Step 5 Preview Released: 600B Parameter MoE Model Highlights Agentic Tasks and Financial Analysis Performance

Step 5 Preview is a flagship model for agentic tasks, featuring a sparse MoE architecture with 600B total parameters (27B active per token). It supports a 1M token context window and vision input, achieving a score of 44 on the Artificial Analysis Intelligence Index and claiming lower task costs compared to models of similar intelligence levels.

Coding and Long-Running Performance

It recorded scores of 67.7 on DeepSWE v1.1, a software engineering benchmark, and 49.0 on StepCodeBench. In GPU kernel optimization tasks on NVIDIA H100, it achieved 508 TFLOPS within a 24-hour limit, surpassing Claude Opus 5 (493 TFLOPS). Additionally, in the Pokémon Red environment, it sustained over 3,000 turns and 6 million tokens of interaction without specialized optimization, progressing through approximately 1/3 of the main storyline.

Financial and Expert Knowledge Analysis

It scored 66.4 on the FrontierFinance benchmark, outperforming GLM-5.3 (64.1) and GPT-6 Astra (55.0), but falling short of Claude Opus 5 (69.7). It can perform 950 web fetches and collect 300,000 monthly records in a single agent action, executing complex domain-specific analyses such as generating workbooks that analyze 17 sheets when reviewing diesel surcharges.

Comparison with Competing Models

In overall reasoning and coding capabilities, there remains a significant gap compared to top-tier frontier models like GPT-6 and Claude Opus 5. It recorded lower scores than competing models in areas such as GPQA Diamond (93.5%) and DeepSWE v1.1 (67.7%).