Step 5 Preview vs DeepSeek V4.1 Flash: Higher Intelligence Ceiling or 60% Cheaper Output?
Compare StepFun's Step 5 Preview and DeepSeek's V4.1 Flash for coding agents, long-context work, and production cost — raw capability vs cost efficiency.
Short Answer
Compares StepFun's Step 5 Preview and DeepSeek's V4.1 Flash — two September 2026 Chinese flagship-class releases — on intelligence, cost, context/output limits, agentic coding and openness, using launch coverage, official sources, pricing trackers and a direct price comparison. VERSIONS COMPARED: Step 5 Preview (600B-A27B sparse MoE, 1M ctx, 64K out) · DeepSeek V4.1 Flash (552B MoE, 8B/16B asymmetric active, 1M ctx, 384K out)
Our Recommendation
AA index 44 vs 39 and 27B active params vs 8B/16B — the bigger reasoning budget for sustained complex tasks, at a higher token price.
Output is ~40% and input ~29% of Step 5 Preview's price (peak-window rates), and cache hits cost a fraction — the default economics pick.
384K max output vs Step 5 Preview's 64K — a 6x larger ceiling for one-shot long documents or code dumps.
NOT A STRICT EITHER/OR — THE TWO CAN BE COMBINED
Comparison at a Glance
| Dimension | Step 5 Preview | DeepSeek V4.1 Flash |
|---|---|---|
| Released | 2026-09-20 (StepFun) | 2026-09-10 (DeepSeek) |
| Architecture | Sparse MoE, 600B total / 27B active | MoE 552B, Causal-Encoder-Decoder; 8B in / 16B out active |
| Context window | 1M tokens | 1M tokens |
| Max output | 64K tokens | 384K tokens |
| Input modalities | Native text + vision (image/video) | Native text + image (vision understanding) |
| Intelligence anchor | AA index 44 (top-3 open globally) | AA index 39 (above GPT-6 Luna's 37) |
| Coding anchor | Tuned for AI coding / SWE / finance agentic work | SuperCLUE agentic terminal coding (zh): 50.51 @ ¥1.86/task |
| API price (input, cache miss) | ¥7 / M tokens | ¥2 / M tokens (~29% of Step 5) |
| API price (input, cache hit) | ¥0.7 / M tokens (raised from ¥0.35 on 2026-10-14) | ¥0.04 / M tokens |
| API price (output) | ¥20 / M tokens | ¥8 / M tokens (~40% of Step 5) |
| Open weights | Weights slated to open-source on 2026-10-15 | Open-weight model |
| Cost positioning | ~1/8 the per-task cost of Claude Opus 5 | ~1/3 the price of DeepSeek V4 Pro |
SOURCE: Baidu Baike (Step 5 Preview / DeepSeek-V4.1 Flash entries) · StepFun release & OpenRouter coverage · DeepSeek official site · Sina Finance price comparison (2026-10-13) · jumohub & OrcaRouter pricing. USD reference for V4.1 Flash: ~$0.3/M input, ~$1.2/M output. Percentage ratios (11% / 29% / 40%) are Sina Finance's own calculation against Step 5 Preview's then-current rates; the cache-hit ratio was computed against the pre-raise ¥0.35 rate.
Why Each One Wins
Intelligence & Reasoning Ceiling
- Step 5 Preview: Scores 44 on the Artificial Analysis composite index, reported as top-3 among open models globally, with 27B active parameters — a substantially larger per-token compute budget than V4.1 Flash's asymmetric 8B/16B.
- DeepSeek V4.1 Flash: Scores 39 on the same Artificial Analysis index (still above GPT-6 Luna's 37) despite activating only 8B on input and 16B on output — exceptional efficiency per unit of compute.
Cost & Production Economics
- Step 5 Preview: ¥7/M input (cache miss), ¥20/M output, and ¥0.7/M cached input after a 2026-10-14 price raise (cached input doubled from ¥0.35). StepFun also moved to peak/off-peak pricing and weekly credit limits.
- DeepSeek V4.1 Flash: Roughly ¥2/M input and ¥8/M output at peak rates (≈$0.3/M and ≈$1.2/M), with cache-hit input around ¥0.04/M — measured by Sina Finance at about 29% and 40% of Step 5 Preview's input and output rates.
Context Window & Output Length
- Step 5 Preview: 1M-token context with a 64K max output, positioned for long-running agentic sessions rather than very large single responses.
- DeepSeek V4.1 Flash: Also 1M-token context, but up to 384K max output — about 6x Step 5 Preview's ceiling for a single generation, and it supports a thinking mode.
Agentic Coding & Specialized Work
- Step 5 Preview: Explicitly optimized for AI coding, software engineering, professional knowledge work, and finance, and marketed at ~1/8 the per-task cost of Claude Opus 5; it hit #1 on OpenRouter Trending within a day of launch.
- DeepSeek V4.1 Flash: Posts 50.51 on SuperCLUE's Chinese agentic terminal-coding evaluation (V4.1-Flash max) at ¥1.86 per task — a concrete cost-per-task anchor — and beat its own V4 Pro on release.
Multimodal & Ecosystem
- Step 5 Preview: Native text and vision input (image/video). Available via OpenRouter, with time-limited free access through OpenCode Zen, WorkBuddy (Oct 12-25) and agent tools like Cline/Hermes; weights scheduled to open on 2026-10-15.
- DeepSeek V4.1 Flash: Native multimodal visual understanding (text + image in, text out), first-party API and DeepSeek Harness integration, priced at ~1/3 of V4 Pro.
Trade-offs
Choose Step 5 Preview, you give up…
- Roughly 2.5x the output price (¥20/M vs ¥8/M) and ~3.5x the input price (¥7/M vs ¥2/M)
- A 64K max output, which caps long single-pass generation
- Pricing momentum risk: cached input doubled to ¥0.7/M on 2026-10-14, and peak/off-peak billing plus weekly credit limits were introduced
- Preview-stage model: behavior and rates are still moving ahead of the 2026-10-15 weight release
Choose DeepSeek V4.1 Flash, you give up…
- About 5 points lower on the Artificial Analysis index (39 vs 44)
- A much smaller active compute budget (8B in / 16B out vs 27B), which can matter on the hardest multi-step tasks
- Step 5 Preview's finance/professional-work tuning and its ~1/8-of-Opus-5 per-task positioning
Supporting Evidence
2026-10
2026-10
2026-10
2026-10
2026-09
2026-09
2026-09
2026-09
2026-10
2026-09
Limitations & Known Caveats
- Both models launched within ten days of each other (09-10 and 09-20); benchmarks, pricing and production behavior are very new and still moving.
- Step 5 Preview is explicitly a preview-stage model, and its weights were still slated to open-source on 2026-10-15 at time of writing — capability and access terms may change.
- Step 5 Preview's cached-input rate doubled on 2026-10-14 (¥0.35 → ¥0.7/M), and StepFun simultaneously introduced peak/off-peak pricing and weekly credit limits; any cost comparison has a short shelf life.
- The 11% / 29% / 40% ratios come from a single Sina Finance calculation, and the cache-hit ratio was computed against the pre-raise ¥0.35 rate — against the new ¥0.7 rate the cache-hit gap is far wider (~6%).
- The AA index figures (44 and 39) are vendor- and press-reported, not from an independent benchmark run reproduced here.
- Step 5 Preview's coding claims rest largely on vendor positioning (~1/8 of Opus 5 per task) and OpenRouter traction; DeepSeek's coding anchor (SuperCLUE 50.51 @ ¥1.86/task) is the more directly measured of the two.
- No independent side-by-side reproduction was performed for this page; figures are compiled from published sources.
Alternatives
DeepSeek V4 Pro · GLM-5.3-Flash · Qwen3.8-Flash · Kimi K3 · Claude Opus 5
Bottom Line
Which should you choose?
- Your workload is sustained, complex agentic coding or finance/professional analysis where the intelligence ceiling matters most.
- You want the higher Artificial Analysis score (44 vs 39) and a 27B active-parameter budget.
- You want in on open weights dropping 2026-10-15, or you can exploit the current free-access windows.
- Your token spend is small enough that ~2.5x output pricing is not the binding constraint.
- Cost per token is a dominant constraint and you run high-volume or long-running calls.
- You need very long single responses — 384K max output vs 64K.
- You want native image input with a settled first-party API and stable pricing (~1/3 of V4 Pro).
- You prefer a model with a concrete, cheaply-measured coding cost anchor (¥1.86/task on SuperCLUE).
Frequently Asked Questions
Which model is smarter?
Step 5 Preview, by the Artificial Analysis composite index — 44 vs 39 — and it activates 27B parameters per token versus V4.1 Flash's asymmetric 8B/16B.
Which is cheaper to run?
DeepSeek V4.1 Flash, substantially: about ¥8/M output vs ¥20/M and about ¥2/M input vs ¥7/M at peak rates — roughly 40% and 29% of Step 5 Preview's rates.
Which handles longer outputs?
DeepSeek V4.1 Flash — up to 384K output tokens vs Step 5 Preview's 64K, though both accept 1M tokens of input context.
Are both open weight?
DeepSeek V4.1 Flash is already open-weight. Step 5 Preview's weights were scheduled to be open-sourced on 2026-10-15, after this comparison was compiled.
Do both accept images?
Yes. Step 5 Preview takes native text plus vision input (image/video); DeepSeek V4.1 Flash accepts text and image input with native visual understanding.
Did Step 5 Preview just get more expensive?
Yes. Effective 2026-10-14 00:30, cached-input pricing doubled from ¥0.35/M to ¥0.7/M, with peak/off-peak pricing and weekly credit limits also introduced. Uncached input (¥7/M) and output (¥20/M) were unchanged.
How Airdix Makes Decisions
Independence guarantee: recommendations are based only on criteria and evidence. A product can pay for visibility, but never for a recommendation. Every claim is traced to a source and marked with a verification date. Recommendation ≠ Paid Placement
Compares StepFun's Step 5 Preview and DeepSeek's V4.1 Flash — two September 2026 Chinese flagship-class releases — on intelligence, cost, context/output limits, agentic coding and openness, using launch coverage, official sources, pricing trackers and a direct price comparison.
Architecture, context and pricing from Baidu Baike entries, DeepSeek's official site and StepFun release coverage; cross-checked against jumohub and OrcaRouter pricing trackers; the headline cost ratios come from Sina Finance's 2026-10-13 comparison.
Step 5 Preview released 2026-09-20; DeepSeek V4.1 Flash released 2026-09-10. No independent side-by-side reproduction was performed here.
Artificial Analysis Intelligence Index · SuperCLUE agentic terminal coding (zh) · OpenRouter Trending · vendor per-task cost positioning
Run the same agentic-coding task and a long-form generation task through both APIs, then compare pass rate / quality against your actual token bill at your own cache-hit ratio.
Very recent releases with moving pricing; preview-stage model on the StepFun side; key ratios are single-source (Sina Finance); AA index scores are vendor/press-reported.
OCT 2026 · content quality review