Chinese company StepFun (阶跃星辰) announced Step 5 Preview on September 20, 2026 — a flagship model for agentic tasks based on a sparse MoE architecture: 600 billion parameters, of which 27 billion are active per token, a 1 million token context window, and input for text, images, and video. The model is already available via API under the name step-5-preview, with open weights promised for October 15, 2026. On the Artificial Analysis Intelligence index, Step 5 Preview scored 44 points, maintaining, according to the company, a cost significantly lower than comparable models.

image

What happened

StepFun published the announcement and opened access to the step-5-preview model via platform.stepfun.ai. In a single request, the model accepts up to 60 images and videos in MP4, QuickTime, and Matroska formats up to 128 MB in size, and outputs up to 64k tokens. The company showed results on agentic benchmarks: DeepSWE v1.1 — 67.7, ProgramBench — 80.5, ALE-CLI — 29.5, FrontierFinance — 66.4; here Step 5 Preview is first among open-weight models, though it lags behind closed GPT-6 Astra and Claude. On Terminal-Bench v4, the gap is more noticeable: 33.3 versus 57.9 for Claude Opus 5. Along with the model, StepFun released its own StepCodeBench benchmark on 553 real repositories, 9 task categories, and 33 programming languages, and demonstrated two 24-hour autonomous runs: optimization of an MLA GPU kernel on NVIDIA H100 with a result of 508 TFLOPS versus 493 TFLOPS for Claude Opus 5, and autonomous post-training of Qwen3-30B-A3B, which raised accuracy on AIME24 from 53.3% to 60%.

Context

The sparse MoE architecture (Mixture-of-Experts, "mixture of experts") means that not all parameters work on each token, but only a part — for Step 5 Preview, this is about 4.5% of the total volume, which reduces inference cost. The mechanism itself is not new; the novelty is in the combination: 27 billion active parameters combined with a 1 million token context, video input, and a claim to frontier agentic tasks — this is a deliberate bet on workloads where long inference cost dominates. The 24-hour autonomous operation demonstrations set a different format for public proofs: not single questions, but hours-long agentic work with feedback. Publishing tables with losses is a rare signal of honesty for vendors, although the set of benchmarks and metrics is still chosen by the company itself. Reference points for comparison in the open segment remain GLM-5.3 and Kimi K3, and closed ones — GPT-6 Astra and Claude.

Why this matters for the industry

For the industry, this is the entry of another player into the frontier of agentic models with a specific date for opening weights. The pressure mechanism is MoE economics: 27 billion active parameters out of 600 billion provide frontier-level quality at a significantly lower inference cost, and StepFun directly builds its argument around the shift of the "Pareto frontier" — the boundary of the compromise between intelligence and price. If the numbers are confirmed, this is pressure on the pricing models of OpenAI and Anthropic and a reduction in the cost of agentic infrastructure for startups. For builders, what matters more is not the "flagship" title, but a ready-made framework for a new class of products — long autonomous agents with 1M context and video input. After the opening of weights, self-hosting, fine-tuning for domain tasks, and distillation into smaller models will appear, and price pressure on closed APIs will likely increase.

Why this matters for users

You can try the model right now: step-5-preview is available via API on platform.stepfun.ai, including 1M context support, video input, and connection to Claude Code via Step Plan. A practical check for the evening — run your real agentic scenario on it: code review of a repository, video analysis, or a report on a hundred sources, and then compare with current providers by quality and cost of one agentic run, collecting your own latency measurements. It makes sense to check the results against StepFun's public tables, where both victories among open weights and lags behind closed models are visible. Those who follow agentic pipelines and coding assistants should compare the model with GLM-5.3 and Kimi K3. After the opening of weights, the possibility of local launch and fine-tuning for your own tasks will appear.

What is still unknown / limitations

The key figures are still vendor's and have not undergone independent verification. The claimed shift of the "Pareto frontier" relies on the formulation "significantly lower than comparable models" without a single price figure in open materials — this is a marketing statement, not measured economics. 44 points on Artificial Analysis Intelligence is an assessment of a third-party index, the methodology of which differs from vendor benchmarks. In the available materials, there are no published prices, limits, and delays, so it is too early to draw conclusions about the real cost of agentic runs, and production implementation is reasonable to consider only as shadow or pilot traffic. The selection of tasks and metrics in StepCodeBench is controlled by the vendor, there is no independent audit, and checking contamination and reproducing the results of 24-hour runs will only be possible after the opening of weights on October 15, 2026.

Sources

Author

Look at AI, editorial team