On 20 September, Chinese AI lab StepFun released Step 5 Preview: a 600-billion-parameter sparse mixture-of-experts model with a one-million-token context window, multimodal input (text, image, and video), and API access open from day one. The headline is not that it beats the frontier — it doesn’t — but where it lands and what it costs: benchmark results within a few points of the strongest closed models, at roughly a tenth of typical frontier pricing, with open weights promised for 15 October.
The technical substance
The architecture is the now-standard economics trick of sparse MoE, pushed further than most: 600B total parameters, but only 27B active per token. In practice that means the model stores frontier-scale knowledge while paying inference costs closer to a mid-size model — which is exactly what makes the pricing possible. The specs that matter for developers:
- Context: 1M tokens in, up to 64k tokens out — enough to hold a mid-sized codebase or a full day of agent history in a single window;
- Multimodal input: text, image, and video;
- Pricing on StepFun’s API: $1.00 per million input tokens on a cache miss, $0.05 on a cache hit, and $2.70 per million output tokens, with reasoning tokens billed as output;
- Intended use: StepFun explicitly targets software engineering, long-horizon agent tasks, professional knowledge work, and finance.
That cache-hit price deserves a second look. Agentic workloads re-read the same context over and over — the system prompt, the tool definitions, the accumulated history. At $0.05 per million cached input tokens, the economics of an agent that loops hundreds of times over a large shared context change qualitatively: the context stops being the dominant cost.
Where it lands against the frontier
On the aggregate Artificial Analysis Intelligence Index, Step 5 Preview scores 44 — below the closed frontier, but in serious company. On DeepSWE v1.1, the software-engineering benchmark, StepFun reports 67.7 at high reasoning effort: a hair above Kimi K3 Max (67.5) and GLM-5.3 Max (66.9), and clearly behind GPT-6 Astra Max (74.1) and Claude Opus 5 Max (74.0).
Read those numbers with the usual caution — they are vendor-reported, benchmark suites saturate and get gamed, and “Preview” in the name means the final model may differ in both directions. But the shape of the result is consistent with the broader 2026 pattern: Chinese open-weight labs (DeepSeek, Moonshot, Zhipu, and now StepFun’s latest) arriving six-to-twelve months behind the closed frontier and pricing at a level the frontier cannot match without burning margin.
Why this matters for the future of IT
Two reasons, one economic and one structural.
Economically, the marginal price of “good-enough intelligence” keeps collapsing. A 67-68 on a hard software-engineering benchmark was frontier territory not long ago; it now costs $1 per million tokens and will shortly be downloadable. For most production workloads — code review, document processing, internal agents, ETL with judgment — the question is no longer “can a model do this?” but “what is the cheapest model that clears the bar?” Systems built on that assumption will look very different from systems built around one expensive oracle.
Structurally, the promised open-weight release (15 October, if StepFun delivers) continues the pressure that open weights put on closed labs: every capability that becomes freely self-hostable stops being a moat and starts being infrastructure. For companies in Việt Nam and elsewhere with data-residency constraints — including the kind now written into Vietnamese law — a frontier-adjacent model you can run in your own datacenter is not a curiosity; it is the compliance-friendly default.
The open questions are real: whether the weights actually ship on schedule and under what licence, how the model behaves outside curated benchmarks, and whether a 1M-token window holds up in quality at depth rather than merely accepting the input. Worth watching on 15 October.
Sources
- StepFun Launches Step 5 Preview: A 600B-Total, 27B-Active MoE Model With 1M Context — MarkTechPost
- StepFun launches a 600B agent model at $1 per million input tokens — RuntimeWire
- StepFun launches Step 5 Preview with 600B parameters, 1M context, and open weights coming October 15 — DataStudios
- Step 5 Preview: Specs, Price, Benchmarks, and the Gaps — CellCog