ModelsDeepSeek V4 Pro vs V4 Flash: Same Window, Different Brain
Both ship 1M context and an MIT licence. Pro activates 49B parameters per token, Flash 13B. Where that single difference decides which one you should run.
Model families usually differentiate their tiers on context length. DeepSeek did not. V4 Pro and V4 Flash both ship 1M-token windows, both under MIT, released together on 23 April 2026. The entire difference is how much of the model runs on each token.
That makes this an unusually clean comparison. You are not trading context for cost, or licence for capability. You are trading exactly one thing: reasoning depth.
The one number that differs
V4 Pro activates roughly 49B parameters per token out of a 1.6T total. V4 Flash activates about 13B. Everything downstream — price, throughput, and where each model falls over — comes from that ratio.
On price, Flash lists around $0.14 per million input tokens and $0.28 per million output. Pro sits meaningfully higher, in the region of $0.44 and $0.87. Roughly three times on both sides.
Because the context window is identical, both models can read the same enormous inputs. The question is only how well they reason about what they read.
Why "cheaper model, more retries" is not automatically wrong
The standard objection to a cheap model is that it fails more, so you retry, so you do not save anything. That is sometimes true and often not, and the deciding factor is whether failures are detectable.
If a failure is caught by a test suite, a schema validation, or a compiler, then a cheap model plus a retry is genuinely cheaper — you pay three times less on the successes and only occasionally pay twice. If failures are silent and land in a pull request that a human has to catch, the cheap model is more expensive in the only currency that matters.
So the real question is not which model is better. It is whether your pipeline can tell when it has been given a bad answer. Evaluating agent reliability covers building that check, and it is worth doing before choosing a tier.
Where the gap widens
Flash's disadvantage is not uniform. It concentrates in a predictable place: step count.
In a single-turn task, a slightly weaker model gives a slightly worse answer. In a twenty-step agent loop, each step's output is the next step's input, so per-step error compounds rather than averages. A model that is a few points behind on single-turn evaluation can be dramatically behind on completion rate over a long run.
This is why the two models can look nearly identical on a quick trial and diverge sharply in production. Trials are short. Production is not.
The practical consequence: evaluate Flash on your longest tasks, not your typical ones. If it holds up at twenty tool calls it will be fine at three.
Where Flash is simply the right answer
Work where the difficulty is volume rather than depth.
- Reading. With a 1M window at $0.14 per million input tokens, feeding Flash an entire subsystem costs cents. Pro reads the same input three times dearer for no benefit if the question is "where is this implemented".
- Bulk extraction and classification. Thousands of low-difficulty items, where a three-fold price difference dominates.
- First-pass triage. Flash reads everything and decides what deserves Pro's attention.
Running both
Because they share a licence, a window and an API shape, routing between them is unusually cheap to implement — no prompt reshaping, no separate credentials.
The pattern that works: default to Flash, escalate to Pro on an explicit trigger. Good triggers are mechanical — a failing test, a turn-count ceiling, a schema validation error. Bad triggers are vibes.
Instrument the escalation rate from the start. The common failure is that the threshold drifts, everything escalates, and you are paying Pro prices with extra latency. Model routing and fallbacks covers the mechanics; agent cost control patterns covers keeping it honest.
How to choose
Take twenty tasks that span your real range, including your longest. Run each through both. Do not average the results — look at where Flash failed.
If the failures cluster on a measurable property, such as everything past eight tool calls, that is your routing rule and you should run both. If the failures are scattered with no pattern, routing will not help and you should pick one model on total cost of completed work.
How to benchmark LLMs on your own work has the harness, and when a cheap model is enough covers setting the threshold without guessing.
Common questions
Do DeepSeek V4 Pro and Flash have the same context window?
Yes. Both ship 1M tokens, which is unusual — most families cut context on the cheaper tier. The entire difference is active parameters: roughly 49B for Pro against 13B for Flash.
Is V4 Flash good enough to replace Pro?
For retrieval, bulk classification and short bounded tasks, frequently yes. For long unattended agent loops, usually not — per-step error compounds across steps, so the gap widens sharply with task length.
Are both models MIT licensed?
Yes. The whole V4 family ships under MIT, so you can fine-tune, redistribute and build a commercial inference product on either with no revenue threshold or separate agreement.
