DeepSeek V4.1 Flash vs DeepSeek V4 Pro: The Handover DeepSeek Scheduled, Then Cancelled

DeepSeek V4.1 Flash shipped on 2026-09-10, and DeepSeek V4 Pro was supposed to be gone by now. The plan, announced two days before the Flash release and firmed up on release day, was that on 2026-09-14 at 12:00 Beijing time every request to deepseek-v4-pro would quietly reroute to the newer, cheaper deepseek-flash and be billed at Flash rates. DeepSeek cancelled that on 2026-09-11 after developers objected, and on 2026-09-14 the deadline passed with V4 Pro still up and its billing unchanged. So this is not a launch comparison — the model on the left is eight days old and the model on the right is a year-old flagship in its fourth month of general availability. It is a comparison of two models whose maker published benchmarks saying the cheap one won, then discovered its users disagreed, and backed down. That sequence is the most useful thing anyone has published about either model this month, because it is the exact decision you are about to make yourself.

What actually happened, in order

The timeline matters more than any single benchmark here, because the reversal is the story and the announcement was deliberately quiet.

• 2026-09-09 — DeepSeek tells users that before V4.1 Pro arrives, V4 Pro requests will be routed to V4.1 Flash and billed at Flash rates. The message is that Flash has comprehensively surpassed Pro on performance, cost, speed and task-completion time.

• 2026-09-10 — DeepSeek V4.1 Flash goes GA on app, web and API. Weights land on Hugging Face under MIT. The API model name becomes deepseek-flash. The previous-generation deepseek-v4-flash and deepseek-v4-flash-vision-exp are retired outright, with their legacy IDs temporarily routed to V4.1 Flash. The Pro shutdown is pushed to 2026-09-14, 12:00 Beijing time (04:00 UTC).

• 2026-09-11 — DeepSeek reverses. In response to user demand, it says, V4 Pro API service will continue after 2026-09-14 with billing unchanged, and the automatic switch to V4.1 Flash is cancelled.

• 2026-09-14 — the deadline passes. V4 Pro is still serving at $0.66 / $1.98 per million tokens off-peak.

The asymmetry in that sequence is the whole article. Flash users were migrated whether they wanted it or not — the old V4 Flash ID now answers with a different model. Pro users got a reprieve, and the reason is not that V4 Pro benchmarks better. It is that production systems had been tuned against it: prompts, agent scaffolds, eval harnesses and reproducibility assumptions that a backend swap invalidates without any code change on the caller's side. DeepSeek's comparison page still lists both models today, side by side, which tells you the company intends to serve both.

Side by side: the two SKUs

Everything below is DeepSeek's own published specification unless a source is named. Prices are per million tokens, off-peak, with the peak rate in parentheses — DeepSeek peaks between 01:00 and 04:00 and again between 06:00 and 10:00 UTC, Monday to Friday, and all other hours are off-peak at half the peak rate.

• API model name — deepseek-flash (DeepSeek-V4.1-Flash) vs deepseek-v4-pro (DeepSeek-V4-Pro-0813).

• Input, cache miss — $0.15 ($0.30) vs $0.66 ($1.32).

• Input, cache hit — $0.003 ($0.006) vs $0.022 ($0.044).

• Output — $0.60 ($1.20) vs $1.98 ($3.96).

• Context / max output — 1M tokens / 384K on both. Identical.

• Image input — supported natively on Flash; listed as not supported on V4 Pro.

• Concurrency limit — 2,500 on Flash against 500 on V4 Pro.

• Reported parameters — Flash: 552B total, vendor figure, with a credible community challenge to it (more on that below). V4 Pro: 1.6T total, ~49B active, which Artificial Analysis also lists and the vLLM recipe page confirms.

• Active budget per token — Flash activates roughly 8B at prefill and 16B at decode by design, which is the point of its architecture; Pro activates ~49B per token.

• Licence — MIT open weights on both.

• GA date — Flash 2026-09-10; V4 Pro 0813 2026-08-13, after an April preview.

The price gap is the entire argument

Input is 4.4× more expensive on Pro. Output is 3.3×. Cache hits — the tier that dominates a long agent run with a stable system prompt — are 7.3×. Because peak doubles every tier on both sides, the ratio is identical at peak; the gap is not a discount artefact.

Put a workload on it. Take a coding agent doing 10M input tokens a month with a 70% cache-hit rate and 2M output tokens. On V4.1 Flash at off-peak that is $0.45 of fresh input, $0.021 of cached input and $1.20 of output: $1.67. On V4 Pro at off-peak it is $1.98, $0.154 and $3.96: $6.09. Roughly 3.6×, on a workload that is deliberately unremarkable.

That is DeepSeek's own list, and it is the price you actually pay here: OrcaRouter passes provider list rates through at 0% markup, so a DeepSeek price move lands on our side the same day rather than being repackaged. Both models sit on the same key, which matters for the section further down — the one about testing a migration you no longer have to make.

Where DeepSeek V4.1 Flash genuinely wins

The strongest evidence for Flash is not DeepSeek's benchmark table. It is Artificial Analysis, which is independent and read directly from its model pages.

• Intelligence Index — Flash (Reasoning, Max Effort) scores 40, ranked #6 of 113 models. V4 Pro 0813 (Reasoning, Max Effort) scores 36, ranked #7. Same index, same effort setting, four points apart, with the cheaper model ahead.

• Output speed — 214.4 tokens/s against 94.4 tokens/s. That is 2.3×, and it is measured, not claimed.

• Time to first token — 1.21s against 1.74s.

• DeepSWE v1.1 — 74.2 for Flash on the tracked leaderboard, first place, ahead of GPT-6 Astra at 74.10, Claude Opus 5 at 74.00 and Gemini 3.8 Flash at 73.70. V4 Pro 0813's 62.7 on the same benchmark is DeepSeek's own figure. The margin at the top is 0.2 points, which is a tie, not a win — but an 11.5-point gap over Pro is not a tie.

• Serving efficiency — Flash's decoder derives its global KV from the encoder's final hidden state instead of recomputing it per layer, which is what produces the asymmetric 8B-prefill / 16B-decode activation. SemiAnalysis's InferenceX profile puts the KV cache at roughly 890 bytes per token — about a quarter of V4 Flash's — with persistent KV storage down to roughly one eighth. That is the reason the price is what it is, and it is also why the model is available at 2,500 concurrency when Pro caps at 500.

• Vision on the served API — not just in the weights. DeepSeek's own pricing page lists image input as supported for Flash and not supported for V4 Pro, and DeepSeek describes it as its first flagship with native multimodal understanding. This is the first DeepSeek flagship where reading a screenshot needs no separate endpoint.

All the other headline numbers you will see quoted — Terminal-Bench 2.1 at 90.6, GPQA Diamond at 90.9, Codeforces 3471 — are DeepSeek's own, unreproduced by us, and should be read with that in mind.

Where DeepSeek V4 Pro is still the right call

It would be easy to write V4 Pro off after a week like that. The reversed deprecation says otherwise, and the spec sheet says otherwise too.

Pro carries 1.6T total parameters and activates ~49B per token, against Flash's 16B at decode. Whatever the index says, per-token capacity is a real resource, and the workloads where it shows up are the long-horizon ones: multi-hour agent runs, large refactors, tasks where an early wrong turn costs the whole trajectory. A four-point index gap measures the average of ten evaluations, not the tail of your distribution.

Pro is also the known quantity. It has been generally available since 2026-08-13 after an April preview, and the 0813 build gained its performance from post-training rather than architecture — same parameter count, same shape as the preview, different behaviour. That is four months of community tooling, published agent harnesses, prompt patterns and failure modes. Flash has been GA for eight days, and the ecosystem around it is correspondingly thin. If your team's reliability comes from knowing exactly how a model fails, Pro is where that knowledge lives.

And service did not stop. DeepSeek's commitment is explicit and its billing is unchanged, the weights are MIT, and V4 Pro is still in our catalogue at the same list rate. The cancelled deprecation is not a stay of execution — it is a statement that DeepSeek intends to keep serving the tier.

Three things to hold against both models

Guardrails, because this decision is expensive to get wrong.

• The parameter count is contested. 552B is DeepSeek's figure. A credible community analysis argues the released checkpoint is 748B — the 552B transformer backbone plus the ~196B Engram conditional-memory table, which is a lookup structure consulted at two points in the network rather than a layer the signal flows through. The published download is 510.3 GB across 48 safetensors shards of FP8 dense weights with FP4 routed experts. Both numbers can be right at once, depending on whether you count retrieval separately from computation. Treat 552B as vendor-reported and check the disk footprint before you plan a deployment.

• A leaderboard score is a screen, not a contract. DeepSWE v1.1's 74.2 is a harness benchmark. A self-published audit by EyesTech Systems Lab claims these Flash-class scores collapse when moved to isolated, real-world multi-file repositories — putting DeepSeek V4.1 Flash at 31.4% on SWE-bench Pro against 74.2% headline, a larger drop than the frontier models in its own comparison. That audit is not independent — the author sells a tool for detecting exactly the harness exploitation he describes — and we have seen no replication of it. It is still the right posture: verify on your own repos before you migrate.

• Verbosity is a cost line. Artificial Analysis measured V4.1 Flash generating 250M output tokens across its Intelligence Index run, against a median of 140M for comparable open-weights models, and calls it very verbose. Output is the expensive direction in this price table, so a model that thinks out loud eats into its own savings. On a reasoning-heavy workload, budget for token count, not just token price.

One more thing, stated plainly so nobody plans around a rumour: DeepSeek V4.1 Pro is not released. The cancelled migration was framed as a stopgap "before V4.1 Pro launches," and nothing about that has changed.

Pick DeepSeek V4.1 Flash if

• Your workload is high-volume, latency-sensitive, or cost-bounded — classification, extraction, routing, summarisation, chat, most codegen.

• You need image input on the same endpoint as text. This is the first DeepSeek flagship where that is a single call rather than a second model.

• Your prompts are cacheable. At 7.3× on cache hits the gap is widest exactly where agent traffic lives, and a stable system prompt is most of a long run's input.

• You are starting fresh this week and have no harness tuned to a specific model's quirks. There is nothing to protect.

• You want the higher concurrency ceiling. 2,500 against 500 is a five-fold difference in how much you can push through before you are queuing.

Pick DeepSeek V4 Pro if

• You already have production tuned against it. The reversal means you have time — use it to test rather than to avoid the question.

• Your tasks are long-horizon and failure-expensive, and you have measured that ~49B active parameters per token buy you something Flash's 16B does not, on your data rather than on a leaderboard.

• You need predictable, text-only behaviour with no vision path to reason about, or you have eval baselines that a model swap would invalidate mid-study.

• Your access pattern is low-volume and high-stakes, where 3.6× on a small bill is not the deciding term.

If you already built on DeepSeek V4 Pro

The useful thing that changed on 2026-09-11 is that your migration stopped being a deadline and became a decision. That is strictly better for you and strictly worse for your inbox, because the decision still has to be made and now nobody is making it for you.

Start by pricing it. The 3.3–4.4× gap is the number you are leaving on the table, and the worked example above turns it into a monthly figure in about a minute. Then test it in the only way that counts: mirror a slice of production traffic to Flash, keep Pro on the primary path, and diff the outputs on your own tasks. Because both models answer on the same key at provider list price with automatic failover, that shadow run is a routing change rather than a second integration, and the router can fail a request back to Pro without you writing the fallback. Teams burning tokens on a migration they did not ask for are the reason the routing layer exists in the first place.

Do not assume Flash is a drop-in. It is a different architecture — a 40-layer causal encoder–decoder with asymmetric activation — and it is more verbose at reasoning. Prompt-level behaviour will not port cleanly in either direction, so measure the migration on outputs, not on the index.

What to watch

Three things decide how this pairing looks in a month. First, whether V4.1 Pro ships and restores a real Pro tier — the entire cancelled migration was explicitly a stopgap for it, so DeepSeek's roadmap still has a flagship-shaped hole. Second, whether the reprieve survives DeepSeek's next price move; a company willing to schedule a silent reroute once has learned that it cannot, not that it should not. Third, whether anyone outside DeepSeek reproduces the Flash benchmark numbers on unmodified repositories, which is the one result that would settle the efficiency-versus-capacity argument this whole comparison turns on. Until then, the honest position is the one the reversal itself implies: Flash is the better default for anything new, Pro is the better bet for anything already built, and DeepSeek has just told you it will serve both while you work out which one you are.