Step 5 Preview

Eesel AI provides an analysis of Step 5 Preview as StepFun's (阶跃星辰) new flagship model for complex work and agentic tasks, with emphasis on software engineering and finance. The article explains the sparse MoE design at 600B total / 27B active parameters, detailing how the router selects a small subset of expert su The piece surfaces StepFun's vendor chart claim of 65% lower cost than GLM-5.3 and Kimi K3 at comparable intelligence, flagging it as a pitch rather than an independent verdict. It contextualizes Step 5 Preview within the broader Step model family and notes StepFun's existing open-weights footprint on Hugging Face, whi

StepFun released Step 5 Preview with a one-million-token context window designed for software engineering, finance, and long-horizon knowledge work, accessible only through StepFun's hosted OpenPlatform and API at launch. The sparse MoE architecture holds roughly 600 billion total parameters while activating about 27 billion per token, accepting text, images, and video inputs and returning up to 64,000 output tokens. Developers can configure reasoning effort, use prompt caching and streaming, and connect application-supplied tools via tool calls and structured output. Reported API pricing is $1 per million uncached input tokens, $0.05 for cached input, and $2.70 per million output tokens; StepFun posted solid coding scores, though GPT-6 Astra and Claude Opus 5 scored higher on cited hardest tests. The model can operate developer-supplied tools but cannot independently reach a user's files or external services; StepFun coordinated 950 web fetches in one internal test, signaling the need for retries and safeguards in production agent stacks. Open weights are reportedly scheduled for October 15.

There's An AI For That describes Step 5 Preview as StepFun's newest flagship model announced September 20, 2026, built on a sparse MoE architecture with 600B total / 27B active parameters arranged across 92 transformer layers in a narrow-deep configuration. It supports a 1M-token context window for both input and outpu The listing documents capability for long-horizon agentic work, including coordinating hundreds of sequential tool calls (around 950 web fetches) within a single agent run, and cites benchmark performance on FrontierFinance, DRACO, DeepSWE v1.1, and ProgramBench. API access is reported at $1.00 per million input tokens

MarkTechPost reports on StepFun's release of Step 5 Preview, confirming the model's positioning as a flagship for agentic work targeting software engineering, professional knowledge work, and finance, with the launch centered on a cost-efficiency Pareto frontier framing rather than raw leaderboard claims. The article c The article lists canonical specifications including model ID step-5-preview, a 1M-token context window, text/image/video input with text output, and reasoning effort options of low, medium, and high, alongside streaming, tool calling, JSON Mode, JSON Schema, and prompt caching support. It highlights the sparse MoE des

AIbase reports on Step 5 Preview as StepFun's new flagship base model positioned for real-world Agentic tasks, built on a sparse MoE architecture with 600B total parameters, 27B active per token, native support for up to a 1M-token context window, and direct processing of text and visual inputs. The article notes an Ar It highlights that single-task cost is reported at roughly one-eighth of Claude Opus 5, with strong benchmark showings on ALE-CLI, FrontierFinance, DRACO, and GDPval-AA v2 (using Artificial Analysis results as of September 19). The piece also covers StepFun's StepCodeBench spanning 553 independent code repositories, 9

ExplainX provides a detailed third-party analysis of Step 5 Preview's "Advancing the Pareto Frontier" launch on September 20, 2026, contextualizing the claim within Artificial Analysis methodology. The article confirms the 600B total / 27B active sparse MoE architecture, 1M-token context with vision support, and the he The piece notes that Step 5 Preview does not beat GPT-6 Astra or Claude Opus 5 outright on the Intelligence Index, with those models remaining higher at much higher per-task cost, and that open weights are promised for October 15, 2026. It positions Step 5 Preview next to GLM-5.3, Kimi K3, and DeepSeek V4, with current

StepFun released Step 5 Preview on September 20, 2026, at 03:15 UTC, positioning it as a flagship model for agentic and long-horizon knowledge work. The architecture is a 600B-parameter sparse mixture of experts with approximately 27B parameters activated per token, paired with a 1M-token context window and 64k-token output. Supported inputs include text, images, and video, and the company frames the launch under the title "Advancing the Pareto Frontier," emphasizing comparable intelligence at lower cost rather than top-line benchmark dominance. Pricing on StepFun's own API is set at $1.00 per million uncached input tokens, $0.05 on cache hits, and $2.70 per million output tokens, with reasoning tokens counted as output. On StepFun's published benchmark table, Step 5 Preview runs level with Kimi K3 and GLM 5.3 on most rows while trailing GPT-6 Astra and Claude Opus 5 on the hardest coding and agent benchmarks; the comparison uses StepFun's High mode against rivals' Max modes, a fairness caveat noted by analysts. Open weights are promised for October 15, 2026, but no license or technical report was named at announcement.

StepFun's official product page introduces Step 5 Preview as a new flagship model for agentic work, built on a sparse Mixture-of-Experts architecture with 600B total parameters, 27B active per token, a 1M-token context window, and vision input. The page frames the model as advancing the Pareto frontier between intellig The supplied excerpt includes StepFun's vendor benchmark tables covering DeepSWE v1.1, StepCodeBench, ProgramBench, Terminal-Bench v4, Agents' Last Exam (ALE-CLI), GDPval-AA v2.1, FrontierFinance, and DRACO, comparing Step 5 Preview (High) against Kimi K3, GLM-5.3, GPT-6 Astra, and Claude Opus 5. Methodology notes spec