Step 5 Preview

StepFun's new flagship for agentic work, announced September 20, 2026 under the banner "Advancing the Pareto Frontier": the pitch is frontier-adjacent intelligence at a fraction of the cost rather than a new top score. A 600B total / 27B active sparse MoE with a 1M-token context, 64K max output, text, image and video input, three reasoning-effort levels, tool calling, JSON Schema output, and prompt caching. Available through StepFun's products and API as step-5-preview; StepFun says the weights will be released on October 15 with no license named yet, so for now it is a closed preview with an open-weight promise. API pricing is $1.00/M input ($0.05 on cache hits) and $2.70/M output including reasoning tokens.

Artificial Analysis scores it 44 on Intelligence Index v4.3 (#26 of 202 at filing, reasoning mode) at $0.72 per index task, on AA's intelligence-versus-cost frontier for its tier. StepFun's own table runs its High mode against rivals' Max modes: DeepSWE v1.1 67.7 (Kimi K3 67.5, GLM-5.3 66.9, GPT-6 Astra 74.1, Claude Opus 5 74.0), Terminal-Bench v4 33.3 (GLM-5.3 41.9, Astra 57.9), Agents' Last Exam CLI 29.5, GDPval-AA v2.1 Elo 1566, ProgramBench 80.5, and the internal StepCodeBench 49.0 avg@4 across 553 repositories and 33 languages. Finance is the stated specialty: FrontierFinance 66.4 (Opus 5 69.7) plus three internal live-search, valuation, and deep-research suites. Three 24-hour long-horizon runs are the most interesting evidence: optimizing an H100 MLA kernel from scratch to 508 TFLOPS (Opus 5: 493), automated post-training that lifted a Qwen3-30B-A3B base from 53.3% to 60% on AIME24, and 3,000+ turns of Pokémon Red reaching three gym badges. In one research task it coordinated 950 web fetches in a single agent action. Successor to the open Step-3.7-Flash line; no technical report yet.