DeepSeek V4.1 Flash vs Gemini 3.1 Pro: One of These Is Still a Preview

The most useful thing to know about DeepSeek V4.1 Flash against Gemini 3.1 Pro is not printed on either spec sheet. DeepSeek V4.1 Flash is a generally available model, released on 10 September 2026 with MIT-licensed weights you can download. Gemini 3.1 Pro has been in preview since 19 February 2026 and is still served under a preview build name roughly seven months later. That asymmetry does not decide which model is better, but it decides what kind of commitment you are making when you build on one of them, and it is the frame for everything below.

Both hold a million tokens of context. Both accept images. The differences that actually separate them are the price curve above 200,000 input tokens, the input modalities, the output ceiling, and the fact that one of these two is a file you can own and the other is an API.

Where the two actually stand

• Price per 1M tokens, up to 200K context — DeepSeek V4.1 Flash $0.15 in / $0.60 out vs Gemini 3.1 Pro $2.00 in / $12.00 out. That is 13 times the input price and 20 times the output price.

• Price per 1M tokens, above 200K context — V4.1 Flash unchanged at $0.15 / $0.60 vs Gemini 3.1 Pro $4.00 in / $18.00 out. The input multiple widens to 27× at exactly the point where long-context work lives.

• Cached input — V4.1 Flash $0.003 per 1M off-peak vs Gemini 3.1 Pro $0.40 per 1M. Two orders of magnitude, on the line item that decides whether a re-read context is affordable.

• Context window — 1M tokens vs 1M tokens. Level.

• Maximum output — V4.1 Flash 384K tokens vs Gemini 3.1 Pro 65K tokens. Six times the ceiling.

• Input modalities — V4.1 Flash takes text and images vs Gemini 3.1 Pro takes text, images, audio, video and file uploads.

• Weights — V4.1 Flash MIT, downloadable vs Gemini 3.1 Pro closed, API only.

• Release state — V4.1 Flash generally available since 10 September 2026 vs Gemini 3.1 Pro in preview since 19 February 2026.

• Observed latency to first token — V4.1 Flash 2.63 s at p50 and 9.05 s at p95 vs Gemini 3.1 Pro 10.00 s at p50 and 10.00 s at p95, on OrcaRouter's own 7-day telemetry. The expensive model's median wait is at the telemetry ceiling, and its tail does not move off it.

Read the list without the prices and the shape is familiar from every open-weights-versus-frontier comparison: the downloadable model wins on output ceiling, ties on context, and is ahead on observed latency. The closed model wins on modalities, and on those it wins outright. Nothing here explains a 13× input multiple on its own.

The 200K cliff is the pricing story

Gemini 3.1 Pro's headline rate is not a single rate. Prompts up to 200,000 input tokens bill at $2.00 in and $12.00 out per million; a prompt that crosses that threshold bills at $4.00 and $18.00. DeepSeek V4.1 Flash does not have a tier — it bills $0.15 and $0.60 whether the request is 2,000 tokens or 900,000, with the one caveat that DeepSeek doubles its rate during peak hours and runs entirely off-peak at weekends.

That tier boundary lands in the worst possible place for long-context work, because long-context work is by definition the work that crosses it. A worked comparison makes it concrete. Take a pipeline making 20,000 calls a month, each averaging 250,000 input tokens and 4,000 output tokens — a document-analysis job over large files.

• DeepSeek V4.1 Flash at off-peak rates: 20,000 × 250,000 is 5,000M input tokens at $0.15 per million, or $750.00. Output is 20,000 × 4,000, which is 80M tokens at $0.60 per million, or $48.00. Total $798.00.

• Gemini 3.1 Pro at its above-200K tier: the same 5,000M input tokens at $4.00 per million is $20,000.00, and 80M output tokens at $18.00 per million is $1,440.00. Total $21,440.00.

That is a 27× difference in monthly spend on identical traffic, and it is not the multiple you would have guessed from the headline numbers. The headline multiple is 13×. The multiple you actually pay for long-context work is 27×, because the tier boundary is exactly where your prompts sit.

What the preview label actually costs you

A preview build is a preview build across the industry: it carries no published stability guarantee. The vendor has not committed to keeping the endpoint at that behaviour, that price, or that name. That is not a criticism of Gemini 3.1 Pro — it is what the label means, and it is why the preview suffix has survived seven months rather than being a two-week formality.

For a prototype this is nothing. For a production path it is a specific, quantifiable risk: you are betting that the build you tuned against stays put. The contrast with the other side of this comparison is structural rather than rhetorical. DeepSeek V4.1 Flash is GA, and its weights are MIT-licensed and downloadable, which means that even if the hosted API changed its terms tomorrow, the model itself would still be in your hands. A closed preview model gives you neither the GA commitment nor the escape hatch. If your workload can run on either, that difference is worth more than a benchmark point.

Where Gemini 3.1 Pro is the better instrument

The modality gap is not a rounding error, and it is the one dimension on this list where the cheap model cannot be substituted at any price. Gemini 3.1 Pro takes audio and video as first-class input, plus file uploads. DeepSeek V4.1 Flash takes text and images. If your pipeline ingests a call recording, a screen capture, or a video review, DeepSeek V4.1 Flash is not a cheaper option — it is not an option. The 13× multiple is irrelevant to a task one model can do and the other cannot.

There is a second, quieter case. Gemini 3.1 Pro has been publicly available, in some form, since February, and it is the model with the longer production history — more people have hit its edges, and its behaviour on real traffic is better documented than a twelve-day-old release's is. For output-heavy interactive work where a ten-second median wait is acceptable because the job runs in the background, that track record is the argument, and it is a real one. It is not a latency argument: on the telemetry above, the cheaper model is the faster of the two at both the median and the tail.

And there is the question of what the preview label means for that history. Seven months of availability under a preview build name is longer than most models get, and it is also seven months without a GA commitment. DeepSeek V4.1 Flash is twelve days old at the time of writing and its independent record is thin: the one recent third-party sweep it appears in, the Agents on Rails agentic-coding board published on 21 September 2026, placed it at 17% on maximum effort, mid-field. Twelve days is not enough time for a model's reputation to mean anything, in either direction — which is the honest reason a buyer might prefer the older model here, and it has nothing to do with speed.

Routing on context length, because that is the real decision

The decision this comparison implies is not "pick one." It is a rule with two clauses: long-context, high-volume work goes to DeepSeek V4.1 Flash, and anything with audio or video in it goes to Gemini 3.1 Pro. The awkward part is that this rule is easy to state and annoying to implement, because the two clauses usually live in different vendors with different keys, different SDKs and different rate limits.

Both models are reachable from a single OrcaRouter key, which turns the rule into a routing rule rather than branching code in your application — you can key the decision on request context length directly, which is the dimension that actually drives the cost difference here. OrcaRouter passes provider list pricing through at 0% markup, so the tiered rates above are Google's own and the DeepSeek peak schedule is DeepSeek's own, with a vendor price change live on our side the same day. And because one side of this pairing is a preview build, automatic failover is worth putting behind it: if the build is revised under you, the request lands on the other model rather than on an error page.

That last point is the practical version of everything above. The 27× multiple on long-context work is the reason to route rather than choose, and the preview label on one of the two models is the reason to have somewhere for the request to go when the build changes.

What to watch

• Whether Gemini 3.1 Pro leaves preview. A GA release would remove the stability caveat and change the shape of this comparison more than any price move would.

• Whether the above-200K tier changes. The tier boundary is doing more work in the cost arithmetic than the headline rate is.

• Whether DeepSeek V4.1 Flash accumulates an independent record. A model with MIT weights, a 384K output ceiling and a 1M context at $0.15 per million input is an unusual package; whether the capability is there is a question no public benchmark has answered yet.

Until the first of those lands, the accurate summary is this: DeepSeek V4.1 Flash is roughly thirteen times cheaper on input and twenty-seven times cheaper on long-context input than Gemini 3.1 Pro, it is the only one of the two you can download, and it is the only one of the two with a GA commitment behind it. Gemini 3.1 Pro is the model with the longer production history and is the only one of the two that can read audio and video. Route accordingly.