Private AI

Nano GPT logo
NanoGPT
Nano GPT logo
NanoGPT
Nano GPT logo
NanoGPT
Search
 
 
 
 
All conversations

Create

Chat
Studio

Explore

Models
Pricing
AI Apps
Blog

Library

Gallery
Workspace
Usage
API
Suggestions and bugs
Terms

|

Privacy

|

Refunds
Balance
Subscription
Teams
Referrals
Presets
Settings
FAQPress & mediaBrand assetsBenchmarksEarn crypto
  1. Models
  2. Text
  3. Stepfun

Step 5 Preview

Step 5 Preview is StepFun's 600B sparse MoE frontier model for production-scale agents, activating 27B parameters per token. It is built for software engineering, long-horizon tool use, research, professional knowledge work, and finance, with native text, image, and video understanding and a 1M-token context window. ⚠️ Note: This model routes through StepFun, so privacy and logging guarantees may be limited.

  • Reasoning
  • Vision
  • Video Input
  • Tool Calling
  • Parallel Tool Calls
  • Structured Output

Added Sep 20, 2026

Model weights

Pricing

Auto routing · per 1M tokens
Input
$1.00
Output
$2.70
Cache read
$0.05
Compare provider prices

Specifications

Context window
1M
Max output
65.5K
Parameters
600B / 27B
Total / active
Avg output (7d)
1.5K tokens
Longer than 87% of models

Benchmarks

Sourced from Artificial Analysis.

Intelligence Index

28.5

Intelligence Index: 28.5
Better than 46% of models compared

Agentic work

AutomationBench-AA

Workflow automation with guardrail penalties

51.0%

Better than 64% of models compared

AutomationBench-AA Tasks Completed

Fully completed workflows without guardrail violations

17.7%

Better than 62% of models compared

AA-Briefcase v1.1

Agentic knowledge work (Elo)

1425 Elo

Better than 77% of models compared

GDPval-AA v2.1

Economically valuable tasks (Elo)

1594 Elo

Better than 87% of models compared

Document reasoning

GDP.pdf

Professional PDF reasoning: all-pass rate

14.8%

Better than 54% of models compared

AA-LCR v1.1

Long context reasoning with updated grading

88.3%

Better than 99% of models compared

MLCR-AA

Medical long-context reasoning

18.3%

Better than 63% of models compared

Reasoning

HLE

Humanity's Last Exam

46.5%

Better than 93% of models compared

CritPt

Research-level physics reasoning

20.9%

Coding

Terminal-Bench v4.0

Practical coding and terminal tasks

33.3%

Better than 81% of models compared

SciCode

Python programming for scientific computing

58.9%

Better than 93% of models compared

Knowledge

AA-Omniscience Accuracy

Proportion of correctly answered questions

36.4%

AA-Omniscience Hallucination Rate

Rate of incorrect answers among non-correct responses

31.3%

Legacy benchmarks

AA-LCR (unversioned / legacy)

Long context reasoning evaluation

88.3%

Better than 99% of models compared

GDPval-AA (unversioned / legacy)

Economically valuable tasks

54.7%

Last updated Oct 30, 2026

Artificial Analysis

Providers

Loading provider options…

Related text models

Compare Step 5 Preview with similar models from the same provider or model family.

Step 3.7 Flash Thinking

stepfun/step-3.7-flash:thinking

Step 3.7 Flash Thinking is StepFun's high-efficiency multimodal MoE model with visible reasoning enabled for deeper agentic coding, long-context reasoning, tool use, and native image/video understanding. ⚠️ Note: This model routes through StepFun, so privacy and logging guarantees may be limited.

Step 3.5 Flash 2603

stepfun-ai/step-3.5-flash-2603

Step 3.5 Flash 2603 is optimized for high-frequency agentic and coding workflows with improved token efficiency and faster reasoning. NOTE: This model runs via StepFun, which may log and train on your prompts.

Step 3.5 Flash

stepfun-ai/step-3.5-flash

StepFun's most capable open-source reasoning model with visible reasoning traces. Built on a sparse Mixture-of-Experts architecture with 196B total parameters and only 11B active per token, it achieves frontier-level performance in math, logic, and agentic coding while reaching up to 350 tokens/sec. Supports 256K context. NOTE: This model runs via StepFun, which may log and train on your prompts.

OrcaCyber Zero 1.5

orca/orcacyber-zero-1.5

OrcaRouter's frontier cybersecurity model, post-trained for authorized security research: vulnerability research and reproduction, exploit development, penetration testing, and security auditing. Supports a 1M-token context window, optional reasoning, tool calling, and structured output. Offensive work is permitted only within an authorized security engagement.

Particle-1

particle-ai/particle-1

Particle-1 is Particle AI’s coding and data-engineering model, post-trained on DeepSeek V4.1 Flash. It accepts text and images, supports tool calling, and offers a shared 1,048,576-token context window with up to 131,072 output tokens, including reasoning. Reasoning defaults to high and can be turned off. JSON formatting is supported; strict JSON-schema enforcement is not guaranteed.

DeepSeek V4.1 Flash 0910 Uncensored

deepseek/deepseek-v4.1-flash-0910-uncensored

Uncensored DeepSeek V4.1 Flash with text and image input, optional reasoning, and tool calling. The date label is retained for compatibility; the served checkpoint can vary with availability.