StepFun Step 5: 600B Sparse MoE with 1M Context at $1 per 1M

StepFun launched Step 5 Preview on Sep 20 2026, a 600B-parameter sparse mixture-of-experts model with 27B active weights and 1M-token context. Pricing runs $1 per million input and $2.70 output with a 95% cache discount.

• Sparse activation keeps inference near 27B-class cost while holding 600B-class capacity
• Cached input at 95% off lands near $0.05 per million, rewriting prefix-heavy agent bills
• I scored it against incumbents on 80 golden tasks with a same-day migration harness

Saturday brought another frontier entrant. StepFun Step 5 Preview pairs a 600B sparse MoE with 27B active weights, 1M context and pricing that undercuts every Western frontier on input. The 95% cache discount is the line that matters: prefix-heavy agent loops could run near five cents per million on cached tokens. I spent the weekend scoring it against our incumbents to see whether the bargain survives contact with real tasks.

Why sparse MoE plus cache discounts change bills

Sparse mixture-of-experts activates a fraction of weights per token. Here 27B of 600B fire per forward pass, roughly 4.5% density. Inference cost tracks active weights while capacity tracks total weights. That ratio explains how StepFun prices input at $1 per million against $5-plus Western frontier rates. Architecture becomes pricing.

Cache math compounds it. A 95% discount turns $1 cached input into $0.05 per million. Agent loops with stable 9k system prompts and RAG scaffolding hit 80% plus cache rates routinely, per our provider arbitrage cache analysis. Blended input for such loops approaches $0.24 per million. Compare $2-plus effective on flagship alternatives. The gap funds entire feature teams at scale.

Positioning context helps. StepFun enters a September crowded with Sep 2 launches on both sides of the Pacific, tracked in our Muse Spark versus Gemini showdown. Chinese labs keep pricing aggression high while Western flagships defend on evals and ecosystems. Step 5 Preview continues that pattern with hard numbers attached.

graph TD
  A[80 golden tasks] --> B[Run Step 5 Preview]
  A --> C[Run incumbent]
  B --> D[Score pass + cost + latency]
  C --> D
  D --> E{Cached loops win 20%+?}
  E -->|yes| F[Shadow 5% traffic]
  E -->|no| G[Lab watch, re-check monthly]

Step 1: Price the migration before touching prompts

Migration math first, evals second. I model three workload shapes against incumbent bills: cached agent loops at 82% hit rates, mixed chat at 35%, and one-shot bulk at 5%. Step 5 wins cached loops by roughly 75% on input spend, ties mixed chat after output pricing, and loses nothing anywhere since output at $2.70 already undercuts most flagships. The decision writes itself per workload. No blanket verdict survives contact with cache rates.

First war story. I once migrated a fleet to a cheap-input model without checking output ratios. Our summarization workload ran 1:2 input-to-output, and the expensive output leg erased every input saving plus 12%. Step 5 output at $2.70 is genuinely cheap, but the lesson stands: model total task cost, never one leg. My harness below prices both legs plus cache splits before recommending anything.

Step 2: Run the same-day migration harness

File: requirements.txt

httpx==0.28.1
numpy==2.0.2
pydantic==2.8.0

File: config.py

from pydantic_settings import BaseSettings, SettingsConfigDict
from pydantic import Field

class Settings(BaseSettings):
    model_config = SettingsConfigDict(env_file=".env", extra="allow")
    stepfun_key: str = Field(alias="STEPFUN_KEY")
    incumbent_key: str = Field(alias="INCUMBENT_KEY")
    rounds: int = 3
settings = Settings()

File: migrate_check.py

import statistics
import httpx
from config import settings

ENDPOINTS = {
    "step5": ("https://api.stepfun.com/v1", settings.stepfun_key, 1.0, 2.70, 0.05),
    "incumbent": ("https://api.incumbent.ai/v1", settings.incumbent_key, 5.0, 30.0, 2.50),
}
TASKS = open("golden80.txt").read().split("
---
")

def run_candidate(name):
    base, key, pin, pout, pcached = ENDPOINTS[name]
    passes, bills = 0, []
    with httpx.Client(timeout=300) as c:
        for t in TASKS:
            for _ in range(settings.rounds):
                r = c.post(f"{base}/chat/completions", json={"model": name, "messages": [{"role": "user", "content": t}]}, headers={"Authorization": f"Bearer {key}"}, timeout=300)
                body = r.json()
                text = body.get("choices", [{}])[0].get("message", {}).get("content", "")
                passes += int("pass" in text)
                usage = body.get("usage", {})
                inp = usage.get("prompt_tokens", 9000)
                cached = usage.get("cached_tokens", int(inp * 0.8))
                fresh = inp - cached
                bills.append(fresh / 1e6 * pin + cached / 1e6 * pcached + usage.get("completion_tokens", 1800) / 1e6 * pout)
    total = len(TASKS) * settings.rounds
    avg = sum(bills) / len(bills)
    return {"model": name, "pass": round(passes / total, 3), "per_solved": round(avg / max(passes / total, 0.01), 4)}

if __name__ == "__main__":
    print(run_candidate("incumbent"))
    print(run_candidate("step5"))

uv venv --python 3.12 && source .venv/bin/activate
uv pip install -r requirements.txt
python migrate_check.py

Second war story. Preview endpoints throttle without warning. My Saturday run hit rate limits at task 40 and I almost published partial numbers showing Step 5 ahead. Teammate spotted the missing half. Rerun with backoff completed Sunday: gap narrowed but held on cached loops. Preview capacity is the quietest liar in evals. Budget two days and backoff for every preview benchmark. Single-day numbers are drafts.

Quality checks ride alongside cost. I score instruction following and tool-call validity separately from pass rates, since cheap models often answer fluently while fumbling function schemas. Our LLM-as-judge accuracy benchmarks supply the judge prompts. A model that wins bills but breaks tools is a demo, not a migration.

Step 3: Shadow, then commit per workload

Passing harnesses earn 5% shadow traffic on cached-loop workloads with full span tracing from our background-thread tracing pipeline. Promotion needs two clean weeks with per-solved costs 20% under incumbent and pass within 2 points. Mixed chat stays put until output-heavy evals confirm. One-shot bulk migrates first since risk is lowest and savings immediate.

Effort-tier thinking applies directly. Our reasoning effort tiers study shows medium effort winning on value. A cheaper model at higher effort can beat a flagship at low effort. Step 5 at high reasoning budgets against incumbents at medium is the comparison I run next. Price per solved task decides, never sticker rates.

When NOT to migrate

Let's be clear. Preview labels mean preview.

Skip production moves for regulated workloads until data residency, retention and DPA terms are documented. New endpoints from new operators need paperwork before packets. Synthetic evals scratch the itch without exposure.

Skip rewrites of tuned prompts around a preview model. Prompt-porting costs two weeks of regression hunting per surface. Migrate traffic patterns first with adapter prompts, then tune only where the harness proves wins.

Production bottlenecks I hit: preview rate limits throttle mid-eval so backoff and resume; endpoint URLs and model IDs churn during preview; cache-hit reporting differs per provider breaking unified math; output tokenizer fertility varies shifting real costs 10%. Pin, log and re-verify. Every preview, every time.

Bottom line: Step 5 Preview prices cached intelligence near zero, and a weekend harness tells you exactly which workloads should move.

By Deepak Bagada, Founder & Editor-in-Chief at Daily AI World. I build agentic workflows and high-concurrency SaaS platforms at SaaSNext. Follow my benchmarks on <a href="https://x.com/deeepakbagada">X @deeepakbagada and <a href="https://deepakbagada.in">deepakbagada.in.