Step 5 Preview is live on Featherless

StepFun released the open weights for Step 5 Preview today: 600B total parameters with 27B active per token, built for long-horizon agent work, at about 1/6 the cost per task of Claude Opus 5.5. It is live on Featherless now at 256K context in FP8.

What’s new

Step 5 Preview is StepFun’s new flagship, and the design choice that stands out is depth. It stacks 92 transformer blocks, as deep as Qwen3.8-Max and Kimi K3, while using less than a quarter of their total parameters and less than a third of their active ones. StepFun calls this “narrow but deep”: agents spend most of their time reading long tool outputs rather than writing long chains of thought, so the model needs enough layers to reason implicitly over a big prefill.

Long context is handled by Sparse GQA. The full-attention layers pick which parts of the history to read, and StepFun groups the index into 8-token blocks, which cuts the cost of indexing and top-k selection by 8x at million-token lengths. The native window is 1M tokens, and it takes text and images as input.

StepFun points it at three kinds of work:

• Coding agents. Fixing bugs in unfamiliar codebases, building features across files, and debugging until the changes run and pass checks.
• Frontend from a design. Turning a screenshot, mockup or PRD into a working web page, or building a playable browser game from a folder of assets.
• Finance work. Reading annual reports and operating data, breaking down revenue, profit, cash flow and risk, and handing back a report, spreadsheet or deck.

The full architecture and training write-up is in StepFun’s Inside Step 5 Preview.

The numbers

Artificial Analysis has already measured it. Step 5 Preview scores 44 on the Intelligence Index, and it costs $1.03 per task to run their full evaluation. Claude Opus 5.5 at its max setting scores 58, at $5.98 per task. That is the trade: not the top of the leaderboard, but frontier-level agent work at about a sixth of the price.

What we built with it

We had early access, so we put it to work.

Feather Pop. We used Step 5 Preview to port an entire game codebase to TypeScript, and it runs in your browser. Play it here.

Sugar Data. We asked Step 5 Preview to write M&A investment theses for 100+ AI and tech companies: who each one should buy, who could buy it, and why. Each memo cost around 6 cents. Browse the deals, or read how we tested it against other open models.

How to run Step 5 Preview

Currently on Featherless, Step 5 Preview serves at 256K context with FP8 quantization, with vision enabled. The model’s native window is 1M tokens, and we’re working to expand ours toward it.

It runs on the same OpenAI-compatible API that developers already use, the same one that serves the other 40,000+ open models we host.

from openai import OpenAI

client = OpenAI(
    base_url="https://api.featherless.ai/v1",
    api_key="YOUR_FEATHERLESS_KEY",
)

response = client.chat.completions.create(
    model="stepfun-ai/Step-5-Preview",
    messages=[
        {
            "role": "user",
            "content": "We're acquiring a company. We earn $800M of net income on 400M shares trading at $50. The target earns $90M and we're paying $2.4B: half new debt at 7%, half new shares issued at $50. Tax rate is 25%, and we expect $40M of pre-tax synergies in year one. Is the deal accretive or dilutive to EPS, and how much in synergies do we need to break even?",
        }
    ],
)

print(response.choices[0].message.content)

If you already run Step 5 Preview in Cline, Kilo Code, OpenCode or Hermes Agent, switching is three settings: your Featherless key, the base URL above, and the model ID stepfun-ai/Step-5-Preview.

Per-token rates are $1 per 1M input tokens, $0.05 per 1M cached input, and $2.70 per 1M output.

And, as is the same with all of your api calls on featherless, no logs, so your prompts and completions stay yours.

Try it out today – https://featherless.ai/models/stepfun-ai/Step-5-Preview

Going beyond

Are you interested in running Step 5 Preview at even larger scales, or at its full 1M context? Are your unpredictable per-token bills exploding? Check out our dedicated gpu clusters, where our team can help tune the full inference stack exactly for your needs, including quantization, batching, fine-tuning, and whatever else you need. We have GPUs, and you need compute. Learn more below, and feel free to schedule a call, where we’ll help tackle your biggest AI problems.

Get Dedicated GPU at $7.6k/month → https://featherless.ai/lp/dedicatedgpu