I'm not aware of other AI labs that released base checkpoint for models in this size class. Qwen released some base models for 3.5, but the biggest one is the 35B checkpoint.
thanks for the info. before running the bench i only tried it in arena.ai type of tasks and it was not impressive. i didn't expect it to be that good at agentic tasks
I will try and add it. But I doubt it works well because Mimo V2 Pro is beaten by stepfun even at performance leaderboard (price is not a factor in this leaderboard), so I expect MiMo V2 Flash to perform even worse.
According to openrouter.ai it looks like StepFun 3.5 Flash is the most popular model at 3.5T tokens, vs GLM 5 Turbo at 2.5T tokens. Claude Sonnet is in 5th place with 1.05T tokens. Which isn't super suprising as StepFun is ~about 5% the price of Sonnet.
the real surprising part to me is that, despite being the cheapest model on board, stepfun is often able to score high at pure performance. Other models at the same price range (e.g. kimi) fails to do that.
Tried the free version on OpenRouter with pi.dev and it's competent at tool calling and creative writing is "good enough" for me (more "natural Claude-level" and not robotic GPT-slop level) but it makes some grave mistakes (had some Hanzi in the output once and typos in words) so it may be good with "simple" agentic workflows but it's definitely not made for programming nor made for long writing.
another thing from the bench I didn't expect: gemini 3.1 pro is very unreliable at using skills. sometimes it just reads the skill and decide to do nothing, while opus/sonnet 4.6 and gpt 5.4 never have this issue.
I ran 300+ benchmarks across 15 models in OpenClaw and published two separate leaderboards: performance and cost-effectiveness.
The two boards look nothing alike. Top 3 performance: Claude Opus 4.6, GPT-5.4, Claude Sonnet 4.6. Top 3 cost-effectiveness: StepFun 3.5 Flash, Grok 4.1 Fast, MiniMax M2.7.
The most dramatic split: Claude Opus 4.6 is #1 on performance but #14 on cost-effectiveness. StepFun 3.5 Flash is #1 cost-effectiveness, #5 performance.
Other surprises: GLM-5 Turbo, Xiaomi MiMo v2 Pro, and MiniMax M2.7 all outrank Gemini 3.1 Pro on performance.
Rankings use relative ordering only (not raw scores) fed into a grouped Plackett-Luce model with bootstrap CIs. Same principle as Chatbot Arena — absolute scores are noisy, but "A beat B" is reliable. Full methodology: https://app.uniclaw.ai/arena/leaderboard/methodology?via=hn
I built this as part of OpenClaw Arena — submit any task, pick 2-5 models, a judge agent evaluates in a fresh VM. Public benchmarks are free.
Cheapest just isn't a very useful metric. Can I suggest a Pareto-curve type representation? Cost / request vs ELO would be useful and you have all the data.
TBH that was my initial thought too, but I found some problem using this approach:
Essentially I'm using the relative rank in each battle to fit a latent strength for each model, and then use a nonlinear function to map the latent strength to Elo just for human readability. The map function is actually arbitrary as long as it's a monotonically increasing function so it preserves the rank. The only reliable result (that is invariant to the choice of the function) is the relative rank of models.
That being said, if I use score/cost as metrics, the rank completely depends on the function I choose, like I can choose a more super-linear function to make high performance model rank higher in score/cost board, or use a more sub-linear function to make low performance model rank higher.
That's why I eventually tried another (the current) approach: let judge give relative rank of models just by looking at cost-effectiveness (consider both performance and cost), and compute the cost-effectiveness leaderboard directly, so the score mapping function does not affect the leaderboard at all.
both are shown in battle detail page already. Time is shown in Scores table. Number of tokens are shown in Cost details at the bottom of the Scores. (I thought most people just want to see cost in USD so I put token details at the bottom)
I would have liked aggregated results instead. Expanding 300 tables is a bit tiresome. But I guess that is easy with AI now. Here is a scatter plot of quality vs duration
But I just noticed that this ranking is meaningless because it conflates model quality with provider uptime.
Claude Haiku has a higher average quality than Claude Opus, which does not make sense. The explanation is that network errors were credited with a quality score of 0, and there were _a lot_ of network errors.
> The explanation is that network errors were credited with a quality score of 0, and there were _a lot_ of network errors.
all network error, provider error, openclaw error are excluded from ranking calculation actually, so that is not the reason.
Real reason:
The absolute score is not consistent across tasks and cannot be directly added/averaged, for both human and LLM. But the relative rank is stable (model A is better than B). That is exactly why Chatbot Arena only uses the relative rank of models in each battle in the first place, and why we follow that approach.
a concrete example of why score across tasks cannot be added/averaged directly: people tend to try haiku with easier task and compare with T2 models, and try opus with harder task and compare with better models.
another example: judge (human or llm) tend to change score based on opponents, like Sonnet might get 10/10 if all other opponents are Haiku level, but might get 8/10 if opponent has Opus/gpt-5.4.
So if you want to make the plot, you should plot the elo score (in leaderboard) vs average cost per task. But note: the average cost has similar issue, people use smaller model to run simpler task naturally, so smaller model's lower cost comes from two factor: lower unit cost, and simpler task.
methodology page contains more details if you are interested.
some kind of top-level metric like avg tokens/task would be useful. e.g. yes stepfun is 5% the price of sonnet, but does it use 1x, 10x or 1000x more tokens to accomplish similar tasks/median per task. for example I am willing to eat a 20% quality dive from sonnet if the token use is < 10% more than sonnet. if token use is 1000x then that's something I want to know.
It’s too late once it’s happened. I was curious, then when I saw the site looked vibecoded and you’re commenting with AI, I decided to stop trying to reason through the discrepancies between what was claimed and what’s on the site (ex. 300 battles vs. only a handful in site data).
> and it doesn't disminish the quality of the work.
It does affect incoming people hearing about the work.
I applaud your instinct to defend someone who put in effort. It's one of the most important things we can do.
Another important thing we can do for them is be honest about our own reactions. It's not sunshine and rainbows on its face, but, it is generous. Mostly because A) it takes time B) other people might see red and harangue you for it.
all 300+ battle data are available at https://app.uniclaw.ai/arena/battles, every single battle is shown with raw conversional history, produced files, judge's verdict and final scores
Yes, judge is one of opus 4.6, gpt 5.4, gemini 3.1 pro (submitter can choose). Self judge (judge model is also one of the participants) is excluded when computing ranking.
> There's lot of references to "just like LMArena", but LMArena is human evaluated?
Yeah LMArena is human evaluated, but here i found it not practical to gather enough human evaluation data because the effort it take to compare the result is much higher:
- for code, judge needs to read through it to check code quality, and actually run it to see the output
- when producing a webpage or a document, judge needs to check the content and layout visually
- when anything goes wrong, judge needs to read the execution log to see whether partial credit shall be granted
if you look at the cost details of each battle (available at the bottom of battle detail page), judge typically cost more than any participant model.
if we evaluate with human, i would say each evaluation can easily take ~5-10 min
39 comments:
StepFun is an interesting model.
If you haven’t heard of it yet there’s some good discussion here: https://news.ycombinator.com/item?id=47069179
Since that discussion, they released the base model and a midtrain checkpoint:
- https://huggingface.co/stepfun-ai/Step-3.5-Flash-Base
- https://huggingface.co/stepfun-ai/Step-3.5-Flash-Base-Midtra...
I'm not aware of other AI labs that released base checkpoint for models in this size class. Qwen released some base models for 3.5, but the biggest one is the 35B checkpoint.
They also released the entire training pipeline:
- https://huggingface.co/datasets/stepfun-ai/Step-3.5-Flash-SF...
- https://github.com/stepfun-ai/SteptronOss
thanks for the info. before running the bench i only tried it in arena.ai type of tasks and it was not impressive. i didn't expect it to be that good at agentic tasks
why do half the comments here read like ai trying to boost some sort of scam?
Missing from the comparison is MiMo V2 Flash (not Pro), which I think could put up a good fight against Step 3.5 Flash.
Pricing is essentially the same: MiMo V2 Flash: $0.09/M input, $0.29/M output Step 3.5 Flash: $0.10/M input, $0.30/M output
MiMo has 41 vs 38 for Step on the Artificial Analysis Intelligence Index, but it's 49 vs 52 for Step on their Agentic Index.
I will try and add it. But I doubt it works well because Mimo V2 Pro is beaten by stepfun even at performance leaderboard (price is not a factor in this leaderboard), so I expect MiMo V2 Flash to perform even worse.
Yet when I tried it it did absymal compared to Gemini 2.5 Flash
what kind of tasks did you try?
According to openrouter.ai it looks like StepFun 3.5 Flash is the most popular model at 3.5T tokens, vs GLM 5 Turbo at 2.5T tokens. Claude Sonnet is in 5th place with 1.05T tokens. Which isn't super suprising as StepFun is ~about 5% the price of Sonnet.
https://openrouter.ai/apps?url=https%3A%2F%2Fopenclaw.ai%2F
> the most popular model
It was free for a long time. That usually skews the statistics. It was the same with grok-code-fast1.
Exactly. When I read the headline I thought: "Ofc it is, its free."
I should have clarified I didn't use the free version...
the real surprising part to me is that, despite being the cheapest model on board, stepfun is often able to score high at pure performance. Other models at the same price range (e.g. kimi) fails to do that.
It looks like Unsloth had trouble generating their dynamic quantized versions of this model, deleted the broken files, then never published an update.
people aren't just using Claude models any more? that's nice to see
i like StepFun 3.5 Flash, a good tradeoff
Tried the free version on OpenRouter with pi.dev and it's competent at tool calling and creative writing is "good enough" for me (more "natural Claude-level" and not robotic GPT-slop level) but it makes some grave mistakes (had some Hanzi in the output once and typos in words) so it may be good with "simple" agentic workflows but it's definitely not made for programming nor made for long writing.
it's actually pretty good at openclaw type of tasks for non technical users: lots of tool calls, some simple programing
Yeah this kind of stuff. I have no experience with OpenClaw though.
another thing from the bench I didn't expect: gemini 3.1 pro is very unreliable at using skills. sometimes it just reads the skill and decide to do nothing, while opus/sonnet 4.6 and gpt 5.4 never have this issue.
I ran 300+ benchmarks across 15 models in OpenClaw and published two separate leaderboards: performance and cost-effectiveness.
The two boards look nothing alike. Top 3 performance: Claude Opus 4.6, GPT-5.4, Claude Sonnet 4.6. Top 3 cost-effectiveness: StepFun 3.5 Flash, Grok 4.1 Fast, MiniMax M2.7.
The most dramatic split: Claude Opus 4.6 is #1 on performance but #14 on cost-effectiveness. StepFun 3.5 Flash is #1 cost-effectiveness, #5 performance.
Other surprises: GLM-5 Turbo, Xiaomi MiMo v2 Pro, and MiniMax M2.7 all outrank Gemini 3.1 Pro on performance.
Rankings use relative ordering only (not raw scores) fed into a grouped Plackett-Luce model with bootstrap CIs. Same principle as Chatbot Arena — absolute scores are noisy, but "A beat B" is reliable. Full methodology: https://app.uniclaw.ai/arena/leaderboard/methodology?via=hn
I built this as part of OpenClaw Arena — submit any task, pick 2-5 models, a judge agent evaluates in a fresh VM. Public benchmarks are free.
Cheapest just isn't a very useful metric. Can I suggest a Pareto-curve type representation? Cost / request vs ELO would be useful and you have all the data.
TBH that was my initial thought too, but I found some problem using this approach:
Essentially I'm using the relative rank in each battle to fit a latent strength for each model, and then use a nonlinear function to map the latent strength to Elo just for human readability. The map function is actually arbitrary as long as it's a monotonically increasing function so it preserves the rank. The only reliable result (that is invariant to the choice of the function) is the relative rank of models.
That being said, if I use score/cost as metrics, the rank completely depends on the function I choose, like I can choose a more super-linear function to make high performance model rank higher in score/cost board, or use a more sub-linear function to make low performance model rank higher.
That's why I eventually tried another (the current) approach: let judge give relative rank of models just by looking at cost-effectiveness (consider both performance and cost), and compute the cost-effectiveness leaderboard directly, so the score mapping function does not affect the leaderboard at all.
Could you add a column for time or number of tokens? Some models take forever because of their excessive reasoning chains.
both are shown in battle detail page already. Time is shown in Scores table. Number of tokens are shown in Cost details at the bottom of the Scores. (I thought most people just want to see cost in USD so I put token details at the bottom)
I would have liked aggregated results instead. Expanding 300 tables is a bit tiresome. But I guess that is easy with AI now. Here is a scatter plot of quality vs duration
https://i.imgur.com/wFVSpS5.png
and quality vs cost
https://i.imgur.com/fqM4edw.png
But I just noticed that this ranking is meaningless because it conflates model quality with provider uptime.
Claude Haiku has a higher average quality than Claude Opus, which does not make sense. The explanation is that network errors were credited with a quality score of 0, and there were _a lot_ of network errors.
> The explanation is that network errors were credited with a quality score of 0, and there were _a lot_ of network errors.
all network error, provider error, openclaw error are excluded from ranking calculation actually, so that is not the reason.
Real reason:
The absolute score is not consistent across tasks and cannot be directly added/averaged, for both human and LLM. But the relative rank is stable (model A is better than B). That is exactly why Chatbot Arena only uses the relative rank of models in each battle in the first place, and why we follow that approach.
a concrete example of why score across tasks cannot be added/averaged directly: people tend to try haiku with easier task and compare with T2 models, and try opus with harder task and compare with better models.
another example: judge (human or llm) tend to change score based on opponents, like Sonnet might get 10/10 if all other opponents are Haiku level, but might get 8/10 if opponent has Opus/gpt-5.4.
So if you want to make the plot, you should plot the elo score (in leaderboard) vs average cost per task. But note: the average cost has similar issue, people use smaller model to run simpler task naturally, so smaller model's lower cost comes from two factor: lower unit cost, and simpler task.
methodology page contains more details if you are interested.
some kind of top-level metric like avg tokens/task would be useful. e.g. yes stepfun is 5% the price of sonnet, but does it use 1x, 10x or 1000x more tokens to accomplish similar tasks/median per task. for example I am willing to eat a 20% quality dive from sonnet if the token use is < 10% more than sonnet. if token use is 1000x then that's something I want to know.
Please don’t use AI to write comments, it cuts against HN guidelines.
sorry didn't know that. Here is my hand writing tldr:
gemini is very unreliable at using skills, often just read skills and decide to do nothing.
stepfun leads cost-effectiveness leaderboard.
ranking really depends on tasks, better try your own task.
It’s too late once it’s happened. I was curious, then when I saw the site looked vibecoded and you’re commenting with AI, I decided to stop trying to reason through the discrepancies between what was claimed and what’s on the site (ex. 300 battles vs. only a handful in site data).
Too late for what? For you? maybe. There are many others that are okay with it and it doesn't disminish the quality of the work. Props to the author.
> Too late for what? For you? maybe.
Maybe? :)
> There are many others that are okay with it
Correct.
> and it doesn't disminish the quality of the work.
It does affect incoming people hearing about the work.
I applaud your instinct to defend someone who put in effort. It's one of the most important things we can do.
Another important thing we can do for them is be honest about our own reactions. It's not sunshine and rainbows on its face, but, it is generous. Mostly because A) it takes time B) other people might see red and harangue you for it.
all 300+ battle data are available at https://app.uniclaw.ai/arena/battles, every single battle is shown with raw conversional history, produced files, judge's verdict and final scores
Thanks! Is the judge an LLM? There's lot of references to "just like LMArena", but LMArena is human evaluated?
> Is the judge an LLM?
Yes, judge is one of opus 4.6, gpt 5.4, gemini 3.1 pro (submitter can choose). Self judge (judge model is also one of the participants) is excluded when computing ranking.
> There's lot of references to "just like LMArena", but LMArena is human evaluated?
Yeah LMArena is human evaluated, but here i found it not practical to gather enough human evaluation data because the effort it take to compare the result is much higher:
- for code, judge needs to read through it to check code quality, and actually run it to see the output
- when producing a webpage or a document, judge needs to check the content and layout visually
- when anything goes wrong, judge needs to read the execution log to see whether partial credit shall be granted
if you look at the cost details of each battle (available at the bottom of battle detail page), judge typically cost more than any participant model.
if we evaluate with human, i would say each evaluation can easily take ~5-10 min
Fair enough, yeah, agent evals are hard especially across N models :/
Thanks for replying btw, didn't mean any disrespect, good on you for not getting aggro about feedback
I appreciate honest feedback, best way to learn :)
>Other surprises: GLM-5 Turbo, Xiaomi MiMo v2 Pro, and MiniMax M2.7 all outrank Gemini 3.1 Pro on performance
This has also been my subjective experience But has also been objective in terms of cost.