Step 5 Preview is a large-scale language model designed to provide high-tier intelligence at a lower cost than many frontier AI offerings. The model is built on a sparse Mixture-of-Experts (MoE) architecture, featuring a total of 600 billion parameters, though only 27 billion parameters are active per token. It supports a 1-million-token context window and includes vision input capabilities. By utilizing the MoE architecture, the model attempts to deliver the reasoning power of a massive model with the inference efficiency of a much smaller one.
The model's positioning is defined by its performance relative to other high-end models. It scores 44 on the Artificial Analysis Intelligence Index, a score that matches Kimi K3 and slightly trails GLM 5.3. In terms of pricing, it is positioned as a cost-effective alternative to the most expensive frontier models, potentially making it a viable replacement for developers who do not require the absolute peak performance of OpenAI or Anthropic models but need more capability than smaller, cheaper options. Its agentic capabilities have been demonstrated in complex environments, such as sustaining progress for over 3,000 turns in a Pokémon game without specific optimizations, reaching roughly one-third of the main story.
Despite its technical specifications, the model has faced scrutiny regarding its output style and the authenticity of its promotional materials. Some users have criticized the model's writing style as 'waffly,' noting that it may produce excessive padding and long reasoning traces that can detract from the final answer. Furthermore, community members pointed out inconsistencies in the promotional demo videos; one observer noted a visible error log in the video's thinking trace suggesting the AI was attempting to access a project that was already completed. While the model is scheduled for an open-weights release on October 15, these observations highlight the trade-offs between high-parameter reasoning and the practical usability of the model's generated content.