Model

Step 5 Preview

CLOSED

StepFun · Step family · released Sep 20, 2026

StepFun's new flagship for agentic work: a 600B-total / 27B-active sparse MoE with a 1M-token context and text/image/video input, launched API-only around 2026-09-20 with open weights promised for October 15. StepFun's own table runs it at High effort against rivals at Max; it ties or edges Kimi K3 and GLM-5.3 on DeepSWE v1.1 (67.7) and GPQA (93.5) but trails them on Terminal-Bench v4 (33.3 vs GLM-5.3 41.9) and several agent rows. Priced at $1.00/$2.70 per Mtok. StepFun's comparison set contains no 'Step-4', 'Kimi K3.1' or other model the corpus added on 2026-09-23. Release date is secondary-sourced: StepFun pages print no date, and 2026-09-20 comes from MarkTechPost coverage (https://www.marktechpost.com/2026/09/20/stepfun-launches-step-5-preview/).

ReasoningCodingVisionFunction callingTool useAgentic

2693.5

Elo · rank #23

Parameters

600B

Active params

27B (MoE)

Context

1M tokens

Architecture

Sparse MoE, 600B total / 27B active, 1M-token context, text/image/video input; reasoning effort low/medium/high

License

Proprietary for now (open weights promised for October 15, 2026)

Languages

—

API price (in/out)

$1 / $2.7

Modalities

text · vision · video

Benchmark results

Bar shows position within the tracked field; marker = field best

AutomationBenchAgents44.0%#9

best: DeepSeek-V4.1-Flash · 54.8%

BrowseCompAgents88.7%#5

best: Kimi K3 · 91.2%

CyberGymCoding84.7%#4

best: MiMo-V2.6-Flash · 95.1%

DeepSWECoding67.7%#13

best: Gemini 4 Argon · 77.9%

GDP.PDFVision14.8%#6

best: GPT-5.6 Sol · 40.0%

GPQA DiamondReasoning93.5%#11

best: GPT-6 Astra · 96.0%

Humanity's Last ExamReasoning46.5%#17

best: Claude Opus 5.5 · 67.7%

JobBenchAgents59.0%#7

best: Claude Opus 5 · 65.7%

MCP AtlasAgents85.6%#1

best: this model · 85.6%

MMMU-ProVision76.0%#27

best: Claude Opus 4.7 · 85.5%

OfficeQA ProAgents60.3%#4

best: Claude Fable 5.1 · 69.0%

SciCodeCoding58.9%#1

best: this model · 58.9%

SWE-Atlas Codebase QnACoding63.6%#1

best: this model · 63.6%

τ³-BankingAgents42.5%#1

best: this model · 42.5%

Terminal-Bench 2.0Coding85.0%#12

best: DeepSeek-V4.1-Flash · 90.6%

Terminal-Bench 4.0Agents33.3%#13

best: Claude Sonnet 5.5 · 70.6%

Toolathlon (Verified)Agents74.1%#7

best: GLM-5.3-Flash · 78.4%

Run it locally

Closed weights — available via API only. No local deployment.

Input / M tok

Output / M tok

$2.7

Step family

Elo progression across releases

Step 3.7 FlashMay 20262182.4 Step 5 PreviewSep 20262693.5

Resources

DOCShttps://platform.stepfun.ai/docs/en/guides/models/step-5-preview APIAvailable via StepFun API

API price $1/$2.7 · each benchmark row carries its own source badge (see methodology)