Step 5 Preview reaches a 44 benchmark score at lower API prices

Chinese AI lab StepFun has released Step 5 Preview through its first-party API. The flagship reasoning model succeeds Step 3.7 Flash, and StepFun says open weights will follow on October 15.

Artificial Analysis gives Step 5 Preview a score of 44 on its composite Intelligence Index, compared with a median of 24 among comparable models. That ties Kimi K3 (max) and falls one point behind GLM-5.3 (max) and Qwen3.8 Max. Its listed API prices are substantially lower than those peers.

600B parameters, 27B active

A mixture-of-experts model routes each token through part of the network rather than activating every parameter. The 27B active figure approximates per-token computation. The full 600B count remains relevant to storage, memory, and distributed serving requirements.

A 44 score at roughly one-third the task cost

Artificial Analysis reports spending $924.68 to run Step 5 Preview across its Intelligence Index, equivalent to about $0.72 per task. Kimi K3 (max) posts the same score at approximately $2 per task, and GLM-5.3 (max) costs about $2.01 per task for a score of 45.

Step 5 generated roughly 164 million output tokens during the evaluation, close to Kimi K3’s 170 million. Per-token rates account for most of the cost gap: Kimi charges $3 per million input tokens and $15 per million output tokens, compared with StepFun’s $1 and $2.70. MiMo-V2.6-Pro is the cited exception, combining a higher index score with a lower cost of $0.13 per task.

Production costs will vary with prompt length, output length, cache reuse, concurrency, and provider discounts. The benchmark figures offer a controlled comparison rather than a forecast for every application.

Reasoning drives the score

Step 5 Preview’s largest gains over Step 3.7 Flash appear on demanding reasoning tests. It scores 46% on Humanity’s Last Exam, an expert-level multidisciplinary benchmark, and 21% on CritPt, which tests physics reasoning. Those results improve on its predecessor by 25 and 19 percentage points, respectively, and exceed the reported scores for GLM-5.3 (max) and Qwen3.8 Max on both evaluations.

Broad factual knowledge also contributes to the composite score. Step 5 reaches 42% accuracy on AA-Omniscience, ahead of the 753B-parameter GLM-5.3 (max) at 34% and behind the 2.8T-parameter Kimi K3 (max) at 48%. Parameter comparisons require caution because training data, architecture, and active computation differ across models.

Agents expose the gap

Evaluations involving tools, terminal work, office-style tasks, and multi-step automation produce weaker relative results. Step 5 trails Qwen3.8 Max and GLM-5.3 (max) across all four agent-oriented benchmarks cited by Artificial Analysis.

Higher scores are better across these tests. Percentage results represent benchmark success rates, and Elo scores are relative ratings that should be compared only within the same evaluation.

Frequent answers bring factual risk

Artificial Analysis reports that Step 5 attempts 68% of AA-Omniscience questions and hallucinates on 43% of those attempts. GLM-5.3 answers 55% of questions and records a 30% hallucination rate among its attempts. Step 5 consequently reaches 16 on the combined AA-Omniscience Index, only two points above GLM-5.3’s 14 despite its higher raw accuracy.

Applications that publish factual answers without review should evaluate the model on domain-specific data and add retrieval, citations, confidence thresholds, or human approval. The reported tendency to answer rather than abstain makes raw accuracy an incomplete reliability measure.

Reasoning jobs fit; agents lag

The first-party API produced 92.8 output tokens per second in Artificial Analysis testing. End-to-end latency will also depend on time to first token, prompt size, concurrency, network conditions, and provider load.

Open weights bring deployment questions

Until the scheduled weights arrive, Step 5 Preview remains primarily an API product. A self-hosted deployment will depend on the release license, checkpoint formats, quantization support, routing implementation, and available inference tooling.

The model’s 27B active parameters reduce computation per token, but serving infrastructure still needs to store or distribute the 600B-parameter network. Self-hosted economics will therefore depend on memory capacity, interconnects, batching, and utilization rather than StepFun’s API prices alone.