← Back to blog
StepFun AI Review 2026: Models, Pricing & Which to Pick

StepFun AI Review 2026: Models, Pricing & Which to Pick

Nathan Cole

By Nathan Cole

MyClaw Editorial

MyClaw

Run Best-in-Class AI Agents Now

Run OpenClaw or Hermes Agent on managed MyClaw hosting, so your AI agent stays online, updated, and ready for real work.

AI Takeaway

  • What is StepFun AI? StepFun (阶跃星辰) is a Chinese AI lab shipping Step-series language, multimodal, and speech models plus an OpenAI-compatible API and Step Plan credits for coding and agent tools.
  • Which StepFun AI model should you use first? Start with step-3.7-flash when you need native image or video plus agent or coding loops; choose step-3.5-flash (or the 2603 agent-tuned ID when listed) for fast text-only reasoning and high-volume tool calls.
  • How does StepFun AI pricing work? You either pay API tokens per million (with cheaper cache hits) or subscribe to Step Plan monthly Credits for supported IDE and agent tools — verify live platform tables before you budget; credits do not roll over.
  • Is StepFun free on MyClaw? No forever-free StepFun quota: MyClaw AI Chat offers limited guest access with automatic routing, while signed-in Chat plans and always-on agent hosting bill separately from any StepFun API or Step Plan usage.
  • When does a MyClaw agent beat raw StepFun chat? Use Chat for a quick model try when the ID appears in your picker; move to managed OpenClaw hosting when the same StepFun-backed loop must stay online, keep tools and memory, and run without your laptop open.

StepFun at a Glance

StepFun (阶跃星辰) builds Step-series models, an OpenAI-compatible Open Platform API, and consumer 阶跃 AI apps — separate from MyClaw. In 2026 you mainly choose multimodal agents on Step 3.7 Flash or text-speed loops on Step 3.5 Flash. Paths include API tokens, Step Plan Credits for IDE/agent tools, select open-weight Flash checkpoints, or a soft MyClaw Chat / managed-agent try.

DetailStep 3.7 FlashStep 3.5 Flash family
API IDsstep-3.7-flashstep-3.5-flash, step-3.5-flash-2603
RoleFlagship multimodal reasoningText reasoning / agent-coding speed
Context256K256K
ModalitiesNative image + video; toolsText only; tools
Best forUI/screenshot loops, agents with eyesHigh-volume text tool calls
Open weightsApache-2.0 Flash (infra-heavy MoE)Hosted API / plan path first

For broader Chinese-lab context without a bakeoff here, see the Chinese AI model roundup — this review stays on StepFun lineup, pricing, and which SKU fits.

Flagship vs Speed Text SKUs

step-3.7-flash is the flagship: sparse MoE (~198B total / ~11B active), native multimodal, tools, and three reasoning_effort levels (low / medium / high). step-3.5-flash and agent-tuned step-3.5-flash-2603 stay text-only at a lower first-party API sticker. Treat them as different products — modality is the split, not the shared “Flash” name.

StepFun Open-Sources Step3-VL-10B Multimodal Model - Pandaily

StepFun AI Model Lineup That Matters in 2026

Step 3.7 Flash for Multimodal Agents and Coding

Use Step 3.7 Flash when the agent must see. Drop screenshots, UI captures, charts, or short video into the same OpenAI-compatible loop that already calls tools — no bolted-on vision sidecar. It targets high-throughput reasoning, agents, and coding with a 256K window. Set reasoning_effort to low for extract-and-rewrite, medium as default, and high for deep math or code plans.

Apache-2.0 open weights exist if you can host a 100B+ class sparse MoE; most teams stay on the hosted API or Step Plan.

Step 3.7 Flash - Intelligence, Performance & Price Analysis | Artificial  Analysis

Step 3.5 Flash When Text Speed Is Enough

Step 3.5 Flash covers reasoning, math, software engineering, deep research, and tool orchestration — still 256K, no native image or video. Pick it when the loop is chat, diffs, and tool results only. When listed, step-3.5-flash-2603 is the high-frequency agent variant in the same text pricing class. If the ticket never needs eyes, start on 3.5 Flash to protect latency and spend.

Step Explore, Audio, and Router Notes

Step Explore (step-explore) is application-gated Agent/Coding capacity on Step Plan, with up to about 1M context for approved interactive use — UID-dependent; batch jobs can fall back to 3.7 Flash at 256K. Adjacent SKUs cover speech (StepAudio 2.5 / 3 preview), step-router-v1, and image-edit (step-image-edit-2 retires from Step Plan on October 10, 2026). Map breadth only — not your 3.7-versus-3.5 default.

StepFun AI Pricing Without the Guesswork

API Tokens, Cache Hits, and Rate Tiers

First-party platform pricing (re-checked 2026-09-19 Asia/Shanghai) bills per million tokens:

ModelInput (cache miss)Input (cache hit)Output
step-3.7-flash$0.20$0.04$1.15
step-3.5-flash / step-3.5-flash-2603$0.10$0.02$0.30

Cache hits are the real lever — repeated system prompts and tool schemas should hit, or “cheap Flash” math evaporates. Rate limits scale with cumulative top-up (V0–V5). Speech and image meters are separate; do not budget audio from the Flash row. Re-check live tables before you lock a forecast.

Step Plan Credits for IDE and Agent Tools

Step Plan is a Credit subscription for coding and agent tools (OpenClaw, Claude Code, Cursor-class agents, and siblings) — a different wallet from pay-as-you-go API balance. Credits convert at roughly $1 ≈ 7M Credits. Flash Mini / Plus / Pro / Max differ mainly in monthly Credit volume; boosters refill mid-cycle on their own 30-day clocks. Monthly Credits do not roll over.

Point IDE clients at https://api.stepfun.ai/step_plan/v1 — not the standard API URL. Use Step Plan when IDE Credit budgeting beats raw tokens; otherwise stay on API pay-as-you-go. MyClaw Chat plans and agent hosting stay on separate invoices from StepFun spend.

Honest Strengths and Limits

Pros:

  • Strong agent and coding positioning with native multimodal on Step 3.7 Flash
  • Clear dual commercial paths — API tokens versus Step Plan Credits — documented first-party
  • Open-weight Flash options under Apache-2.0 for teams that can self-host
  • Three reasoning efforts to trade depth versus latency on the same 3.7 Flash ID

Cons:

  • Verbose or high-effort runs can erase “cheap token” savings — measure per job
  • Step Explore and some plan SKUs are gated or inventory-dependent
  • Self-hosting a 100B+ class MoE stays infrastructure-heavy despite “Flash” naming
  • MyClaw Chat availability of StepFun IDs may vary by workspace — trust the live picker

Guest Chat is capped like other models — limited guest access, automatic routing, a single guest chat, no card to start — per AI Chat pricing. A try path, not forever-free StepFun tokens.

Skip the Setup, Run Best-in-Class AI Agents Now

Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.

Get Started

Which StepFun Model to Pick for Your Job

Pick Step 3.7 Flash When…

Choose step-3.7-flash when inputs include screenshots, product UI, whiteboards, receipts, charts, or short video inside an agent or coding loop — or whenever you want one ID that both sees and calls tools. Lock multimodal tickets in your harness before you compare text-only SKUs.

Pick Step 3.5 Flash When…

Choose step-3.5-flash — or step-3.5-flash-2603 when listed — for text-only high-volume agents, cost-sensitive tool calls, and IDE loops that never attach images. Score with tests and diffs at matched settings so “faster” cannot mask unfinished work.

Skip or Escalate When…

Skip inventing Explore or Audio IDs your account is not approved for. If another lab’s Flash-band speed story already fits a measured ticket, see the sibling GLM 5.3 FlashX review. Move to always-on private runtime when laptop sleep and session loss hurt more than model quality — that is hosting, not a new API key.

Try StepFun Workflows in MyClaw Chat or an Agent

Use Chat for a fast try when the model appears; use managed OpenClaw when the loop must stay up. Chat plans, agent hosting, and StepFun spend stay separate invoices.

Step 1: Open Chat and Check the Live Picker

Open MyClaw AI Chat and select a StepFun / Step 3.7 Flash-style ID if listed in your workspace picker. If missing, run the closest listed Chinese or auto-routed model and note the swap. Guests get a limited try under automatic routing; sign in for a fuller picker and history.

Get Started

Step 2: Freeze One Measurable Job

Pin one coding, multimodal, or tool-loop ticket with a written pass/fail rule before the first token streams. Optionally A/B step-3.7-flash versus step-3.5-flash at matched effort and tool settings. Score with tests, diffs, and artifacts — not vibes.

Step 3: Keep Winning Loops Online on Managed OpenClaw

When stable prompts and tools should survive sleep, move them into an always-on OpenClaw agent on MyClaw and configure the StepFun provider API or Step Plan endpoint when your workspace supports it. Soft next step: compare Chat allowance versus hosting pricing — hosting covers the cloud computer; AI token usage is not included.

Final Call: Match the SKU to the Job

Default to Step 3.7 Flash for multimodal agents; Step 3.5 Flash (or 2603 when listed) for fast text agents; Step Plan when IDE Credit budgeting beats raw tokens. Re-check platform price tables and your MyClaw picker before you standardize.

Start with one frozen Chat try, note substitutions, and host the loop when always-on beats a browser tab.

Skip the Setup, Run Best-in-Class AI Agents Now

Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.

StepFun AI Review 2026: Models, Pricing & Which to Pick | MyClaw.ai