paid for roughly once per token in a multi-turn session when prompt caching is working properly.
The difference to the provider's costs between:
- Submitting 100 back-to-back requests, each adding 1k tokens to the context
- Submitting 1 100k token context request
Is negligible. Yet the consumer is going to pay a lot more in the former case for all those repeatedly submitted input tokens that you get charged the cached rate for (but which required no additional compute on the provider side).
DeepSeek V4.1 has an insanely optimized KV-cache that takes around 900MB per million tokens, making it feasible to move to disk between requests. Makes sense they're cheap
Plus there's complicated cluster-wide KVCache distribution schemes, built around RDMA networking, etc etc.
By far the most expensive hardware cost now of inference is the memory cost. KVCache is still using it. Though obv GPU compute is using more energy
It's not just like... "stick it in Redis, all done!"
As shown in the OpenRouter links you provided, there are over 30 distinct providers of these open-weights models. They can't raise prices because it's a near-perfect marketplace and if they do, someone will just lower them to compete as long as marginal revenue >= marginal cost. There's no moat or other economic factors to allow for price control.
Think about mobile users. If you compare the raw cost for MVNOs vs MNOs, they have basically the same cost. There is no special moat around text messages, voice calls, data. By law, MNOs have to give MVNOs access to resell their network access; it's literally the same network. Yet MVNOs are 2x-6x cheaper than MNOs. How is that possible? Trickery. MNOs hide fees, bundle "features", lock people into exorbitant prices, and increase costs over time. Nobody today sees the price per minute of calls, text, data. If you look really, really closely, you can compare data caps, international vs domestic. But almost nobody does.
Now back to LLMs: Normal users do not pay API cost, they pay subscription cost. They're going to be the vast majority of users. No normal user is going to look at API provider costs. They're going to pick one of a handful of subscription providers. Those subscription providers will hide the real costs behind the scenes, making deals with inference providers, who will in turn hide those prices for competitive advantage. They will raise prices, and it'll be buried in the subscription fees. This has been happening for a decade with mobile subscriptions, and it's already happening now with LLM subscriptions. (There is no price-list for current LLM subscriptions; it's a total mess that's constantly changing)
Enterprise and government are large-scale users, and they'll pay exactly the same way they pay for compute right now (LLM cost is literally compute cost). Some will self-host, some will pay API pricing, some will rent, some will use "team" or "business" subscriptions. This is also all happening right now. AWS is making bank from Bedrock purely because people in their ecosystem would rather spend money through a known vendor with marketplace credits, and that's just one of a hundred different examples.
To put it in simpler terms: AI is just servers. Normal people do not pay for servers, they pay a service provider for services. Businesses pay for servers, but in many different ways, and definitely not the cheapest options.
Ask HN: Any guesses on which model/provider is Space Bunny Alpha on OpenRouter? - https://news.ycombinator.com/item?id=49898971 - Sept 2026 (1 comment)
H3, on other hand, is phenomenal
And often they also don't mention how verbose it is, that the same text splits into more tokens than in competing models, how slow it is, that the pricing is temporary, that the cost per task actually isn't that good despite being both cheap and capable.
So, I can quickly judge that this is an interesting option only if cheaper or faster than DeepSeek.
Based on Terminal Bench 4.0, I'd also guesstimate behind GPT 5.6 Sol, so of course behind the SOTA as expected.
For example, here you see they are comparing against Opus 5 and GPT-5.6 Sol, when Opus 5.5, Sonnet 5.5, GPT-6 Astra, and GPT-6.1 Sol would be the latest SOTA models from those labs to compare against.
Also because a lab's own benchmarks tend to run about 6.4% better than third party runs of the same benchmark. Source: https://whatmodel.sh/claims This can happen due to a number of factors, and doesn't necessarily indicate anything nefarious. It could just be that they run multiple times while fixing issues that aren't related to the model, and have a selection bias towards stopping when the score is on the higher side. But since you see that sometimes their own scores are lower than the independent runs, some of the variation is explained just by variance in the model solving problems, since LLMs are nondeterministic there can be variation from run to run and you don't see labs always cheating, just potentially a bit of selection bias about when they stop tweaking their benchmark setup and publish the results.
And yes, there are independent third party runs of benchmarks for just this reason. The most popular with broad coverage of a range of benchmarks is https://artificialanalysis.ai/ but there are a number of others as well.
Being pedantic - LLMs are deterministic software, that is chaotic and typically deployed with a PRNG seed. An LLM ran using the same seed will produce the same output consistently (theoretically - in practice this isn’t always the case because of changing factors wrt the execution environment given how sensitive/chaotic models are).
It is possible, if you control the execution environment sufficiently, and are careful not to use certain kinds of parallelism which can lead to nondeterminism, and you set the seed deterministically and ensure that the prompts are byte identical, you can operate them deterministically.
But no one actually does all of that. Almost everyone uses cloud hosted APIs where the provider could change any number of variables at any time. Everyone wants faster cheaper inference so uses inference engines which may trade off a certain amount of determinism for more speed. Different versions of harnesses and chat templates and benchmark software may introduce small differences in text.
And even though the cloud providers have in the past offered seed and system fingerprints to try to give people a way to control or detect non-determinism (see https://developers.openai.com/cookbook/examples/reproducible...), that has all been deprecated and removed in the newer APIs.
So yeah, while in theory it's possible that if you control the whole stack you can get deterministic results, that's not how anyone is using it, and there is no way for third party benchmarks to ensure deterministic results of hosted models, the latest APIs don't even support it at all.
This whole "well, actually, LLMs are deterministic if you hold them right" is a red herring. In any actual real-world use, they are basically always non-deterministic.
And of course, there are reasons that providers have dropped these attempts at providing deterministic results; because the results are chaotic, perturbed by small changes in input, it's almost never useful to try to make them act deterministic or detect when non-determinism has happend. Instead, it's better to think about them non-deterministically, and look at the distribution of results, rather than just considering a single pass or fail on a single test to be meaningful.
Found the tweet: https://x.com/karpathy/status/2109361546505966046?lang=en
As a prospective user, none of of the benchmarks details will matter to you, but the family of models and settings they peg themselves against can give you a hint of where you might want to evaluate them withing your own workflows.
As an enthusiast or lay researcher, the more interesting thing is to look at their techniques and any innovations they're exploring in training, inference, etc. Those are usually going to be details tucked below the fold and won't be summarized in a benchmark chart.
As anybody else, why do you even care about this in the first place?
This doesn't just apply to AI. You should know better than to take marketing announcements at face value, nor trust that most "influencers" or "journalists" or "commenter" are repeating and discussing those announcements are doing a lot of work as they buzz over what's said.
Don't be a victim to PR spin, but also don't discount potentially interesting work just because it got channeled through a marketing team. Just look for the information you need amidst the noise or move on if there's none there.
*crafted by LLMs with a list of PR guidelines in a markdown file
If you are doing coding only then official benchmarks are mostly right, but as many people are starting to do for instance marketing work with LLMs now there’s very little correlation between coding and creativity for instance.
https://artificialanalysis.ai/models/comparisons/mimo-v2-6-p...
For me, the DeepSeek 4.1 models are currently the ultimate cost-effective model the rest have to match with, especially if you are not on a subscription plan with heavy discount.
GPT Sol 6.1 is however, for me the most efficient model, both in terms of time to finish and token usage.
This model was released as a "stealth model" in OpenCode Go for a while as "Space Bunny" (they mention this in the blog). So people have been using without knowing what the model was exactly, but it failed to impress anyone as far as I can tell. This is despite the model being served for free (with some asterisks). GLM 5.3 Flash and DeepSeek V4 Flash, even before last week's update, seemed to be much preferred. Stealth releases happen often with OpenCode subscriptions, and GLM 5.3 Flash (stealth name "Ox Alpha") impressed a lot for example.
The perverse incentives around benchmarks are now so bad that benchmark scores are effectively meaningless. Models can score highly and still be crap to use. It doesn't matter whether it's Minimax or Google, or anyone really.
I am not sure. The people who actually care enough to look at benchmarks will run evals, and find out quickly. If you post silly benchmark scores too often, you will simply be ignored. You are paying with your reputation.
i don't know about that. it seems like the silly benchmarks keep getting paid attention to.
I can run Qwen Flash Next on hopes, dreams, DDR4, and ancient data center GPUs at 200 t/s :)
Lately I prefer GLM5.3-Flash since I feel it is noticeably smarter and I have the hardware for it, but FN's performance has me looking forward to Qwen4.
But once in a while it gets into long reasoning loops, which is annoying. It seems less with the Swift model but I haven't benchmarked the difference between that one and the original.
1 - https://bench.killswitch-lang.org/
https://github.com/Niko1221/Strata
They genuinely try to improve usability, without spending time on marketing or benchmaxxing.
Except for v4 0731, which was a regression, and probably the reason they released v4.1 ahead of schedule.
Got it to do a code review on a couple of codebases where I knew roughly what the problems were and it underperformed even with a prod in the correct direction. Its code output is average, and it's more expensive than glm-5.3-flash, I don't see myself using it when it comes off free.
2. Add "Flash" to the end
3. ?
4. Profit
DHS now plays AI whack a mole, apparently.
[0] https://public-inspection.federalregister.gov/2026-21447.pdf
I've become incredibly skeptical of new model announcements until I can find evidence of someone else using it for something effective. I'd love to see an announcement that wasn't structured around attracting investors with too much money and too little domain knowledge.
This MiniMax model is clustered with Haiku 5.5 so (combined with the criticism on here that it was lacklustre while free on opencode go) I don't find it interesting, but every announcement has a chance at pushing that frontier.
I was benchmarking affordable LLMs for japanese->english translation recently and the winner was Kimi K3 - there are a ton of interesting models in the wake of the flagships.
I personally enjoy seeing how different models view problems. Take big plans from big models, then see how other agents reinterpret a stance. See which directions they go. I love the diversity of views they produce.
It's not really relevant what the numbers are. I really really enjoyed talking with MiniMax M3 a lot. And M2.7. it had a good world view, wrote the situation up incredibly fluently. It thought out loud well. These were pleasant uses for me. The realities, the world's they could see felt very grounded. Good stuff, to me.
I use it at 60+ tok/second on a single 24GB card.
How dare you, sir/ma'am! the LLM spent a whole 0.9 tokens creating the numbers and graphs!!
I have both platform.deepseek.com (hooray off-peak pricing) and the $100/m Claude subscription and boy do I find myself mostly using DS-flash pretty much every time.
There's something to be said for the speed with which it gets word done - I take fewer "it's compiling" breaks, and lose my train of thought less frequently. What's crazy is that the speedy work it does is pretty much always correct - which is a reason why I don't feel the need to spin up claude-code and "ask the expert".
This speed & good-work combo is the DS signature IMO
I use Opus 5.5 at work and it does this much less (I know, different classes). Maybe it is to be expected of the flash models.