Step 5 Preview vs DeepSeek V4.1 Flash: Six Times the Input Price for Four and a Half Index Points

Step 5 Preview costs $1.00 per million input tokens. DeepSeek V4.1 Flash costs $0.15. On output the gap is $2.70 against $0.60, and on cache hits it is $0.10 against $0.003 — a thirty-three-fold spread on the line that dominates a long agent loop. For that money Step 5 Preview scores 44 on Artificial Analysis's Intelligence Index against DeepSeek V4.1 Flash's 39.46, and finishes an index task at roughly $0.71 against DeepSeek's $0.27. Six and a bit times the input price, two and a bit times the task cost, four and a half points. That is the arithmetic, and it is not a favourable one for StepFun's model. The reason to keep reading is that the two models differ on three lines that have nothing to do with the price, and two of those lines are the ones that would get a model rejected in a procurement review. None of this is a reason to avoid the model, only a reason to know the history: the documented AI incidents and agent failures behind models in this class are collected in the OrcaRouter AI Incident Archive where each entry carries its own sources, severity, confidence grade and disputed flag.

Both are Chinese sparse mixture-of-experts models with million-token windows and open weights, released eleven weeks apart — DeepSeek V4.1 Flash on September 10, 2026 and Step 5 Preview on September 20, 2026. They are the closest pair in this model's entire comparison set, which makes the differences that survive the parallelism the interesting ones. They are the licence, the output ceiling, and video.

Where these two actually part company

• Input price — Step 5 Preview $1.00 per million tokens on StepFun's platform vs DeepSeek V4.1 Flash $0.15 off-peak, doubling to $0.30 in DeepSeek's weekday peak windows

• Output price — $2.70 vs $0.60 off-peak, $1.20 at peak

• Cache-hit input — Step 5 Preview $0.10 per million, a 95% discount, vs DeepSeek V4.1 Flash $0.003 off-peak and $0.006 at peak, a 98% discount, and the widest ratio anywhere on either card

• Time-of-day pricing — none on StepFun vs DeepSeek doubles all three meters between 01:00 and 04:00 and again between 06:00 and 10:00 UTC on weekdays, excluding Chinese public holidays

• Context — 1M tokens for both; DeepSeek publishes the figure as 1,048,576

• Maximum output — Step 5 Preview 64k tokens per StepFun's documentation vs DeepSeek V4.1 Flash 384,000 tokens, six times the ceiling

• Modality — Step 5 Preview takes text, images and video vs DeepSeek V4.1 Flash text and images, with no video path

• Architecture — Step 5 Preview 600B total, 27B active vs DeepSeek V4.1 Flash 552B total, 16B active, so the cheaper model also activates 41% fewer parameters per token

• Licence — StepFun Model License 1.0, a custom grant with attribution and separate-commercial-licence thresholds vs DeepSeek V4.1 Flash under MIT, an OSI-approved licence with no threshold to read

• Independent score — 44 for Step 5 Preview against 39.46 for DeepSeek V4.1 Flash at maximum effort, both on Intelligence Index v4.3.2

• Independent cost per completed index task — roughly $0.71 vs $0.27

Read the last two rows as a pair and the per-point arithmetic falls out: about 1.6 cents per index point for Step 5 Preview against about 0.7 cents for DeepSeek V4.1 Flash. The cheaper model is also the cheaper model per unit of measured capability, which is unusual — discount models usually buy their score with verbosity or reasoning tokens. DeepSeek V4.1 Flash emitted 88,574 output tokens per index task, well inside the field's norms. Step 5 Preview emitted roughly 160 million output tokens across the full suite and drew a "very verbose" label from the evaluator. StepFun's model is not buying its four and a half points with efficiency.

The peak-hour doubling is a smaller problem than it sounds, and the output ceiling is a bigger one

DeepSeek's peak pricing gets discussed as though it were a trap, and it mostly is not. It applies between 01:00 and 04:00 and again between 06:00 and 10:00 UTC on weekdays, which is early morning in Europe and the small hours in the United States, and it excludes weekends and Chinese public holidays entirely. A batch job can be moved. A European team whose traffic peaks at 08:00 UTC pays double for its busiest hours, which is the real cost of the arrangement, but even doubled DeepSeek is cheaper per token than Step 5 Preview at its flat rate. The structural point is different: StepFun's card has one column and DeepSeek's has two, and a budget built on one column is a budget that will be wrong on somebody's real traffic mix. The honest framing is that DeepSeek's peak window is a scheduling problem and StepFun's flat rate is a simplicity feature, and neither of those is worth several times the input price on its own.

The output ceiling is the line that actually breaks things. Step 5 Preview's documentation caps a single response at 64k tokens. DeepSeek V4.1 Flash allows 384,000 — six times as much — and the practical consequence is not about long prose. It is about agent runs whose final artefact is a large file, a full-file refactor, a generated schema or a long structured extraction. A 64k ceiling means those tasks have to be decomposed into multiple calls with an orchestrated hand-off between them, which is engineering work you would otherwise not do. This is the one row on the card where the cheaper model is buying you something you cannot get from the more expensive one at any price, and it is the reason a team with that workload should not be reading the rate card at all.

Video input is real, and it is worth less than it sounds

Step 5 Preview accepts video, and DeepSeek V4.1 Flash does not. That is a genuine capability difference and it is the strongest single argument for StepFun's model in this comparison. StepFun documents the accepted path in detail: MP4, QuickTime and Matroska containers, passed as a URL, as Base64, or through a Files API reference, with URL-based MP4s under 128MB and a recommended length under five minutes. Images are capped at 60 per request, and both come with low and high detail settings.

What that buys you is one model instead of a pipeline. A workflow that ingests a screen recording, a lecture capture or a dashcam clip and then reasons over what happened needs either a video model or a frame-sampling harness you built yourself; a model that takes the video directly collapses that into a single call. What it does not buy you is a general video understanding capability you can plan a product around, because nobody has published an independent evaluation of Step 5 Preview's video comprehension. The capability is documented by the vendor and unscored by anyone else, and the four-and-a-half-point index gap it is attached to is entirely text and image. Treat the video row as a reason to test, not as a number.

The licence line is where the cheap model is the safe choice

This is the part of the comparison a price table cannot show, and it reverses the usual reading of "you get what you pay for." DeepSeek V4.1 Flash ships under MIT — the OSI-approved licence with no attribution trigger, no revenue threshold, no separate commercial agreement for large operators, and no ambiguity about derivative works. The weights are on Hugging Face, 552 billion parameters total with 16 billion active, and a legal team asked to approve a product built on them has a short conversation.

Step 5 Preview ships under StepFun Model License 1.0, which is a real open-weights grant and is not an open-source licence. It is perpetual, worldwide, royalty-free and includes a patent grant, and it permits commercial use, fine-tuning and derivative works. Two conditions attach. A commercial product or service incorporating the model that passes 100 million monthly active users or $20 million in revenue in any calendar month must display "Powered by StepFun AI" prominently in the product interface and in the user agreement. And a hardware manufacturer or brand owner shipping the model's capabilities in devices, or a model-as-a-service business whose combined revenue exceeds $20 million over any twelve months, needs a separate commercial licence. Internal use that does not expose the model or its capabilities to a third party is squarely inside the grant.

For a startup fine-tuning Step 5 Preview on internal document work, that is fine and the licence costs nothing. For a company that intends to ship a hosted product, it is a threshold to model, and the difference between "we may need to negotiate in eighteen months" and "we never have to think about this" is worth more than a few multiples on a token price to a lot of organisations. StepFun's licence is the same shape other open-weight licences use and it is not an unreasonable document. It is simply a document, where DeepSeek's is a standard.

What each of them is for

DeepSeek V4.1 Flash is the correct default for most workloads that this comparison covers. It is cheaper on every meter, cheaper per completed task, cheaper per index point, more permissive in licence, more than twice as generous on the input window, and it runs on 16 billion active parameters against StepFun's 27 billion, which matters if self-hosting is on the roadmap because the smallest machine that can serve it is smaller. Its disadvantages are a 39.46 index score against 44 and the absence of video. Its peak-window doubling is a scheduling detail, not a structural one.

Step 5 Preview earns its premium in three specific situations and loses everywhere else. You need four and a half more index points and the marginal capability at the top of the board is worth two and a half times the task cost to you. You need video input in the same call as the reasoning. Or you need a flat, single-column rate card with no time-of-day window and no scheduling logic in the pipeline, which is a real operational simplification for a team that would otherwise have to build the scheduler. None of those is a bad reason. All of them are narrower than the rate card makes them look.

There is an undramatic middle path worth naming, which is not choosing. Both of these models are cheap enough to run against each other on your own traffic for a week, and the decision you make from your own per-task numbers will be better than the one you make from a borrowed index — particularly because the two models fail differently, and the failure modes are where the real cost lives. That comparison is also where the video row gets settled honestly: if your pipeline has no video in it, the strongest argument for StepFun's model disappears and the rest of the card is a straight loss.

DeepSeek V4.1 Flash is on OrcaRouter as deepseek/deepseek-v4.1-flash at DeepSeek's own published rates with 0% markup, which has a specific consequence for this model: DeepSeek reprices on a peak and off-peak schedule, and a pass-through means the off-peak column is what you pay the moment it applies rather than at the next contract cycle. If you are running Step 5 Preview somewhere else and want a cheap fallback for the majority of your traffic that is not video, that is one routing rule pointing at a model on the same key — and automatic failover means an unfamiliar endpoint never becomes a single point of failure for a production path. Step 5 Preview itself is not in OrcaRouter's live catalogue; call it through StepFun's own platform, or self-host it from the FP8, BF16 or NVFP4 checkpoints the vendor publishes.

The verdict, in one paragraph you can act on

If your workload is text and images, DeepSeek V4.1 Flash is the better model at every price point on the card, and the four-and-a-half-point index deficit is a real but narrow cost. If your workload needs a 384k-token single response, DeepSeek V4.1 Flash is the only one of the two that can do it. If your legal team wants a standard licence rather than a custom one, DeepSeek V4.1 Flash is again the answer. Step 5 Preview is worth the premium only when you specifically need video input, or when four and a half index points at the top of the independent board changes what your product can do — and if that is the case, note that the model is verbose, which means the per-task premium is roughly 2.6 times rather than the 6.7 times the input line suggests. Test both against your own traffic. The rate card tells you which is cheaper; only your trace tells you whether the expensive one is worth it.