Step-5-Pre
StepFun · Undisclosed (MoE)
The generation that carries StepFun into tier S
A preview of StepFun’s fifth generation, released 1 October 2026 and two generations past the edge-orchestration Step-3: perception 96, documents 95 and agents 94 all sit in tier-S territory, video 91 gives the family long-video QA for the first time, and a composite of 93.0 ranks #10 overall. At ¥4/1M output tokens it lands between GLM-5.3-Flash (¥3) and GLM-5.3 (¥9). The trade-off is explicit: generation 78, no creative work and no open weights.
Last updated Oct 18, 2026Editorial scoring
Bottom line
Eight-dimension breakdown
editorial mapping · 0–100- Step-5-Pre93.0
Weights: perception 20 / documents 18 / charts 14 / video 14 / agents 14 / generation 6 / Chinese 6 / ecosystem 8. Definitions and sources on the methodology page.
Public benchmarks
as reported, not converted| Benchmark / metric | Value | Note |
|---|---|---|
| 模态 | 4 | Text/image/video/audio |
| 输出单价 | ¥4 | Between the two tier-S options |
| 上下文 | 128K | Official spec |
Public scores come from different versions and prompts and are not interchangeable with our composite; they are one evidence input.
Scenario mapping
five stress tests, one scale| Scenario | Score | Observation |
|---|---|---|
| Chinese receipt OCR | 95 · Stable on official forms and receipts; handwriting weaker | |
| Chart data extraction | 93 · Reliable on stacked multi-series charts | |
| 30-min video QA | 91 · Long video is usable for the first time; occasional drift | |
| Mobile GUI control | 94 · Cross-app planning near the top tier | |
| Chinese cultural context | 93 · Solid general context; average on obscure allusions |
Strengths and limitations
Strengths
- Perception 96 matches Kimi K3, and video 91 gives StepFun long-video QA for the first time.
- At ¥4/1M output tokens it sits between the two tier-S options while matching flagship-level perception.
- Same stack as Step-3, with one multimodal interface across edge and cloud, so migration is cheap.
Limitations
- Generation and editing at 78 — effectively no usable text-to-image or Chinese glyph rendering.
- 128K context is small by 2026 standards; whole manuals and long video still need chunking.
- Closed weights with undisclosed parameters, grade B, and third-party reproduction still pending.
Price and real cost
List price
¥4/M tokens
≈ ¥4 / 1M output tokens. For 1,000 images at 1,024px, ~1,100 image tokens each and 300 output tokens, the monthly cost is about ¥2.4 (assuming input at 25% of output price).
Use your volumeAbout Step-5-Pre
Is Step-5-Pre open source?
Licence: Commercial + on-prem licence. No public weights; access is via vendor API or an on-prem licence.
How much does Step-5-Pre cost?
≈ ¥4 / 1M output tokens. It ranks #10 with a composite of 93.0. Real bills depend on resolution, tiling, caching and volume discounts.
What is Step-5-Pre best for?
A flagship-class stand-in for edge-cloud products; Production video understanding; Speech-plus-vision-plus-video devices。
Why does Step-5-Pre score 93.0?
The composite weights eight dimensions: perception 20%, documents 18%, charts 14%, video 14%, agents 14%, generation 6%, Chinese 6%, ecosystem 8%. Its strongest dimension is Visual perception & grounding; its weakest is Generation & editing.
Personal ranking · Not an official score
Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.