--- title: "StepFun open-sources Step 5 Preview — frontier quality at 1/8 the cost of Claude Opus 5?" date: 2026-10-17 category: Companies & Stack site: NeuroAI canonical: https://neuroai.site/a/na-corp-stepfun-step5-open-weights language: en --- # StepFun open-sources Step 5 Preview — frontier quality at 1/8 the cost of Claude Opus 5? > Shanghai startup StepFun released the weights of Step 5 Preview, a 600B-parameter open mixture-of-experts model that it says matches top closed systems on coding and finance while costing far less per task. Open-weight models usually force a trade: capability on one side, cost on the other. StepFun is betting that trade is no longer necessary. Its new Step 5 Preview runs with only 27B parameters active out of 600B, yet it lands inside the global open-model top three on one widely cited index. The pitch is simple — frontier-grade answers for a fraction of the bill — and the timing is deliberate, arriving just as Chinese labs tighten the screws on expensive closed APIs. ## What got released Step 5 Preview was opened through StepFun's API in late September 2026, and on October 15, 2026 the company published the model weights in both BF16 and FP8 precision on Hugging Face and ModelScope. That two-week gap between API access and open weights is becoming a familiar pattern: show developers the model, then let them download it and leave the cloud. The architecture is a sparse mixture-of-experts design with 600B total parameters and only 27B active per token. It supports a 1M-token context window and takes native text and image input. StepFun (阶跃星辰) is a Shanghai large model (大模型) startup that has been shipping open models through 2026, and Step 5 Preview is its most ambitious release to date. Offering both BF16 and FP8 weights matters in practice: FP8 roughly halves the memory footprint versus BF16, making the model runnable on a narrower, cheaper set of accelerators. ## The efficiency claim The number StepFun leads with is cost per task. On the Artificial Analysis Intelligence Index, Step 5 Preview scores 44 points — StepFun says that puts it among the top three open models globally. More striking is the price comparison: a single-task average cost of about US$0.71 (≈ RMB 4.75), against roughly US$5.86 (≈ RMB 39.22) for Claude Opus 5 — about one-eighth. List pricing tells the same story: input at US$1 per million tokens and output at US$2.7 per million tokens. For teams burning tokens on long agent runs, that gap compounds quickly, and it is the kind of arithmetic that moves a procurement decision more than a leaderboard rank. ## Where it is aimed StepFun is positioning the model for AI coding, software engineering, professional knowledge work, and finance — domains where long context and tool use matter more than chit-chat. On the external FrontierFinance benchmark, which covers investment research, Step 5 Preview scored 66.4, which StepFun says ranks second globally. On subsets like ALE-CLI, FrontierFinance, and DRACO, it reports beating Kimi K3 and GLM-5.3 while still trailing the closed GPT-6 Astra and Claude Opus 5. That placement is the strategy: not to beat the very top closed models, but to beat every other open model on the dimensions enterprises actually buy — code, research, and money-related reasoning — at a price no closed lab can match. ## Early traction The model did not wait long to draw a crowd. Within days of the weight release, StepFun said Step 5 Preview reached the number-two spot on OpenRouter's weekly and daily token-usage rankings, with roughly 18.8 trillion tokens consumed across a seven-day window as of October 15, 2026. That is a spike of developer curiosity, not yet proof of production adoption — but for a freshly opened model, it is a loud entrance. ## Where StepFun fits in the open-weight field Step 5 Preview lands in a crowded 2026 open-model field where some Chinese labs have begun gating commercial use of their largest checkpoints while keeping smaller models free. StepFun's choice to release a 600B flagship openly, without a reported revenue threshold, places it in the permissive camp — a deliberate signal to developers deciding where to build. For enterprises evaluating open weights, the differentiator is increasingly not the leaderboard but the absence of a licensing surprise down the road. According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the open-weight race has moved from raw parameters to usable cost per task, and that is what decides adoption. ## Honest limitations - The Intelligence Index score, the 1/8 cost comparison, and the benchmark results are vendor-reported or drawn from Artificial Analysis; we did not re-run them. - "One-eighth the cost of Claude Opus 5" compares list prices on a specific task profile; real bills depend on caching, batching, and workload shape, and the gap narrows or widens accordingly. - The OpenRouter ranking reflects developer interest, not production deployment, and 18.8 trillion tokens in a week is a launch spike that may not persist. - We did not inspect the released weights or the specific license terms on Hugging Face and ModelScope directly. ## What readers can do now - If you build coding or finance agents, benchmark Step 5 Preview against Qwen, Kimi, and GLM on your own evaluation set before trusting index scores. - Check the Hugging Face and ModelScope model cards for license, context limits, and quantization support before self-hosting. - Verify whether the 27B-active efficiency holds at your context length and batch size — MoE savings often shrink on very long sequences. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-corp-stepfun-step5-open-weights Free to quote with attribution and a link to the original.