Skip to content
NewsModel release

StepFun launches Step 5 Preview, a 600B sparse MoE at $1 per million input tokens with open weights due October 15

· by Pondero Newsdesk

The short version

StepFun shipped the Step 5 Preview API on September 20, 2026, offering a 600 billion-parameter sparse mixture-of-experts model with 27 billion active parameters per token, a 1 million-token context window, and open weights scheduled for October 15.

StepFun launches Step 5 Preview, a 600B sparse MoE at $1 per million input tokens with open weights due October 15

StepFun shipped API access to Step 5 Preview on September 20, 2026, pairing coding and agent benchmarks it positions at the frontier with pricing that, per StepFun, runs at roughly one-seventh the cost of GPT-5.6 Sol at a comparable intelligence-index score. Open weights are scheduled for October 15.

What

Step 5 Preview is a sparse mixture-of-experts model with 600 billion total parameters and 27 billion activated per token. The architecture uses a 92-layer narrow-deep Transformer. Per StepFun, deeper stacks produce longer information paths for multi-hop reasoning during long prefill sequences, the kind agents generate while searching, executing code, and ingesting tool outputs. A 1 million-token context window supports extended sessions, and the model accepts both text and image inputs.

To make million-token sessions cheaper to run, the model uses Sparse Grouped-Query Attention with block-wise token merging. Per StepFun's product page, this cuts attention selection cost to roughly one-eighth of a comparable dense baseline.

On StepFun's own benchmark suite, Step 5 Preview at the High compute tier scored 67.7 on DeepSWE v1.1 (Kimi K3 Max: 67.5; GPT-6 Astra Max: 74.1; Claude Opus 5 Max: 74.0), 49.0 on StepCodeBench (Kimi K3 Max: 43.9; Claude Opus 5 Max: 63.9), and 80.5 on ProgramBench (Kimi K3 Max: 77.8; GPT-6 Astra Max: 85.4). On Terminal-Bench v4, Step 5 Preview scored 33.3, behind GLM-5.3 at 41.9 and GPT-6 Astra Max at 57.9 (per StepFun's Step 5 Preview page).

API pricing opened at $1.00 per million input tokens (cache miss), $0.05 per million tokens (cache hit), and $2.70 per million output tokens. Artificial Analysis placed the model at 44 on its intelligence index, alongside Kimi K3 Max, per Pandaily's coverage.

The Hugging Face repository for Step 5 Preview currently holds only a placeholder file. Full weights are expected on October 15.

Why it matters

The $1.00 cache-miss input price is competitive, but the $0.05 cache-hit rate is the number that changes the economics for agent operators. Long-horizon coding loops that reload the same codebase across turns benefit most: a repeated 500k-token context drops from $0.50 per call to $0.025, roughly a 20x reduction in context cost on cache hits.

The October 15 open-weights date is the second variable to track. Self-hosted deployment of a 600B MoE removes per-token billing entirely for operators with GPU infrastructure, the same dynamic that drove adoption of Kimi K3 and DeepSeek R2. StepFun is following the same playbook: competitive API pricing during a preview window, open weights to follow.

One caveat: the benchmark figures above are StepFun's own and reflect the High compute tier. No independent third-party coding evaluation has published results as of September 21, 2026.

What to watch next

Whether the October 15 open-weights release ships on schedule is the immediate question. Independent benchmark results from third-party labs will provide a clearer picture of where Step 5 Preview actually sits versus Claude Opus 5 and GPT-6 Astra. Pricing after the preview period ends has not been confirmed.

Sources