Step 5 Preview vs DeepSeek V4.1 Pro：このうちの1つは製品ではない

You cannot call D​eepSeek V4.1 Pro. Not on a slow day, not at a higher price, not with a waitlist — there is no endpoint, no model card and no checkpoint anywhere, and there has not been since D​eepSeek named it. S​tep 5 Preview, by contrast, was released by StepFun on September 20, 2026, has been downloadable in three precisions since the morning of release, and has a published rate card, a million-token context window and one independent benchmark score. A comparison table between these two would be a comparison between a model and a name, so this piece is not that. It is the honest version: which of the two you can build on today, what the D​eepSeek tier was supposed to be, and what to actually route in its place while it does not exist.

The reason this matchup gets searched at all is that the two occupy the same slot on paper. Both are advertised as flagship-class mixture-of-experts systems for hard agentic and reasoning work, both quote a million-token context, and both sit in families where the smaller sibling is the one that actually shipped. The difference is what each vendor did after naming its flagship: StepFun published weights, a serving recipe and a price. D​eepSeek published a sentence, then withdrew it.

What D​eepSeek actually said, and what it did next

The single public appearance of the D​eepSeek V4.1 Pro name is in the September 10, 2026 release note for DeepSeek V4.1 Flash — the 552-billion-parameter multimodal model that is the V4.1 line's smallest member. In that note D​eepSeek wrote that from 04:00 UTC on September 14 all requests to deepseek-v4-pro/ would route to V4.1 Flash and be billed at Flash rates, and that this would continue "until V4.1-Pro launches." Developers running production traffic on the pinned Pro name objected to being silently moved onto a different model, and the plan was withdrawn. Nothing has been published about the Pro tier since.

So the correct reading of "D​eepSeek V4.1 Pro" today is: an announced intent, contingent on a launch date that has not been given, in a family where the Pro predecessor is a live, callable, separately-priced model. That predecessor is what the name is standing in for, and for anyone choosing an opponent this week it is the only real one on the D​eepSeek side of the fence.

実際に実行できる比較

Here is S​tep 5 Preview against the callable D​eepSeek flagship, with every row sourced. The D​eepSeek column is DeepSeek V4 Pro, not V4.1 Pro, and that substitution is the whole point of the table.

• Price per million tokens — S​tep 5 Preview $1.00 input, $0.10 cache-hit input, $2.70 output, flat around the clock vs DeepSeek V4 Pro $0.66 input, $0.022 cache-hit input, $1.98 output off-peak, doubling to $1.32 / $0.044 / $3.96 during D​eepSeek's peak hours

• Context window — S​tep 5 Preview 1M tokens vs DeepSeek V4 Pro 1,048,576 tokens

• Maximum output — S​tep 5 Preview 64k tokens in StepFun's documentation, though the model card's own specification table says "up to 1M" vs DeepSeek V4 Pro 384,000 tokens

• Architecture — S​tep 5 Preview sparse MoE, 600B total and 27B active, 352 experts with 8 selected per token across 88 of its 92 layers vs DeepSeek V4 Pro MoE, 1.6 trillion total and 49 billion active

• Input modality — S​tep 5 Preview text, images and video vs DeepSeek V4 Pro text only

• Independent score — S​tep 5 Preview 44 on the Artificial Analysis Intelligence Index, at rank 41 of 219 when that page was captured vs DeepSeek V4 Pro no public equivalent row on the same board

• Median time to first token on OrcaRouter — S​tep 5 Preview not routed, so no row vs DeepSeek V4 Pro 1.89 seconds

• Weights — S​tep 5 Preview FP8, BF16 and NVFP4 repositories, all under StepFun's own account, ungated vs DeepSeek V4 Pro open weights on Hugging Face

出力上限の行には、それだけで一文を割く価値があります。なぜなら、そこは両ベンダー自身の資料が自らと矛盾している唯一の箇所だからです。StepFunのドキュメントには、最大出力64kと書かれています。同じリリースのStepFunのモデルカードは、仕様表に「最大100万トークン」と記載しています。どちらもStepFunの文書であり、どちらも現行のもので、その違いはドキュメント上の制限とアーキテクチャ上の制限の違いです。誰かが200kトークンの補完を投稿するまでは、64kをホスト型APIから得られる数値として扱ってください。

なぜ価格差は見た目よりも小さいのか

DeepSeek V4 Pro is the cheaper model on both ends of the rate card — 34% less on input, 27% less on output — and its cache-hit rate, two cents per million tokens, is an order of magnitude below S​tep 5 Preview's ten. That is a real and large advantage for workloads with a stable prefix: a retrieval pipeline that re-sends the same 300,000-token document set on every call pays $6.60 per million calls on D​eepSeek and $30.00 on StepFun, and nothing else in the comparison closes that gap. The comparison assumes D​eepSeek's off-peak rate, which is the one its documentation leads with; that rate applies outside two windows, 01:00–04:00 and 06:00–10:00 UTC Monday to Friday, and inside them every line on the D​eepSeek card doubles. A workload that runs on United States business hours spends a meaningful share of its week in the peak window, where DeepSeek V4 Pro costs $1.32 input, $0.044 cached and $3.96 output — still cheaper than StepFun on input and now more expensive on cache reads, against a S​tep 5 Preview rate that does not move with the clock.

What closes some of it is verbosity. On the same evaluation suite, S​tep 5 Preview generated 160 million output tokens against a suite median of 85 million — the evaluator's own note calls it "very verbose" — and it does not publish an equivalent token count for DeepSeek V4 Pro, so the honest position is that the cache advantage is measurable and the output-token penalty is measurable on one side only. If your prompts are long, stable and cache-heavy, DeepSeek V4 Pro is the cheaper purchase and the gap is wide. If they are short and highly variable, most of the input is billed as cache misses and the two rate cards land within a third of each other.

The modality split is the other thing price does not capture. DeepSeek V4 Pro is text-only. S​tep 5 Preview reads images and video, and that is not a feature you can price against a text-only model — it is a workflow you either have or you do not. Screenshot triage, chart extraction and video summarisation are out of scope on the D​eepSeek flagship and in scope here, and no discount on input tokens buys them back.

存在しないモデルを回避するルーティング

この対決が提起する実用的な問いは、使う予定だったモデルに名前はあるがエンドポイントがないとき、どうするかというものだ。クライアント内の定数を向け直すのは簡単な部分だ。より難しいのは、決して出荷されなかったティアへのピン留めが、最悪のタイミングで同じ代替の判断を自分自身で下さざるを得ない状況を残すということであり、ここでルーティング層がその存在意義を発揮する。

DeepSeek V4 Pro is on OrcaRouter, and so is the model that replaced it in D​eepSeek's own lineup: the catalogue carries both deepseek/deepseek-v4-pro/ and deepseek/deepseek-v4.1-flash/, which means an A/B against the successor D​eepSeek itself recommends is a change to one string in the model parameter rather than a second contract, a second SDK and a second bill. Requests go to the same O​penAI-compatible endpoint either way, at 0% markup on the provider's list price — $0.66 and $1.98 per million for V4 Pro is exactly D​eepSeek's own rate, passed through untouched, so a vendor price change reaches you the day it happens rather than on the next billing cycle.

That covers the D​eepSeek side of the question and not the StepFun side, because S​tep 5 Preview is not on OrcaRouter's live model list — it is not routed here, and the way to call it is StepFun's own API. What the router gives you in that arrangement is a fallback floor: one endpoint, one key and one bill that carries the models you can call today, against a vendor endpoint you are integrating for the one you cannot get anywhere else. Automatic failover matters more than usual on a model this new, because a primary provider going down at 3am should not mean a production path going down with it.

これらのそれぞれが本当は何のためにあるのか

DeepSeek V4 Pro は、ワークロードがテキストで、プロンプトプレフィックスが安定しており、キャッシュヒット率が高く、384,000トークンの出力上限が重要な場合の選択肢です — 入力のほとんどが毎回同じで、モデルには書き換えではなく拡張が求められる、長コンテキストの文書・コード作業です。タスクのどこにも画像が必要ない場合も、同様に選択肢となります。そのようなワークロードでは、34%の入力割引と2セントのキャッシュレートが決定的であり、V4.1 Proの後継が存在しないことは、固定された、安定した、すでに出荷済みのモデルを支持する論拠であって、反対する論拠ではありません。

S​tep 5 Preview is the pick when the task is agentic and multimodal: a repository-scale coding session with screenshots in the loop, a research run over mixed documents and video, or a finance workflow that has to read the tables in a filing rather than the text around them. StepFun built this model for sustained tool use, published a serving recipe that handles its reasoning and tool-call formats properly, and put the weights out so that a data-residency requirement does not rule it out. It costs roughly half again what DeepSeek V4 Pro costs on input and a third more on output, and on a cache-heavy text workload that premium is hard to justify. On a workload with images in it there is no comparison to make, because one of the two models cannot see.

And D​eepSeek V4.1 Pro is not a pick at all. When it ships it will presumably be the largest model in the V4.1 family and will be priced above V4.1 Flash; D​eepSeek has said so implicitly by naming a Pro tier at all. Nothing about it is knowable today beyond the one sentence from September 10, and anyone publishing specs for it is publishing an inference. If you have capacity plans that assume it, the safest assumption available is that it arrives later than you hope and costs more than you budgeted.

短いバージョン

Two models get compared here and only one exists. S​tep 5 Preview was released on September 20, 2026, is callable on StepFun's platform at $1.00 and $2.70, reads text, images and video, and has an open-weights story that turned out to be true all along. D​eepSeek V4.1 Pro is a tier name in a release note that D​eepSeek has not followed with a model card, an endpoint or a date. For a text workload with a stable prefix, the model to compare S​tep 5 Preview against is DeepSeek V4 Pro, which is cheaper on every line, callable today, and available alongside the rest of the catalogue on one key. For anything with an image in it, there is no D​eepSeek comparison to run.

数字からは分からないのは、それが現場で何か問題を起こしていないかどうかだ。このクラスのモデルに関する、記録されたAIインシデントやエージェントの障害は、OrcaRouter AIインシデントアーカイブに集められており、各項目にはそれぞれの情報源、深刻度、信頼度グレード、異議ありフラグが付されている。