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 V4.1 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 used Space Bunny from some providers that offered it for free and it was a very much a "you get what you pay for" experience. If the release is comparable, I'll be sticking with Deepseek v4 for sure.
I’m one of the unimpressed. I tested SpaceBunny intensely on a research project for a week.
The model wasn’t horrible at coding, it just made a ton of initial mistakes—but then would catch most of them later on. It can get work done if coaxed, but its peers are just cheaper and/or easier to work with.
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.
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.
Opus 5.5 is according to AA cost per task, 23x more expensive than Deepseek with max effort (which is dumb, you shouldn't use) and 13x more expensive on xhigh.
Claude sub has a 60x subsidy, so a $20 sub gives you $1200 in API.
Based on this math, Claude sub is a much better deal than Deepseek. And it's a stronger model of course.
> Claude sub has a 60x subsidy, so a $20 sub gives you $1200 in API.
This is only true if you max out the plan, which requires working around awkward 5-hour and weekly limits. For side projects, sure, you can use a Ralph loop or whatever and just continue when the window resets. For work, I can't afford to simply stop working when I hit my limit. And using Claude at API prices is prohibitively expensive.
Personally, I have a $20/month OpenAI subscription that I supplement with DeepSeek V4.1 Flash (currently using CoralBricks as the provider since it's fast and I'm grandfathered into their $0 cache read pricing). It's working out rather well so far. Even when I use DS all day as my primary model, I only pay about $1/day, which is roughly equivalent to what I pay for the sub anyway (but with no limits).
I am sorry, but if you're given $1200 in api for free I wouldn't consider that as working around flakey limits especially when you only need to max out the 5 hour limit 5 times a week.
I don't understand what you're saying. If I hit my 5-hour limit in 2 hours, I still need to get work done for the other 3. I'm not going to sit there and wait for my limit to reset.
I pay API rates via OpenRouter / Fireworks, and I find it challenging to hit the $20 mark.
If I was to do anywhere near that kind of work on Anthropic, I'd find myself running out of tokens, waiting days to continue working, and leaning on their el-crap models to try to avoid running out again. This happens all the time.
This way I get to leverage DS for nearly all my work, benefit from its far superior latencies, and I still get to lean on OpenAPI / Anthropic or whatever SOTA models I want when I need them.
The only time that really falls down is if I'm doing huge amounts of real novel work, which hasn't been the case for a while.
Unfortunately, Fireworks is not keeping up with new models the last few months. When we selected them as a vendor, they were crushing it, new models were available in 1-2 days. Now they want us to "request" them in Discord, instead of making them all available.
I use OpenCode at home, this is one of the reasons.
Their Ember-1 model (Kimi K3 tuned to think less) hallucinates to high hell, I have no trust of it, nowhere to provide feedback, despite their launch post saying they would like feedback. That they loudly and proudly claim their internal devs notice no difference is a huge red flag for me. It says there is too much eyes free coding, I noticed the hallucinations and false claims of task completion immediately.
> isn't that what ZDR is for? (assuming you trust that)
ZDR is:
1. A pinky-promise you have no way of verifying
2. A contract term, which means its still subject to the good-old "lawful request or court order" clauses
3. Subject to arbitrary provider exceptions ... see Anthropic and Fable
What you really want (assuming you can't afford your own servers and datacentre space) is secure enclave processing:
Just because anthropic charges absurd consumer API prices, doesn't mean their subscription is a 'subsidy' lmao.
I'm a heavy Claude/anthropic user, I max out my 20x max plan in 1-2 days at this point. The OpenCode Go $40 plan with DeepSeek flash is like 2x the value of the Claude max 20x plan for how much I actually get out of it, for 1/5 the price.
DeepSeek literally like 10x better value than Claude/Anthropic.
I'm the same way re: bottoming out my 20x Max plan in ~2 days. I'm having better luck lately, though. What helped was using Opus 5.5 for planning and orchestration, and Haiku 5.5 for all of the simple stuff. Claude doesn't default to this -- the default is that it uses the same model for everything -- you have to give it specific instructions.
My math takes that into account. It's 20x more expensive but you get 60x subsidy on subs. I.e you can do more tasks on a $20 claude sub than with $20 in deepseek API.
It's just pointing out how heavily subsidized subscriptions are. I often read about how much value people get from DeepSeek, but when I try to use it, I end up spending the amount I'd pay for a subscription much faster.
DeepSeek is really cheap when you compare API pricing, but it's more expensive than a subsidized Anthropic or OpenAI subscription.
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...
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 within 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" 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.
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.
In many cases they are not deterministic even with a deterministic seed because many runtimes embrace races that interact with floating point rounding to improve latency.
You have to read between the lines (or wait for someone on HN to do it for you).
It's never "cheaper and faster and more intelligent", it's pick 2. Sometimes (as was the case this month), it was "cheaper per token, but reasons five times as much", so your costs go up.
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.
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)
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.
> 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% proportional 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).
That still isn't the whole story. Compare DS 4.1 Flash [1] to GLM 5.3 Flash [2], and consider the effective price. DS's price is lower than GLM on input tokens, but higher than GLM on output tokens. Output price is an order of magnitude higher than input price. So for output-heavy workloads and long reasoning, DS costs more than GLM.[1] https://openrouter.ai/deepseek/deepseek-v4.1-flash#pricing [2] https://openrouter.ai/z-ai/glm-5.3-flash#pricing
Yes the official DS provider has a low cache cost, for now (it's just distributed KV). But people are still concerned about ZDR, sovereign AI, getting more model choice in their subscription price, etc. They want more bang for their buck. So when a new model comes out, the bigger impact is really its performance across providers, not the price at a single provider.
And in about 5 years (or less), API prices will stop being a thing. We all used to pay per-minute for phone calls, texts, data. Nobody does anymore. Virtually every mobile provider is now a flat prepaid or contract fee. Carriers improved their technology (going from analog to digital), and it turned out consumers will spend more money on a packaged subscription than on pay-per-use. That's why everything in the world is a subscription now.
Your modern $70 big wireless plan isn't giving you $70 of value. Similarly, your $20 AI subscription that's a great deal right now, will be less valuable in 5 years, when hardware & models are more efficient. You'll still be able to buy a local AI rig, but it will still be very expensive, due to AI models continuing to be large, and NVidia keeping new hardware expensive, as these things together reinforce their dominance and profits, and our reliance on them.
This also follows another pre-existing trend. Why do people keep having to buy a new iPhone or laptop every other year? Because each year companies make software & data less efficient/more bloated, so people need a more powerful machine to run it, which enables them to make even more bloated software. Almost everything people do with a computer today is what they did 15-20 years ago; it's just a higher bit-rate. Despite this, people keep re-buying the same machine, so companies can keep extracting more wealth.
AI works the same way. 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. This isn't just a theory, btw; the US economy's growth is now almost entirely reliant on this happening.
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.
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.
20 t/s is about the speed you can read at, so slower than that and it always feels like it's buffering. 35 t/s is where it starts feeling decent and 50 t/s feels nice.
I know, it is never enough, more, more, give me more. But the fact is that no one has broken though yet to offer a frontier like open weight over 13 tps in consumer hardware. It would be usable at 13tps for me, especially considering the benefits and advantages of privacy, infinite tokens, availability and no refusals/restrictions.
To this day Minimax and GLM are very strong coders but have the same issue, inference speed. We need a real break through, one that thinks out of the box (Tesla turbines style perhaps).
If all you're going to do is chat with an AI, you don't need a 400B parameter MoE model.
13 TPS is unusable for agentic coding, because if decode is that slow, then prefill is absolutely abysmal, and agentic work live and die by prefill speeds.
I’ve been agentic coding at home with like 15 TPS on average. I don’t mind it at all. I can give my agent a big task or a list of little ones and just… do something else. Why do I need to watch it work?
Seconding this. 2-bit quant of DSv4Flash 0731 (Unsloth) running on lemonade-llamacpp on a Corsair Strix Halo. OpenCode had an issue giving me 400s constantly for the past few weeks, so I just went all-in on local dev with omp. Yes, it takes time, but also: it works quite well. I've since gotten OpenCode to behave, so I'm back to using that quite a bit, but the local machine is really good as a "smol" model in omp, for example.
It doesn’t. Decode is by far the most dominant and expensive part of agentic tasks and once prefill hits a certain achievable speed decode dominates the time in ant agentic flows with reasoning. There is a reason input is so much cheaper :)
Isn’t input so much cheaper because it’s just less work? Prefill is usually 10-20x as fast as generation.
If you have 200k context, but only 200t/s prefill, then you’ll wait multiple minutes when continuing an average session from yesterday. Or compaction - might take 15 minutes.
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
The recent DeepSeek release used Opus 5 as a reference point as well. I don't understand it. It's not like they're referencing superseded releases of other open models.
I think it clearly points to where they expect themselves to be, Opus 5.5 and Astra or Sol 6.1 are models that shouldn't exist at those prices without the price competition.
And no new releases from Chinese labs likely means their models that they could compare against those models are still baking largely because they lack compute so severely they need to train models for far longer, or figure out training on Chinese hardware.
Either way it's clear open models a lagging by ~1-2 months compare to frontier, depending on what frontier means to a person. Honestly most new models are just not that much better than at edge cases, they still fail a lot for the kind of stuff I do, but those edge cases matter a lot for un-interrupted long horizon work.
I personally am starting to see a wall here in llm development but that might just be me, given that people are being forced to scale models up considerably for frontier performance.
It's scary thinking that modern models will likely be at best 2-3% better in benchmarks or only useful in 1-2 unique edge cases compared to base models for 100x cost increase in compute required, it's getting absurd.
I can't imagine how Fable models can retain any value once you have to go 100x larger, while 27-32B models start to catch up to the current frontier for coding or other narrow/focused tasks in a year or so.
These models exist because better models exist, these models can't exist without better models existing, you will never beat the frontier with smaller model sizes. Because all these models are trained again larger models, generally a teacher model of sorts.
Lots of synthetic data, distilled data etc.
The problem is if we can pack a Fable in Qwen 4 27B who will pay subscription prices?
Synthetic data, distilled data, etc., does not necessarily mean data coming from outside of an organization. When we see performance increases with models like Haiku, it is because they are able to leverage the knowledge of larger architectures.
I fully expect that we will see increased performance from smaller architectures (to a point) simply because they are refined using larger ones.
> The problem is if we can pack a Fable in Qwen 4 27B who will pay subscription prices?
This is available to everyone developing these systems. Prices per relative unit of intelligence will continue to decrease. And market economics will continue to drive the cost down.
It's not really if you mean to point to the bitter lesson, it was models/intelligence scales better with general/simple methods with more compute better than it scales with any hand-crafted human program/method.
Ala intelligence is a function of compute, there was no guarantee that you need to 10x compute for 10% better performance or something like that.
If you ask the labs they will say the scaling is fairly linear or atleast sub-quadratic, they definitely won't say it's exponential.
No one was thinking we need to 10x compute from here again for another 10-20% increase in performance.
It's only been since Opus 4.6 that the compute growth vs improvements are starting to feel this way.
For instance GPT-3 vs GPT-4 was practically a 10-100x jump for a 10-100x increase in compute, we went from broken sentences to fully formed essays writing, speech and so on.
GPT-4 vs GPT-5 was just as massive we went from no agentic capabilities to fully agentic, 10-100x for a 10-100x increase in compute usage.
GPT-5 vs GPT-6 was atleast a 10x (upto 100x) increase in compute, but the difference is not that clear sure people say that one is significantly better than the other, but GPT-5.5 was good enough compared to where we are today.
Even on benchmarks compared to GPT-5.2 Codex we have almost 2x'ed the performance on say artificial analysis (53 vs 29).
This is absurd. 10-100x increase in compute for 2-5x performance improvement is a massive gap.
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.
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.
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.
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.
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.
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.
> We need provide SVG code. Need creative. Ensure valid standalone. Pelican on bicycle, perhaps detailed vector. Need answer likely fenced svg. Could include title desc. Need be reasonably concise but image quality.
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.
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 V4.1 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.
The model wasn’t horrible at coding, it just made a ton of initial mistakes—but then would catch most of them later on. It can get work done if coaxed, but its peers are just cheaper and/or easier to work with.
Claude sub has a 60x subsidy, so a $20 sub gives you $1200 in API.
Based on this math, Claude sub is a much better deal than Deepseek. And it's a stronger model of course.
This is only true if you max out the plan, which requires working around awkward 5-hour and weekly limits. For side projects, sure, you can use a Ralph loop or whatever and just continue when the window resets. For work, I can't afford to simply stop working when I hit my limit. And using Claude at API prices is prohibitively expensive.
Personally, I have a $20/month OpenAI subscription that I supplement with DeepSeek V4.1 Flash (currently using CoralBricks as the provider since it's fast and I'm grandfathered into their $0 cache read pricing). It's working out rather well so far. Even when I use DS all day as my primary model, I only pay about $1/day, which is roughly equivalent to what I pay for the sub anyway (but with no limits).
If I was to do anywhere near that kind of work on Anthropic, I'd find myself running out of tokens, waiting days to continue working, and leaning on their el-crap models to try to avoid running out again. This happens all the time.
This way I get to leverage DS for nearly all my work, benefit from its far superior latencies, and I still get to lean on OpenAPI / Anthropic or whatever SOTA models I want when I need them.
The only time that really falls down is if I'm doing huge amounts of real novel work, which hasn't been the case for a while.
I use OpenCode at home, this is one of the reasons.
Their Ember-1 model (Kimi K3 tuned to think less) hallucinates to high hell, I have no trust of it, nowhere to provide feedback, despite their launch post saying they would like feedback. That they loudly and proudly claim their internal devs notice no difference is a huge red flag for me. It says there is too much eyes free coding, I noticed the hallucinations and false claims of task completion immediately.
For some reason it just creeps me out knowing that someone is recording me and reading my conversations.
ZDR is:
What you really want (assuming you can't afford your own servers and datacentre space) is secure enclave processing:See Tinfoil (US)[1] or Privatemode (Germany)[2].
[1] https://tinfoil.sh/ [2] https://www.privatemode.ai/
I'm a heavy Claude/anthropic user, I max out my 20x max plan in 1-2 days at this point. The OpenCode Go $40 plan with DeepSeek flash is like 2x the value of the Claude max 20x plan for how much I actually get out of it, for 1/5 the price.
DeepSeek literally like 10x better value than Claude/Anthropic.
DeepSeek is really cheap when you compare API pricing, but it's more expensive than a subsidized Anthropic or OpenAI subscription.
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 within 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" 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.
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.
It's never "cheaper and faster and more intelligent", it's pick 2. Sometimes (as was the case this month), it was "cheaper per token, but reasons five times as much", so your costs go up.
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.
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.
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.
Edit: You removed model names ig
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% proportional 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
Yes the official DS provider has a low cache cost, for now (it's just distributed KV). But people are still concerned about ZDR, sovereign AI, getting more model choice in their subscription price, etc. They want more bang for their buck. So when a new model comes out, the bigger impact is really its performance across providers, not the price at a single provider.
And in about 5 years (or less), API prices will stop being a thing. We all used to pay per-minute for phone calls, texts, data. Nobody does anymore. Virtually every mobile provider is now a flat prepaid or contract fee. Carriers improved their technology (going from analog to digital), and it turned out consumers will spend more money on a packaged subscription than on pay-per-use. That's why everything in the world is a subscription now.
Your modern $70 big wireless plan isn't giving you $70 of value. Similarly, your $20 AI subscription that's a great deal right now, will be less valuable in 5 years, when hardware & models are more efficient. You'll still be able to buy a local AI rig, but it will still be very expensive, due to AI models continuing to be large, and NVidia keeping new hardware expensive, as these things together reinforce their dominance and profits, and our reliance on them.
This also follows another pre-existing trend. Why do people keep having to buy a new iPhone or laptop every other year? Because each year companies make software & data less efficient/more bloated, so people need a more powerful machine to run it, which enables them to make even more bloated software. Almost everything people do with a computer today is what they did 15-20 years ago; it's just a higher bit-rate. Despite this, people keep re-buying the same machine, so companies can keep extracting more wealth.
AI works the same way. 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. This isn't just a theory, btw; the US economy's growth is now almost entirely reliant on this happening.
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+
https://agent.minimax.io/tools/m3-1-flash-preview
- DeepSeek V4 Flash: 284B / 13B active MoE
- DeepSeek V4 Pro: 1.6T / 49B active MoE
- DeepSeek V4.1 Flash: 552B / 8B input, 16B output MoE
https://jonclegg.github.io/pacman-bakeoff/#minimax-m3.1-flas...
20 t/s is about the speed you can read at, so slower than that and it always feels like it's buffering. 35 t/s is where it starts feeling decent and 50 t/s feels nice.
13 TPS is unusable for agentic coding, because if decode is that slow, then prefill is absolutely abysmal, and agentic work live and die by prefill speeds.
Why wouldn't I want the smartest possible model I can get? I only "chat" about smart person stuff.
If you have 200k context, but only 200t/s prefill, then you’ll wait multiple minutes when continuing an average session from yesterday. Or compaction - might take 15 minutes.
And no new releases from Chinese labs likely means their models that they could compare against those models are still baking largely because they lack compute so severely they need to train models for far longer, or figure out training on Chinese hardware.
Either way it's clear open models a lagging by ~1-2 months compare to frontier, depending on what frontier means to a person. Honestly most new models are just not that much better than at edge cases, they still fail a lot for the kind of stuff I do, but those edge cases matter a lot for un-interrupted long horizon work.
I personally am starting to see a wall here in llm development but that might just be me, given that people are being forced to scale models up considerably for frontier performance.
It's scary thinking that modern models will likely be at best 2-3% better in benchmarks or only useful in 1-2 unique edge cases compared to base models for 100x cost increase in compute required, it's getting absurd.
I can't imagine how Fable models can retain any value once you have to go 100x larger, while 27-32B models start to catch up to the current frontier for coding or other narrow/focused tasks in a year or so.
Lots of synthetic data, distilled data etc.
The problem is if we can pack a Fable in Qwen 4 27B who will pay subscription prices?
I fully expect that we will see increased performance from smaller architectures (to a point) simply because they are refined using larger ones.
> The problem is if we can pack a Fable in Qwen 4 27B who will pay subscription prices?
This is available to everyone developing these systems. Prices per relative unit of intelligence will continue to decrease. And market economics will continue to drive the cost down.
Ala intelligence is a function of compute, there was no guarantee that you need to 10x compute for 10% better performance or something like that.
If you ask the labs they will say the scaling is fairly linear or atleast sub-quadratic, they definitely won't say it's exponential.
No one was thinking we need to 10x compute from here again for another 10-20% increase in performance.
It's only been since Opus 4.6 that the compute growth vs improvements are starting to feel this way.
For instance GPT-3 vs GPT-4 was practically a 10-100x jump for a 10-100x increase in compute, we went from broken sentences to fully formed essays writing, speech and so on.
GPT-4 vs GPT-5 was just as massive we went from no agentic capabilities to fully agentic, 10-100x for a 10-100x increase in compute usage.
GPT-5 vs GPT-6 was atleast a 10x (upto 100x) increase in compute, but the difference is not that clear sure people say that one is significantly better than the other, but GPT-5.5 was good enough compared to where we are today.
Even on benchmarks compared to GPT-5.2 Codex we have almost 2x'ed the performance on say artificial analysis (53 vs 29). This is absurd. 10-100x increase in compute for 2-5x performance improvement is a massive gap.
[1]: https://bitterlesson.ai/ [2]: https://artificialanalysis.ai/?models=gpt-6-astra%2Cgpt-6-lu...
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.
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.
It actually hasn't, it seems: https://news.ycombinator.com/item?id=50151775
Obviously your marketing team is going to use the Pareto frontier where your model has improved the most and compares favorably against competitors.
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
H3, on the other hand, is phenomenal
2. Add "Flash" to the end
3. ?
4. Profit
Unless they come in at half the price (which I think will be impossible) on inference ... they have wasted their training compute imho.
Reasoning traces use truncated English:
> We need provide SVG code. Need creative. Ensure valid standalone. Pelican on bicycle, perhaps detailed vector. Need answer likely fenced svg. Could include title desc. Need be reasonably concise but image quality.
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)
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.
Then maybe they can be merged once one gains a comment?