DeepSeek Retired V4 Pro: What the V4.1 Flash Switch Means for Your Stack (2026)

DeepSeek shipped V4.1 Flash on September 10, 2026, and then made a decision that matters more to anyone with code in production than the release itself. From 04:00 UTC on September 14, the deepseek-v4-pro endpoint stopped serving V4 Pro and started serving V4.1 Flash instead, at Flash prices, and DeepSeek has said this holds until a future V4.1 Pro arrives. The reasoning is blunt: the smaller new model beats the older, larger one on most of the tests DeepSeek runs, so keeping the expensive model alive made no sense. If your service calls deepseek-v4-pro, the model behind that string changed this week without your deploying anything.

That is a rare and clarifying move. Most vendors deprecate slowly and leave the old model reachable. DeepSeek collapsed the tier outright. So the practical questions are the ones this post answers: what is different about the model you are now getting, what does it do to your bill, and what should you check before you trust the swap.

Attribute	DeepSeek V4 Pro (retired)	DeepSeek V4.1 Flash (now serving)
Architecture	1.6T parameter MoE	552B parameter MoE, roughly 8B active on input and 16B on output
Weights	Open	Open weights, MIT license
Context window	1,000,000 tokens	1,000,000 tokens
Input price (per 1M)	about 0.435 dollars	about 0.15 dollars off-peak on a cache miss
Output price (per 1M)	about 0.87 dollars	about 0.60 dollars off-peak
DeepSWE v1.1	62.7	74.2
Terminal-Bench 2.1	87.9	90.6
GPQA Diamond	92.4	90.9
HLE (no tools)	42.7	36.8

What actually changed in the model

V4.1 Flash is smaller and cheaper, and it is not uniformly better. It is better exactly where most production traffic lives and worse in a place that matters for a narrower set of jobs.

On agentic and tool use work it pulls clearly ahead of the model it replaced. DeepSWE v1.1 climbs to 74.2 from 62.7, and Terminal-Bench 2.1 to 90.6 from 87.9. Those are the tasks where an agent has to plan, call tools, and finish a multi step job, which is precisely the workload a cheap high volume model is built to serve. For an agent that runs at scale behind a review gate, this is a straight upgrade: cheaper per call and more likely to complete the job.

On knowledge heavy reasoning without tools it gives ground. GPQA Diamond slips to 90.9 from 92.4, and Humanity's Last Exam drops to 36.8 from 42.7. That pattern is consistent with a smaller mixture of experts that activates only a few billion parameters per token: it reasons well when it can call tools and search, and it has less raw parametric knowledge to draw on when it cannot. If your workload is closed book question answering with no retrieval, the retired larger model was genuinely stronger, and you have lost that option.

What the switch does to your bill

The headline is that the same endpoint got cheaper. V4 Pro billed at roughly 0.435 dollars per million input tokens and 0.87 per million output. V4.1 Flash, off-peak, runs at about 0.15 dollars per million input on a cache miss, about 0.003 on a cache hit, and about 0.60 per million output. Peak hour rates, in the windows DeepSeek defines around 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, roughly double the off-peak numbers. So the cost you actually pay now depends on two things the old flat tier did not force you to think about: when your traffic runs, and how well your prompts cache.

The caching gap is the part worth engineering around. A cache hit at 0.003 dollars per million input tokens is two orders of magnitude below a cache miss. If your prompts share a large stable prefix, a system prompt, a tool schema, a retrieved document set, then structuring calls so that prefix is cached turns the input side of your bill into a rounding error. If your prompts are unique every time, you pay the miss rate and the caching does nothing for you. We took the general version of this apart in AI coding agent pricing for 2026: the sticker rate is rarely the number that describes your spend, and here the spread between cache hit, cache miss, off-peak, and peak is wide enough that two teams on the identical endpoint can see bills that differ by more than ten times.

The open weight angle

V4.1 Flash ships open weights under MIT with no regional restriction, which changes the calculus if you were relying on V4 Pro's open weights for self hosting. The good news is that a 552 billion parameter model is far more tractable to serve than a 1.6 trillion parameter one. The old model was heavy enough that running it locally meant expert offloading and serious hardware, which we walked through in running DeepSeek V4 Pro locally with expert offloading on Apple Silicon. The new model's smaller total parameter count and low active count per token make a self hosted deployment more realistic for teams that want the weights pinned and the inference under their own control, though it is still a large MoE that wants real GPU memory rather than a laptop.

The strategic point is that DeepSeek retiring its own premium tier while keeping the smaller model open is a signal about where the open weight coding market is going. The strongest open model on standard coding tests is still a larger system, and we compared the retired V4 Pro against that class in GLM 5.2 versus DeepSeek V4 Pro on competitive programming. With V4 Pro gone from the API, that comparison now reads as a question about which open weights to actually deploy.

What to do this week

Re-benchmark the endpoint against your own tasks. The model changed, so your prior evaluation of deepseek-v4-pro is stale. Run your real prompts through it and check the two failure directions specifically: agentic tasks should improve, tool free knowledge questions may regress.

Model your cost on cache behavior and traffic timing, not the sticker rate. Measure your cache hit rate, decide whether you can batch non urgent traffic into off-peak windows, and restructure prompts to maximize a shared cached prefix.

Decide whether to pin weights. If you depended on V4 Pro for closed book reasoning quality, or you need a fixed model that will not be swapped under you again, the MIT weights let you self host a known version rather than accept whatever the endpoint routes to next.

FAQ

Is DeepSeek V4 Pro gone for good? The endpoint is serving V4.1 Flash as of September 14, 2026, and DeepSeek has framed this as lasting until a V4.1 Pro ships. The old model's open weights still exist, so you can self host that version, but the hosted V4 Pro tier is retired.

Is V4.1 Flash better than V4 Pro? On agentic coding and tool use, yes, with clear gains on DeepSWE and Terminal-Bench. On knowledge heavy reasoning without tools it is weaker, giving up points on GPQA Diamond and Humanity's Last Exam.

Why is my DeepSeek bill different now? Because the endpoint moved to Flash pricing, which is cheaper on the sticker but split into cache hit, cache miss, off-peak, and peak rates. Your effective cost now depends on caching and when your traffic runs.

Can I still run the old model? Yes, if you self host the open weights. The hosted API will not give you V4 Pro anymore.

When this applies to your stack

If you have DeepSeek in production, this switch is a small reminder of a large risk: when your model is a string pointing at someone else's endpoint, the vendor can change what that string means. The response is not to panic, it is to own your evaluation and your cost model, and to decide deliberately which workloads should run on a hosted endpoint that can move under you and which should run on weights you have pinned yourself. Designing that split, and the routing and caching layer that makes the economics work, is the kind of AI infrastructure work we do at Contra Collective.