Step 5 Preview: Specs, Pricing, Benchmarks and Polish Test

👁 118 reads

Updated regularly. Information current as of October 19, 2026, 11:30 a.m.

Step 5 Preview is the new flagship model from Shanghai-based StepFun: a 600-billion-parameter MoE architecture, with 27 billion parameters active per token, a 1-million-token context window, and support for text, images and video input. It scores 44 on the Artificial Analysis Intelligence Index, matching Kimi K3, and costs $1 per million input tokens and $2.70 per million output tokens. It has been available on OpenRouter since October 8, and StepFun released its weights on October 15. In my test, which we ran in Polish, it scored 65.4 points - roughly on a par with GLM-5.3.

What is Step 5 Preview?

StepFun (Chinese name: 阶跃星辰) is a Shanghai-based startup founded in April 2023 by Jiang Daxin, formerly a Microsoft vice president. In China, it is counted among the “six AI tigers,” alongside Moonshot (Kimi), Zhipu (GLM) and MiniMax. Until now, I knew the company mainly for its low-cost, open Step 3.5 Flash and Step 3.7 Flash models, released under the Apache 2.0 license. Step 5 Preview is the company’s first model aimed at the very top tier - and it skips version 4 altogether.

The developer describes it as a model “for agentic work”: coding, multistep tasks involving tools, and working with documents, with a particular focus on finance. In practice, reasoning is mandatory: the model always “thinks” before answering. Here are the key specs from StepFun and OpenRouter documentation:

• Architecture: Mixture of Experts, 600 billion parameters in total, with 27 billion active per token.
• Context: 1 million input tokens and up to 64,000 tokens per response.
• Input: Text, images (JPG, PNG and WebP, up to 60 per request) and video (MP4, MOV and MKV, up to 128 MB; under 5 minutes is recommended). Text output only.
• Reasoning effort: Low, medium (the default) and high. You cannot turn reasoning off.
• API features: Streaming, tool calls, JSON and JSON Schema modes, and prompt caching.

How to get started

OpenRouter is a straightforward option: one API key gives you access to hundreds of models, with billing in US dollars by card. See my OpenRouter guide for how to create an account and API key.

1. Sign in at openrouter.ai and add funds to your account (the minimum is a few dollars).
2. Visit the stepfun/step-5-preview model page and click “Chat” if you want to try it in your browser first.
3. For API use, set the model ID to stepfun/step-5-preview and use reasoning: {"effort": "low"} if you want to reduce time and cost.
4. Set max_tokens well above the response length you expect - at least 12,000 to 16,000. Reasoning uses up the limit, and blank responses were the most common problem in my test.
5. Alternatively, sign up for StepFun’s international platform at platform.stepfun.ai, which bills in US dollars. Its Chinese platform, platform.stepfun.com, bills in yuan.

Here is an example API call I used in the test (Python, using the requests library). The prompt below is a translation of the Polish test prompt, not a new English-language test:

import os, requests
r = requests.post("https://openrouter.ai/api/v1/chat/completions",
    headers={"Authorization": "Bearer " + os.environ["OPENROUTER_API_KEY"]},
    json={"model": "stepfun/step-5-preview",
          "max_tokens": 16000,
          "reasoning": {"effort": "low"},
          "messages": [{"role": "user", "content": "Translated version of the Polish test prompt: Fix the errors in the text: ..."}]})
print(r.json()["choices"][0]["message"]["content"])

How much does Step 5 cost?

Prices are listed in US dollars. Reasoning tokens are billed as output, so the actual cost of a task depends largely on how much the model “thinks.”

That is cheap compared with many rivals, but not the cheapest. Artificial Analysis estimates that an average task in its index costs $1.03 with Step 5: the model is verbose, generating 160 million tokens in its evaluation compared with a median of 94 million. My 13 lab tasks cost $0.31 in total, including retries for blank responses. For comparison, GLM-5.3 completed the same set for $0.03, while Claude Opus 5.5 cost $0.37. See the AI model pricing ranking for the full price table.

Step 5 benchmark results

The key independent result is its score of 44 on the Artificial Analysis Intelligence Index, ranking 40th out of 226 models. Artificial Analysis currently classifies Step 5 as a closed model because its weights had not yet been released. If it were counted among open models, it would be tied for third, behind MiMo-V2.6-Pro (46) and GLM-5.3 (45), and level with Kimi K3 (44). That is the basis for the “top 3 open models” claim repeated in Chinese media.

I found the developer’s results in tables republished by media outlets (eesel.ai, BenchLM and MarkTechPost), since the model page on stepfun.com cannot be read without a browser. Treat these figures as claims from the developer:

The overall picture is consistent: Step 5 edges out Kimi K3 and GLM-5.3 in coding and finance, but trails GLM in terminal work. It is a few points behind leading Western models such as Opus 5 and GPT-6 Astra. Artificial Analysis reports a speed of 86.8 tokens per second and 2.8 seconds to the first token. In my test, the longest response (6,975 tokens) took 79 seconds, or about 88 tokens per second.

Step 5 in Polish: my test results

On October 9, I ran Step 5 Preview through the same 13 tasks used for the Prompt Lab ranking and the Polish-language model arena: nine writing tasks (an article, product description, email, social post, story, headlines, summary, plain-language rewrite and translation), two proofreading tasks and two quizzes about Polish topics. The test was conducted in Polish. I used low reasoning effort and the same token limits as for the other models. Two judges, Claude Opus 5.5 and GPT-6 Sol, rated the texts blind; I calibrated the results against four anchor models.

Key takeaways:

• Blank responses. For the product description and both proofreading tasks, the model used its entire limit (6,000-8,000 tokens) on reasoning and returned no text, even after three attempts. This was not a Polish-language error; it was the model’s behavior at “low” effort, and the main reason for its low score.
• A higher limit gets a response.I repeated the easy proofreading task once with a 16,000-token limit. The model used 4,902 tokens for reasoning, returned text and corrected 19 of 22 errors. It missed the errors in the original Polish test: [corrupted lettering in the original Polish test], [corrupted lettering in the original Polish test], and [corrupted lettering in the original Polish test].
• It knows facts about Poland. It scored 46/46 on the easy quiz. On the harder quiz, it answered 11 questions before reaching its limit; all 11 answers were correct.
• Short writing is stronger than long writing. The judges gave the apology email 8.5 and the summary and translation 8.0. The story scored 5.0 for phrases equivalent to “the sun from twenty years ago,” while the plain-language text scored 5.3 for wording equivalent to “savings may become less able to buy things.”
• En dashes. The model used 14.8 en dashes per 1,000 words, including in an email subject line. If you publish AI-generated text, you may need to replace them with hyphens.

For the full task descriptions, response excerpts and judges’ ratings, see Step 5 in Polish: results from 13 tasks. The findings are my assessment of a single run.

Open weights released on October 15

StepFun met its deadline: on October 15, 2026, Step 5 Preview weights became available on Hugging Face and ModelScope in FP8 (stepfun-ai/Step-5-Preview, about 650 GB of files) and BF16 formats. Nvidia also released an NVFP4 version. The StepFun Model License 1.0 is free and perpetual, permits commercial use, modifications and derivative models, and requires the license to be included when redistributing the model. StepFun’s hardware recommendations:

Step 5 will not run on a typical home graphics card, even an RTX 4090 with 24 GB of memory. I cover the license and my second Polish-language test from October 19 in Step 5 has open weights. For comparisons with other open models from that week, see the open AI models hub. See local AI for models you can run on your own computer.

Step 5 vs. Kimi K3, GLM-5.3, Qwen and DeepSeek

Prices and AA scores are from the Artificial Analysis model pages as of October 9. The Kimi K3 and DeepSeek results come from the first round of the lab (the same tasks, but a different group of judges), so I treat differences of up to 2-3 points as a tie. See Step 5 vs. Kimi K3, GLM-5.3, Qwen and DeepSeek for details and conclusions.

Who should use Step 5?

• Developers and agent teams looking for low per-token costs and coding results close to Kimi K3. With a 1-million-token context window and caching at $0.05, it could be a low-cost engine for analyzing repositories.
• Businesses that want to host a model themselves - now that the weights are available, provided the license proves as permissive as the one for Step 3.7 Flash.
• Not for Polish-language writing. For text, emails and proofreading, GPT-6 Luna (80.4 points for a fraction of a cent) or Kimi K3 is a better choice. Step 5 scored about the same as GLM-5.3 and also carries a risk of blank responses.

Frequently asked questions

Is Step 5 free?

Not through the API: OpenRouter did not offer a free version as of October 9. Chinese media have reported free access through StepFun apps, but I could not verify that from my location. The weights became available on October 15, so you can run the model yourself if you have a server.

Is Step 5 available in Poland?

As of October 9, I was able to access it through OpenRouter and StepFun’s international platform, platform.stepfun.ai; both bill in US dollars. I did not check whether the consumer StepFun app works with a Polish account. Availability may vary by country, so check with the provider.

Can Step 5 speak Polish?

It understands Polish and writes well in short formats, but scored 65.4 in my Polish-language test - below Kimi K3, Qwen3.8 Max and DeepSeek V4.1 Flash, and roughly level with GLM-5.3.

How much memory do you need to run Step 5 locally?

The weights alone require about 1.2 TB in BF16, about 600 GB in FP8, or about 300-360 GB at 4-bit precision. This is server-class hardware, not a home computer with a single graphics card.

Why does Step 5 return a blank response?

Reasoning is mandatory and counts toward the max_tokens limit. If the limit is too low, the model may use its entire budget thinking and have no room left to answer. Set the limit to 12,000-16,000 tokens or more.

How is Step 5 different from Step 3.7 Flash?

Step 3.7 Flash is a smaller model (about 200 billion parameters, 11 billion active, with a 256,000-token context window) priced at $0.20 / $1.15. Step 5 is three times larger, has a 1-million-token context window, accepts video and costs about five times more per input token.

When will the full Step 5 arrive?

StepFun has not announced a date. “Preview” means the specs and prices could still change.

More about Step 5

• Step 5 Preview launches on OpenRouter - news
• Step 5 in Polish: results from 13 tasks
• Step 5 vs. Kimi K3, GLM-5.3, Qwen and DeepSeek
• Lesser-known AI models: an overview

// read moreStepFun Step 5 Preview Hits OpenRouter: Kimi K3 Rival at $1NewsStep 5 vs Kimi K3, GLM-5.3, Qwen and DeepSeek ComparedBenchmarksLing 3.1 Flash: Ant Group’s 560B Model Specs, Pricing, and TestBenchmarks

Kinetyra