Step 5 Preview: Advancing the Pareto Frontier

Commenters focused on Step 5 Preview’s technical and marketing positioning. nh43215rgb summarized the model specs (sparse Mixture-of-Experts, 600B total / 27B active per token, 1M-token context, vision input, AI Index score of 44, open weights on Oct 15) and suggested the vendor skipped “4” to imply parity with leading models. bethekind highlighted an extended FireRed playthrough as an unconventional benchmark, claiming the model sustained over 3,000 turns and millions of tokens and comparing its run-time performance to Astra. ghoshbishakh praised the vendor’s messaging as positioning the product as “best among cheaper and a bit less performant” alternatives, and Jacques2Marais pointed others to the official pricing page.

Commenters also argued about naming, roadmap, and real costs. Bolwin noted Moonshot has teased K3.1, arguing other Chinese labs might not mimic a version-skip. dannyw described typical naming as marketing-driven, with K3.1 plausibly a post-train iteration. On pricing, sieve estimated cached-read costs of roughly $8-12/day at $0.04/M, while ignoramous corrected the input cache rate to $0.05/M for Step 5 and called the vendor’s subscription plan “ridiculous” relative to heavy token usage, comparing the model’s API economics to Kimi/GLM-level alternatives. The divide centers on whether the release is substantive advance or marketing theater and whether advertised affordability holds up under real workloads.