StepFun API: Chinese AI Models for Coding
If you are looking for “Step 3.7,” the API detail to keep is the suffix. The
current public model is Step 3.7 Flash, and its exact model ID is
step-3.7-flash. StepFun's official catalog does not list bare step-3.7 as a
callable ID.
Step 3.7 Flash combines a 256K-token context window with tool use and native
image and video understanding. StepFun describes it as a sparse mixture-of-
experts model with 198B total parameters and 11B active parameters. Those
architecture numbers are useful background; the contract your code depends on
is the API ID, context limit, response behavior, and rate card.
Direct price and context
StepFun's public rate card was denominated in CNY when checked on September 11,
- Prices are per 1M tokens:
| Meter | Direct price |
|---|---|
| Cached input | ¥0.27 |
| Uncached input | ¥1.35 |
| Output | ¥8.10 |
The vendor does not publish a USD price on that page. Converting the CNY figure
at today's exchange rate can help with an internal estimate, but it does not
turn the result into an official USD quote.
The documented context window is 256K. The same specification page does not
state a separate maximum output, so reserve output headroom and test the limit
your coding agent actually needs.
Curl: a small route check
StepFun exposes an OpenAI-compatible Chat Completions endpoint:
curl https://api.stepfun.com/v1/chat/completions \
-H "Authorization: Bearer $STEPFUN_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"step-3.7-flash","messages":[{"role":"user","content":"Find the race condition in this design."}]}'
Use a short first request. It proves authentication, model access, and response
shape without spending time debugging a large repository payload.
Python with an OpenAI client
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["STEPFUN_API_KEY"],
base_url="https://api.stepfun.com/v1",
)
response = client.chat.completions.create(
model="step-3.7-flash",
messages=[{"role": "user", "content": "Propose a bounded retry policy."}],
)
print(response.choices[0].message.content)
StepFun also documents an Anthropic-compatible Messages endpoint. Choose the
protocol that matches your existing agent runtime, then verify tool calls and
streaming with your own schemas.
AIWave's current public catalog includes step-3.7-flash. To run the same
OpenAI-shaped request through AIWave, change the base URL to
https://aiwave.live/v1 and use an AIWave key. Check AIWave's dated USD rate
card rather than carrying over StepFun's direct CNY numbers. The benefit is one
credential and one billing record across routes; the tradeoff is the gateway's
markup and another dependency in the path.
Multimodal input is a separate test path
Step 3.7 Flash can inspect images and video, which is useful when a coding task
depends on a UI screenshot or recorded failure. Do not assume the text-only
request above proves that path. Test accepted media formats, size limits,
upload lifetime, and token accounting separately. Strip secrets and customer
data from screenshots before sending them, and keep a text fallback for
environments where media upload is unavailable.
A practical coding-agent setup
Do not send an entire repository because the window allows it. Start with a
repository map, the interfaces touched by the task, the current diff, and the
failing tests. Add files when the model identifies a concrete dependency.
Keep a stable system prompt and stable project summary at the front if you want
cache hits. Put volatile material—timestamps, logs, and current tool output—
later. Record whether input was actually billed as cached; an expected hit is
not evidence of a hit.
For production, separate 401, balance, rate-limit, and upstream errors. Retry
only operations that are safe to repeat, with a cap and jitter. Save the model
ID, input/output tokens, latency, and charge beside the request ID. Those fields
tell you whether a cheaper-looking route is actually cheaper for the workload.
Top comments (0)