01What is Step 5 Preview and how does it work?

Step 5 Preview is a new Chinese AI model from Step Fun AI, and it is free for one week across multiple platforms. It is a mixture of experts (MoE) model with 600 billion total parameters and 27 billion active, so it only uses a small portion of its brain for each task. This makes it fast and efficient.

It has a 1 million token context window, meaning it can handle very long inputs, and it also has vision, so it can understand images as well as text. Built by Step Fun AI, it performs similarly to GP6 Luna and Gemini 3.8 Flash on benchmarks. It is not the most frontier model—it won't replace Claude Opus—but it is great for subagents and background automations.

In this episode, Julian Goldie shows how you can use Step 5 Preview for free inside tools like Hermes Agent, a free AI coder, and other platforms. You can build with it, and it integrates into an agentic operating system, allowing you to automate tasks without paying for premium models.

02Key features of Step 5 Preview

  • Free for one week across OpenCode, Klein, Nou's Research, KiloCode, and more.
  • 600 billion total parameters with 27 billion active (mixture of experts).
  • 1 million token context window for long inputs.
  • Vision capabilities to process images.
  • Fast output at 100 tokens per second, faster than some models like Astra 6 or Opus 8.
  • Cheaper than many competitors, costing $0.72 per intelligence task via API (after free week).
  • Open source future – planned release on Hugging Face.

What's new? Step Fun AI skipped Step 4 and went straight from Step 3.7 to Step 5, which is unusual. The model is already integrated into Julian's agentic operating system, Hermes, and is available in a free AI coder.

You can use it alongside other models like Hermes Apollo, Hermes Oracle, and Hermes music.

03How to get started with Step 5 Preview

  1. Access Step 5 Preview on platforms like OpenCode, Klein, Nou's Research, or KiloCode – it is free for one week.
  2. Go to the Hermes Agent interface and select Step 5 as your model.
  3. Alternatively, open the free AI coder and choose Step 5 from the model list.
  4. If you are outside the free week, use the API – it costs $0.72 per intelligence index task.
  5. Connect it to your agentic operating system, like Hermes, to use it for automations.
  6. You can also use it via open router or other providers, as it is available on Nou's Research.

Once set up, you can start chatting or building. In the episode, Julian demonstrates asking it to build something in the chat. Because it is free for a week, you can test it without any cost.

04How to use Step 5 Preview in your workflows

Step 5 Preview is best used as a subagent model for quick answers and background tasks. In the episode, Julian shows it plugged into Hermes Agent, where it can provide fast responses without affecting your local setup, since Hermes Cloud runs on a VPS. You can use it side by side with other models like Hermes Apollo for different workflows.

For example, you can ask it to build something in the free AI coder. It will generate code or complete tasks. Because of its 1 million token context, it can handle large projects.

You can also integrate it into your agentic operating system to automate processes, like sending emails or scraping data.

Use it as a free alternative for non-critical tasks where speed matters. It is not for frontier-level reasoning, but for everyday automations, it works well. In the episode, it is used in a chat to build things and is part of a memory system inside the Agent OS.

05Benchmarks and cost comparison

In the episode, Julian compares Step 5 Preview to other models using indices like Hermes Index and Artificial Analysis. On Hermes Index, it scores 33.89, similar to GP6 Luna. On Artificial Analysis, it sits in the middle, outperforming Gemini 3.8 Flash on intelligence and cost.

Cost-wise, Step 5 Preview is 42% less expensive than Gemini 3.8 Flash. It costs $0.72 per intelligence index task, while Gemini 3.8 Flash costs $1.24. Despite being cheaper, Step 5 Preview outperforms Gemini 3.8 Flash and also beats Kimi K3 Max and DeepSeek v4.1.

However, it is slower than Gemini 3.8 Flash. Gemini outputs 331 tokens per second, while Step 5 does 100. But it is faster than some other models like Astra 6 or Fable 5.

Also, it is verbose, producing more tokens than typical models, which offsets some cost savings but still remains cheaper.

06Tips, limitations, and mistakes to avoid

  • Don't expect frontier performance – it won't replace Claude Opus or GP6 Astra for complex reasoning.
  • Watch for verbosity – it produces long outputs, so you may need to manage token usage.
  • Use it for speed – it is fast for a free model, but not as fast as Gemini 3.8 Flash.
  • Take advantage of the free week – after that, you pay $0.72 per task, which is still cheap.
  • Consider open source – if you have resources, you can run it locally once released.

A common mistake is thinking it can do everything. It is best for subagents and background tasks, not for final, high-stakes decisions. Also, because it is verbose, you may use more tokens than expected, but the lower cost per token makes up for it.

07Who should use Step 5 Preview and why it matters

Step 5 Preview is ideal for developers and AI enthusiasts who need a free, fast model for automations and subagents. It is especially useful for those who want to experiment without spending money. In the episode, Julian highlights how it is already integrated into his AI Profit Boarding community, where members can use it in the agentic operating system.

Why it matters: it shows that Chinese AI labs like Step Fun AI are becoming competitive with Western models like Gemini 3.8 Flash, at a lower cost. This could lead to more affordable AI tools for everyone. Since it is open source, it also allows for transparency and customization.

If you are building automations on a budget, Step 5 Preview is a great option. It is not for everyone, but for those who need quick, reliable, and cheap AI, it is worth trying during the free week.