Step 5 Preview

Architectural Overview & Capability Profile

Step 5 Preview is StepFun’s flagship frontier large language model engineered for multi-step scientific reasoning, enterprise agent orchestration, and complex software engineering. Built on a sparse Mixture-of-Experts (MoE) topology comprising 600 billion total parameters with 27 billion active parameters per token, Step 5 Preview achieves an overall RankLLMs rating of 57.8 / 100, placing it in the top echelon of global proprietary systems.

Deployed via StepFun’s cloud API with a 1,000,000-token context window and native multimodal comprehension (supporting text, high-resolution document scans, and technical diagrams), Step 5 Preview introduces advanced autonomous tool-use mechanisms optimized for the Model Context Protocol (MCP) and long-running multi-turn tool loops.

Benchmark Performance Across Reasoning, Coding & Agentic Suites

In comprehensive head-to-head evaluations comparing Step 5 Preview against GLM-5.3, Kimi K3, GPT-6 Astra, and Claude Fable 5.1, the model demonstrates pronounced technical strengths:

1. Reasoning & Formal Knowledge: Step 5 Preview logs 93.5% on GPQA Diamond, matching Kimi K3 and rivaling GPT-6 Astra (96.1%). On AA-LCR v1.1, it scores 88.3%, proving resistant to long-context distraction and multi-hop logical traps.
2. Coding & Software Engineering: Across software engineering evaluations, Step 5 Preview demonstrates balanced execution:
   • ProgramBench: Scores 80.5% pass rate, closely matching Claude Opus 5 (82.3%).
   • SWE-Marathon v1.1: Reaches 72.7% partial score, outperforming GLM-5.3 (67.4%).
   • DeepSWE v1.1: Achieves 67.7%, on par with Kimi K3 (67.5%) and Claude Fable 5.1 (67.4%).
   • Terminal-Bench v2.1 & v4: Registers 85.0% on TB v2.1 and 33.3% on TB v4.
   • Internal StepCodeBench: Leads open architectures with 49.0% across multi-file generation and unit testing.
3. Agentic Tool Use & Protocol Execution:
   • MCP-Atlas: Reaches 85.6%, confirming robust integration with Model Context Protocol tool servers.
   • Draco: Achieves 83.3%, demonstrating disciplined task completion without infinite loops.
   • BrowseComp: Scores 88.7% on live multi-page web information extraction.
   • Toolathlon-Verified: Reaches 74.1%, executing intricate multi-tool sequences.
4. Cybersecurity & Finance:
   • CyberGym: Reaches 84.7%, demonstrating defensive security patching and vulnerability analysis.
   • FrontierFinance: Scores 66.4%, outperforming GPT-6 Astra (55.0%) and GLM-5.3 (64.1%).

Economics & Inference Efficiency

Step 5 Preview is offered at $0.60 per 1M input tokens and $2.40 per 1M output tokens (with $0.08 / 1M KV cache read). With an average generation throughput of 140 tokens per second, Step 5 Preview provides a balanced option for production enterprise pipelines requiring deep reasoning and MCP tool compliance.