Stepfun Step-2 API 2026: Multimodal LLM & Pricing

Stepfun (阶跃星辰) is a Shanghai AI lab whose Step-2 model is a 1.2-trillion-parameter multimodal flagship accepting text, image, audio and video on a single OpenAI-compatible endpoint. ⚠️ Unlike most China LLM stacks that ship separate vision APIs and bill per modality, Step-2 treats every input as a content part — a video understanding call is priced the same as a text chat. This review covers models, August 2026 pricing, API integration, Step-R reasoning, regional access, competitors and practical limits.

🧠 Model list

💰 Pricing (verified 2026-08-05)

Token rates are in CNY per million tokens. Stepfun is pay-as-you-go from the first production request after the 30-day free window expires.

🔌 OpenAI Python SDK

from openai import OpenAI
client = OpenAI(
    base_url="https://platform.stepfun.com/v1",
    api_key="YOUR_STEPFUN_API_KEY"
)
response = client.chat.completions.create(
    model="step-2",
    messages=[
        {"role": "user", "content": [
            {"type": "text", "text": "Summarize this clip."},
            {"type": "video_url", "video_url": {"url": "https://example.com/clip.mp4"}}
        ]}
    ]
)
print(response.choices[0].message.content)

The same base_url works with any OpenAI-compatible framework adapter. Stepfun also ships a first-party Python SDK with multimodal helpers at github.com/stepfun-ai for callers that need streaming token-usage reports, image-aware caching or the Step-CC completions endpoint.

⚡ Step-R reasoning at ¥8/M

Step-R is Stepfun's answer to OpenAI o1 and DeepSeek R1, with chain-of-thought exposed through the standard streaming interface. The key number: ¥8 input / ¥24 output per million tokens. OpenAI o1 charges $15 / $60 and DeepSeek R1 sits at ¥4 / ¥16 — Step-R is roughly 1/5 of o1 and 2x of R1, but with notably stronger multimodal grounding on Step-2 inputs (when Step-R is paired with Step-2, multimodal context survives the reasoning pass).

For production math/code workloads the model is competitive with R1 on MATH-500 and HumanEval, and its reasoning trace can be inspected before the final answer for debugging or audit. The trade-off: no first-class tool calling — Step-R is text-only, so a tool-using agent must wrap it in a planner model (typically Step-2) and use the two in series.

🆚 vs DeepSeek V3.2 / Qwen3.5-Max / Kimi K3

Step-2's main edge is the single-endpoint multimodal pipeline — a single chat-completions call handles an arbitrary mix of text/image/audio/video. DeepSeek V3.2 wins on raw price (10x cheaper input tokens) but ships text-only and forces callers to integrate a separate VL API. Qwen3.5-Max and Kimi K3 lead on context window (1M tokens) and are better for code-archive or long-doc RAG. OpenAI GPT-5 still leads on ecosystem maturity and tool calling but is unreachable from mainland China without proxying.

⚠️ Limits to know

• 128K context trails Kimi K3 and Qwen3.8-Max (both 1M).
• International endpoint is Hong Kong only; US/EU latency higher than domestic.
• Function calling is beta — production callers wrap Step-2 in a planner layer for tool routing.
• No SOC 2 or HIPAA documentation on the public site — confirm enterprise compliance contractually before deploying patient or financial workloads.
• Model release cadence is moderate; some new models stay in beta longer than competing labs.
• No batch or volume discount beyond the standard prepaid tiers.

🎯 Verdict

Choose Stepfun when you need first-party multimodal LLM access from mainland China on a single OpenAI-compatible endpoint, or when a 1/5-price substitute for OpenAI o1 / DeepSeek R1 reasoning (Step-R) is the deciding factor. The 1M-token contexts of Kimi K3 and Qwen3.8-Max are unbeatable for long-doc RAG; DeepSeek V3.2 is the right pick when text-only at 1/40 the cost dominates. For multimodal production in China, Step-2 is currently the cleanest single-endpoint option in the open-weight cohort.

FAQ

Is Stepfun OpenAI compatible?
Yes. Set the OpenAI client base URL to https://platform.stepfun.com/v1.

Which inputs are supported?
Step-2 and Step-2-mini accept text, image, audio and video on a single chat-completions surface; older models (Step-1, Step-1V) are text+image only.

Is there a free tier?
Yes — 1,000,000 free tokens for the first 30 days, then prepaid pay-as-you-go.

What is Step-R for?
Reasoning (math, code review, multi-step planning) at 1/5 the price of OpenAI o1.

Can it be used in China?
Yes — platform.stepfun.com is directly reachable from mainland China, with no proxy or cross-border routing required.