stepfun-aiReleased Sep 18, 2026

Step-5-Preview

600B total and 27B active parameters; accepts images and video, with FP8 and BF16 weights for deployment.

Source description

Step 5 Preview is StepFun's flagship model for agentic work. It is a sparse Mixture-of-Experts model with 600B total parameters and 27B active per token, a 1M-token context window, and text and image input. It delivers frontier-level performance in software engineering, long-horizon agentic execution with repeated tool calls, and professional knowledge work, with particular strength in finance.

Context window1M

Max output64K

Provider endpoints6

Benchmark results19

Image input Video input Tool calling Structured output Reasoning Prompt caching Open weights

Measured performance

Benchmark results

All leaderboards ↗

Artificial Analysis Intelligence Index ↗4.3.2

43.73

Effort not specified

Independent resultChecked Oct 18, 2026

AA-Briefcase ↗1.1

1,424.32

Effort not specified

Independent resultChecked Oct 18, 2026

GDPval-AA ↗2.1

1,586.63

Effort not specified

Independent resultChecked Oct 18, 2026

AutomationBench-AA ↗AA Index 4.3.2 collection

50.96%

Effort not specified

Independent resultChecked Oct 18, 2026

Terminal-Bench 4.0 · AA ↗4.0

33.33%

Effort not specified · Harness unreported

Independent resultChecked Oct 18, 2026

SciCode ↗AA Index 4.3.2 collection

58.91%AA · Under review

Effort not specified

Independent resultChecked Oct 18, 2026

Humanity’s Last Exam ↗AA Index 4.3.2 collection

46.48%

Effort not specified

Independent resultChecked Oct 18, 2026

GDP.pdf ↗AA Index 4.3.2 collection

14.8%

Effort not specified

Independent resultChecked Oct 18, 2026

CritPt ↗AA Index 4.3.2 collection

20.86%AA · Under review

Effort not specified

Independent resultChecked Oct 18, 2026

AA-Omniscience ↗AA Index 4.3.2 collection

16.38

Effort not specified

Independent resultChecked Oct 18, 2026

88.33%

Effort not specified

Independent resultChecked Oct 18, 2026

LiveBench Overall ↗2.0 · 2026-10-14 (dataset 2026-09-02)

62.01%

Effort: max · single run · equal mean of 7 suites

Independent resultChecked Oct 18, 2026

LiveBench Design ↗2.0 · 2026-10-14 (dataset 2026-09-02)

76.9%

Effort: max · single run

Independent resultChecked Oct 18, 2026

LiveBench Research ↗2.0 · 2026-10-14 (dataset 2026-09-02)

72.8%

Effort: max · single run

Independent resultChecked Oct 18, 2026

LiveBench Browser Use ↗2.0 · 2026-10-14 (dataset 2026-09-02)

24%

Effort: max · single run

Independent resultChecked Oct 18, 2026

LiveBench Coding ↗2.0 · 2026-10-14 (dataset 2026-09-02)

85.6%

Effort: max · single run

Independent resultChecked Oct 18, 2026

LiveBench Agentic ↗2.0 · 2026-10-14 (dataset 2026-09-02)

27.5%

Effort: max · single run

Independent resultChecked Oct 18, 2026

LiveBench Data Analysis ↗2.0 · 2026-10-14 (dataset 2026-09-02)

63.5%

Effort: max · single run

Independent resultChecked Oct 18, 2026

LiveBench Reasoning ↗2.0 · 2026-10-14 (dataset 2026-09-02)

83.8%

Effort: max · single run

Independent resultChecked Oct 18, 2026

Provider measurements

Speed

1 configuration

≈10K input tokens1 requestmedian / 72h

Artificial Analysis · first-party providers · serving endpoint not reported

StepFun

Step 5 Preview

Reasoning Not reported

Output rate: 91.1422 tok/s
Time to first chunk: 2.18962 s

Last checked Oct 18, 2026, 00:17 UTCSource ↗

Methodology 2.2.0 & measurement details

First chunk includes request processing and network time, and may include reasoning.

Prompt: long · ≈10,000 input tokens
Parallel requests: 1
Minimum answer: 1,500 tokens
Aggregation: median over 72 hours
Token counting: o200k_base
Benchmark client: Google Cloud us-central1-a
Serving endpoint: Not reported
Serving region: Not reported
Sample size: Not reported
Window end: Not reported

Official methodology ↗

Access & cost

Prices across providers

6 quotes

Your token mix

USD · listed token rates

Input million tokensOutput million tokens

$1.675

InferenceNet · via OpenRouterinference-net/fp4 · quantization=fp4

$1.675

Featherless · via OpenRouterfeatherless/fp4 · quantization=fp4

$1.675

DeepInfra · via OpenRouterdeepinfra/fp8 · quantization=fp8

$1.675

StepFun · via OpenRouterstepfun/fp8 · quantization=fp8

$1.675

OpenRouter

Catalog quote

Input: $1.00
Output: $2.70
Cached input: $0.10

USD / 1M tokens

1M context64K max outputPrice source ↗Checked Oct 18, 2026

Endpoint ID

stepfun/step-5-preview

DeepInfra

Input: $1.00
Output: $2.70

USD / 1M tokens

1M contextPrice source ↗Checked Oct 18, 2026

Endpoint ID

stepfun-ai/Step-5-Preview

InferenceNet

via OpenRouter

inference-net/fp4 · quantization=fp4

Input: $1.00
Output: $2.70
Cached input: $0.05

USD / 1M tokens

1M context131.1K max outputPrice source ↗Checked Oct 17, 2026

Endpoint ID

stepfun/step-5-preview::inference-net/fp4

Featherless

via OpenRouter

featherless/fp4 · quantization=fp4

Input: $1.00
Output: $2.70
Cached input: $0.10

USD / 1M tokens

262.1K context32.8K max outputPrice source ↗Checked Oct 17, 2026

Endpoint ID

stepfun/step-5-preview::featherless/fp4

DeepInfra

via OpenRouter

deepinfra/fp8 · quantization=fp8

Input: $1.00
Output: $2.70
Cached input: $0.10

USD / 1M tokens

1M context65.5K max outputPrice source ↗Checked Oct 17, 2026

Endpoint ID

stepfun/step-5-preview::deepinfra/fp8

StepFun

via OpenRouter

stepfun/fp8 · quantization=fp8

Input: $1.00
Output: $2.70
Cached input: $0.10

USD / 1M tokens

1M context64K max outputPrice source ↗Checked Oct 17, 2026

Endpoint ID

stepfun/step-5-preview::stepfun/fp8

Model specifications

Developer: stepfun-ai
Modality: Language
Context: 1,048,576 tokens
Maximum output: 64,000 tokens
Weights: Open weights
Released: Sep 18, 2026
Provider listed: Oct 4, 2026
First discovered: Oct 16, 2026

Exact identities

openrouterstepfun/step-5-preview
deepinfrastepfun-ai/Step-5-Preview
artificial-analysis094ea3e3-ce41-419e-8bd1-723389663158
livebenchstep-5-preview

Sources

OpenRouter ↗DeepInfra ↗Artificial Analysis ↗LiveBench ↗Model documentation ↗

Updated Oct 18, 2026