Step 5 Preview Review 2026: StepFun's 600B MoE Flagship Activates Only 27B — Top-Three Open Scores at $1/$2.70
Step 5 Preview Review 2026: StepFun's 600B MoE Flagship Activates Only 27B — Top-Three Open Scores at $1/$2.70
Verdict first: Step 5 Preview (released September 20, 2026) is StepFun's argument that the interesting frontier is no longer raw scale but how much usable intelligence you can wring out of each unit of compute. It is a sparse Mixture-of-Experts flagship with 600B total parameters that activate only ~27B per token, built from 92 transformer layers in a deliberately "narrow and deep" design, with a 1M-token context window and native text-plus-vision input. StepFun reports an Artificial Analysis Intelligence Index of 44, which it says places the model in the global top three open models, plus DeepSWE v1.1 at 67.7% and StepCodeBench at 49.0%. The headline that will travel furthest is cost: $1.00 per million input / $2.70 per million output with cache reads at $0.05 — about a 95% cache discount — and roughly $0.71 per Artificial Analysis task, which StepFun describes as one eighth the cost of Claude Opus 5. The caveats matter as much as the claims: every benchmark is vendor-reported, two of the marquee results (ALE-CLI and FrontierFinance) come with rankings but no published scores, the model is unusually verbose (about 160M output tokens per eval against a ~92M median), and full weights do not arrive until October 15. For teams running long-horizon agents with a real budget line, this is one of the most interesting price-performance entries of the quarter — read with the arithmetic in hand.
Step 5 Preview — 600B total parameters, 27B active, 92 layers, 1M context: frontier-adjacent scores at $1/$2.70 per 1M tokens
What Is Step 5 Preview?
Step 5 Preview is the new flagship base model from StepFun — the Shanghai AI lab founded by former Microsoft global vice president Jiang Daxin, known in China for its consumer assistant Yuewen and its open platform at platform.stepfun.com. It was released and made fully available through the API on September 20, 2026 (Artificial Analysis dates the release to September 18), and it is the third step in a deliberate line: Step 3.5 Flash → Step 3.7 Flash → Step 5 Preview.
That lineage is the point. Where the Flash models were efficiency plays — Step 3.7 Flash famously activated around 11B parameters and raced to the top of throughput leaderboards — Step 5 Preview is the first time StepFun has pushed that same efficiency philosophy into a genuine flagship tier. The brief is not "be the biggest." It is "convert compute into intelligence efficiently enough that frontier-adjacent capability becomes something you can afford to run all day."
The target workloads are explicit: AI coding, software engineering, professional knowledge work and finance — long-horizon agentic tasks rather than chat. StepFun describes it as a "main workhorse model" for real-world agentic jobs. There is no voice or video input here, and no speech or image generation: it is a text-and-vision-in, text-out reasoning and orchestration engine.
The Architecture: Why 92 Narrow Layers
The structural choice worth understanding is depth over width. Step 5 Preview stacks 92 transformer layers in what StepFun calls a "narrow and deep" configuration. The stated rationale is specific: depth gives information a longer propagation path during long prefills, which matters for the implicit multi-hop reasoning that shows up when a model has to hold a large codebase or document set in context and connect facts across it.
The sparsity ratio is the other half of the design. With 600B total parameters and only ~27B activated per token, the model carries the knowledge capacity of a very large network while paying the compute bill of a mid-sized one on every forward pass. That gap is the entire economic story of this release — it is why $1/$2.70 pricing is possible at this capability level at all.
On attention, Step 5 Preview uses Sparse GQA combined with Block-wise Token Merging. StepFun says this cuts the cost of the indexer and top-k selection to one eighth of the original — a meaningful number once you are running 1M-token contexts, because attention bookkeeping rather than raw FLOPs is often what makes long context expensive in practice.
Specs and Availability
| Spec | Step 5 Preview |
|---|---|
| Released | September 20, 2026 (AA lists Sep 18) |
| Architecture | Sparse Mixture-of-Experts |
| Parameters | 600B total / ~27B activated per token |
| Depth | 92 transformer layers ("narrow and deep") |
| Context window | 1M tokens |
| Input modalities | Text + image (native vision) |
| Output modality | Text only |
| Attention | Sparse GQA + Block-wise Token Merging |
| Output speed | ~100 tokens/sec (vendor figure) |
| Model ID | step-5-preview |
| Endpoints | platform.stepfun.com (CN) · platform.stepfun.ai (intl) · studio.stepfun.com (playground) |
| API protocols | /v1/chat/completions, /v1/responses, /v1/messages (Anthropic-compatible) |
| Open weights | Scheduled for October 15, 2026 |
The Anthropic-compatible /v1/messages endpoint is a small detail with outsized practical value: existing agent harnesses built against Claude's protocol — Claude Code, and a growing set of Codex-adjacent and OpenClaw-style tools — can point at step-5-preview with a base-URL and key change rather than a rewrite.
Pricing: The Cache Discount Is the Real Story
| Item | International (USD, per 1M tokens) | Domestic (CNY, per 1M tokens) |
|---|---|---|
| Input (cache miss) | $1.00 | ¥7 |
| Input (cache hit) | $0.05 (~95% off) | ¥0.35 |
| Output | $2.70 | ¥18–20 (sources disagree) |
| Web search (Linkup) | ~$0.013 per call | — |
The $0.05 cache-read price — roughly a 95% discount — is the number that actually changes agent economics. A long-horizon agent loop re-reads the same large context on every turn; at a 1M-token window, the cached input price rather than the headline input price is what you pay most of the time. StepFun pricing that at a twentieth of the miss rate is a deliberate bet that long-running agents, not one-shot chat, are the workload that matters.
There is also a subscription tier, Step Plan, where 1M credits ≈ ¥1 and credits are then converted at model rates:
| Tier | Monthly | Credits |
|---|---|---|
| Flash Mini | ¥49 | 400M |
| Flash Plus | ¥99 | 1600M |
| Flash Pro | ¥199 | 8000M |
| Flash Max | ¥699 | 40000M |
StepFun also runs new-user promotions (a ¥99 plan credit on signup, plus stacking first-call and referral benefits). As always, confirm the live numbers against the official pricing page — domestic output pricing is reported as both ¥18 and ¥20 depending on the source.
Benchmarks: Top-Three Open, All Self-Reported
| Dimension | Result |
|---|---|
| General intelligence | Artificial Analysis Intelligence Index: 44 — vendor claims global top three open models |
| Software engineering | DeepSWE v1.1: 67.7% |
| Coding | StepCodeBench: 49.0% |
| CLI agent tasks | ALE-CLI subset — second only to GPT-6 Astra or Claude Opus 5, leading other open models (no score published) |
| Finance | FrontierFinance — same placement (no score published) |
| Cost efficiency | ~$0.71 per AA Intelligence Index task; vendor says ~1/8 of Claude Opus 5 |
The pattern to notice: the two results with the most impressive framing — ALE-CLI and FrontierFinance — are given only as relative placements with no numeric scores. "Second only to GPT-6 Astra or Claude Opus 5, ahead of other open models" is a meaningful claim, but it is not a reproducible number, and it depends entirely on which models were included in the comparison set. Treat it as directional.
The general caveat: every figure above is StepFun's own, not independently reproduced. There are no third-party leaderboard entries, no community replication runs, and no production cost reports at publication time. The architecture and the price are checkable facts; the rankings are marketing until someone reproduces them.
Reading the "1/8 of Opus 5" Claim Carefully
The claim that travels fastest deserves the most scrutiny, and three qualifications change how it should be read.
- The comparison point is Opus 5's most expensive configuration. Claude Opus 5 at its highest reasoning tier is cited at roughly $2.03 per task. Comparing a model's standard operating setup against a rival's top-tier reasoning setting is a favorable framing — it is not the same as beating the rival's default.
- The metric is a disclosed-poorly average. "$0.71 per task" is a task-level average on the Artificial Analysis Intelligence Index. StepFun has not published the task mix, the token accounting method, or how reasoning tokens are attributed — all of which materially move the number.
- The model is unusually verbose, and that eats the advantage. The evaluation run produced roughly 160 million output tokens, against a median of about 92 million for similarly priced models — roughly 74% more output than peers. At $2.70 per million output tokens, verbosity is not a footnote; it is the difference between the headline price and what you actually pay. Low unit price plus long reasoning partially cancels out.
The fair reading: against Kimi K3 Max at roughly $2 per task, Step 5 Preview's ~$0.71 is a genuine and substantial value advantage — roughly a 2.8× gap on the same metric. The "1/8" figure is a specific, favorable framing rather than a general truth, but it sits on top of a real underlying advantage. Budget on the $0.71-order figure, not the "1/8" headline, and measure on your own task mix before committing.
The Engineering Underneath
The infrastructure work is less quotable than the benchmark table but arguably more predictive of whether the model is pleasant to run in production.
- Long-horizon RL speedup of over 3× end-to-end — meaning post-training on long agentic trajectories, the exact regime this model targets, got dramatically cheaper to iterate on.
- Bit-wise alignment between training and inference with end-to-end overhead under 10%, and asynchronous training logprob deviation held within 1e-2. In plain terms: the numbers you train against are the numbers you serve, which is a common and painful source of silent quality loss at this scale.
- MTP-3 speculative decoding, FP8 MoE, FP8 KV quantisation-aware training, cross-machine expert parallelism, KV cache offloading, and load-aware scheduling — a full stack aimed at serving a 600B MoE economically.
- Sample loss below 1% across training, which at these scales is the difference between a stable run and a wasted cluster allocation.
- Self-evolution: StepFun says it used the model to build its own iterative training data, constructing a million-scale set of verifiable hard tasks. This is the same "model improving model" loop the lab has been signalling across its line.
The Honest Take
Score: 8.0/10 for its category (cost-efficient long-context agentic flagship). As a bid for the absolute capability crown it is not the point of this release, and should not be graded as one.
What genuinely impresses:
- The sparsity economics. 600B total with 27B active is what makes $1/$2.70 possible at this capability tier — the architecture and the price sheet are the same decision.
- The cache price. $0.05 per million cache reads (~95% off) is the clearest signal that StepFun designed this for long-running agent loops rather than chat, and it is where the real savings live.
- 1M context with native vision input and a real attention-cost reduction (indexer and top-k selection at 1/8 the cost) — long context that is not just nominally supported but economically usable.
- Agentic-targeted evaluation. Leading other open models on ALE-CLI and FrontierFinance, even without published scores, points at a deliberate agentic focus rather than benchmark-shaped tuning.
- Protocol compatibility — the Anthropic-compatible
/v1/messagesendpoint makes migration for existing harnesses close to trivial. - A dated open-weights commitment (October 15), which is more than most closed competitors offer at all.
What should temper your excitement:
- Verbosity is a cost, not a quirk. ~160M output tokens per eval against a ~92M median means the effective price is materially higher than the rate card suggests.
- Every benchmark is self-reported, and the two strongest results ship as rankings without numbers. That is a meaningful transparency gap.
- "1/8 the cost of Opus 5" is a favorable framing — highest reasoning tier as the baseline, undisclosed task mix, task-level average.
- No weights until October 15. Private deployment, fine-tuning and on-prem evaluation are all off the table today, and the hardware footprint has not been published — so teams that need to plan GPU capacity cannot.
- Modality ceiling. Text and image in, text out. No audio, no video, no generation of any kind. For multimodal pipelines this is one component, not a platform.
- ~100 tokens/sec output is well below the Flash line's headline throughput, so this is not the model to pick if raw speed is the constraint.
- The generational comparison is not clean. Step 3.7 Flash was an 11B-active efficiency model and Step 5 Preview is a 600B flagship — calling this simply "an upgrade" overstates comparability.
Who should use it: teams running long-horizon agents, large-codebase or large-document workloads, and finance or professional-knowledge pipelines where cost per completed task is the real budget line and the 95% cache discount applies directly. Also anyone already on Claude-protocol harnesses who wants a cheaper backend with a base-URL change.
Who should wait: anyone needing open weights or on-prem deployment (October 15), anyone requiring audio or video input, anyone for whom output verbosity would blow the budget, and anyone making procurement decisions on independently verified benchmarks.
FAQ
What is Step 5 Preview?
StepFun's flagship base model, released September 20, 2026. Sparse MoE with 600B total parameters and ~27B activated per token, 92 transformer layers in a "narrow and deep" design, 1M-token context, native text and vision input with text output. It is aimed at real-world agentic work — AI coding, software engineering, professional knowledge tasks and finance — and full weights are scheduled for October 15, 2026.
How much does Step 5 Preview cost?
$1.00 per million input tokens and $2.70 per million output tokens, with cache reads at $0.05 (~95% off) and web search around $0.013 per call. Domestic pricing is reported at ¥7 per million input (¥0.35 cached) and ¥18–20 per million output, with sources disagreeing on the output rate. Step Plan subscriptions run from ¥49 to ¥699 per month. StepFun puts the average cost per Artificial Analysis task at about $0.71.
Is Step 5 Preview open source?
Not yet — StepFun has committed to releasing full weights on October 15, 2026. Until then it is API-only via platform.stepfun.com (domestic) and platform.stepfun.ai (international), with the model ID step-5-preview. Private deployment is not available and the hardware footprint has not been published.
Can I use Step 5 Preview with Claude Code or other agent harnesses?
Likely yes, with minimal changes. In addition to the OpenAI-style /v1/chat/completions and /v1/responses endpoints, StepFun exposes an Anthropic-compatible /v1/messages endpoint — so harnesses built against Claude's protocol can generally be repointed with a base URL and API key change rather than a rewrite.
Is the "1/8 the cost of Claude Opus 5" claim reliable?
Partly. The baseline is Opus 5's highest reasoning tier at roughly $2.03 per task rather than its default, the cost metric is a task-level average with no disclosed task mix, and Step 5 Preview's verbosity (~160M output tokens per eval vs a ~92M median) partially cancels the low unit price. Against Kimi K3 Max at roughly $2 per task, the ~$0.71 figure represents a real value advantage — but budget on the order of $0.71 per task rather than treating "1/8" as a general rule.
Does Step 5 Preview handle audio or video?
No. Input is text and images; output is text. There is no audio or video input, and no speech, music or image generation. For multimodal pipelines beyond vision, it covers one part of the stack.
How do Step 5 Preview's benchmarks compare to GPT-6 Astra or Claude Opus 5?
On the Agents' Last Exam CLI subset and FrontierFinance, StepFun describes the model as second only to GPT-6 Astra or Claude Opus 5 while leading other open models — but no numeric scores were published for either, so the comparison cannot be checked. Its published numbers are the AA Intelligence Index score of 44, DeepSWE v1.1 at 67.7% and StepCodeBench at 49.0%, all self-reported.
Sources
- PingWest — StepFun releases Step 5 Preview flagship; per-task cost 1/8 of Claude Opus 5 (Sep 20, 2026)
- Tencent News — Step 5 Preview: 600B MoE aimed at long-horizon agents, open-sourcing in October (Sep 20, 2026)
- GeekPark — StepFun Step 5 Preview targets AI coding, finance and professional knowledge work
- Zhidx — A new Chinese MoE flagship enters the global open top three, priced against DeepSeek-V4-Pro
- Vercel AI Gateway — Step 5 Preview API, pricing and playground
- ifeng Tech — StepFun's Step 5 Preview enters the global open-source top three
All benchmark figures are StepFun's own and have not been independently reproduced as of the publication date. Prices and specifications change frequently; verify against the official StepFun open platform documentation before making purchasing decisions.
评论
发表评论