DeepSeek Retires Its 1.6T V4-Pro and Puts the 552B V4.1-Flash in Its Place

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026, applied a new price sheet from 04:00 UTC the same day, and announced that from 04:00 UTC on September 14, 2026 every request naming deepseek-v4-pro will be routed to V4.1-Flash and billed at the V4.1-Flash rate. The new model carries 552B backbone parameters against the 1.6T of V4-Pro, and activates only 8B parameters per token during prefill and 16B during decode. A smaller model is taking the larger one's slot, and DeepSeek's own comparison table shows this is not a clean sweep but a trade made line by line. ASAP works only from figures verifiable in DeepSeek's own documentation and model card to show what the swap gains and what it gives up.

Three lines in the notice change the call path, not the model name

The September 10, 2026 notice in DeepSeek's API documentation settles three things at once. First, DeepSeek-V4.1-Flash is live under the API model name deepseek-flash. Second, the new prices take effect at 04:00 UTC on September 10, 2026. Third, from 04:00 UTC on September 14, 2026, requests addressed to deepseek-v4-pro are all routed to V4.1-Flash and billed at the V4.1-Flash price.

The same notice retires older names. Both deepseek-v4-flash and deepseek-v4-flash-vision-exp are discontinued, and for compatibility those two names temporarily resolve to V4.1-Flash at the Flash price. DeepSeek's stated reason for the swap is that "after extensive testing, V4.1 Flash has comprehensively surpassed V4 Pro in performance, cost, speed, and total time."

In Korea Standard Time the new pricing lands at 13:00 on September 10, 2026 and the V4-Pro reroute at 13:00 on September 14, 2026. A team that never edits a model string still crosses that boundary on a Monday afternoon, and a pipeline that pins versions by name passes through the change without a single code edit.

Whether 552B can stand in for 1.6T is decided item by item inside the table

The instruct comparison table on the Hugging Face model card does not claim V4.1-Flash wins everywhere. The Codeforces rating rises 123 points, from 3348 on V4-Pro to 3471 on V4.1-Flash, and MathArena Apex is effectively level at 65.3 against 65.6. In the other direction, HLE, which is answered without tools, falls 5.9 points from 42.7 to 36.8, and GPQA Diamond falls 1.5 points from 92.4 to 90.9.

The base model table points one way only. HumanEval runs 69.5 on V4-Flash and 76.8 on V4-Pro before reaching 79.4 on V4.1-Flash, GSM8K runs 90.8 and 92.6 before 93.0, and BigCodeBench runs 56.8 and 59.2 before 60.6. All three peak on the new model even though the backbone shrank from 1.6T to 552B, less than a third of the size.

Agentic scores move the same way. The instruct figures published on the card include DeepSWE v1.1 at 74.2 and Terminal-Bench 2.1 at 90.6. Multimodal rows are new to this generation, with MMMU-Pro at 56.5, CVBench at 77.9, and DocVQA at 95.6. A Flash line that had been text-only now carries native visual input.

Output pricing falls to a third of V4-Pro and by 9 percent against V4-Flash

Per DeepSeek's pricing document, off-peak rates for deepseek-flash per 1M tokens are $0.003 for cached input, $0.15 for cache-miss input, and $0.6 for output. Peak rates are exactly double at $0.006, $0.3, and $1.2. Peak hours run 01:00 to 04:00 and 06:00 to 10:00 UTC Monday through Friday, and every other hour is off-peak.

V4-Pro's off-peak rates were $0.022 for cached input, $0.66 for cache-miss input, and $1.98 for output. Line by line, the new model costs one seventh, one fourth, and one third as much. Moving the same traffic over drops the bill below a third for any workload weighted toward output tokens.

Against the outgoing Flash the cuts are far narrower. Compared with V4-Flash off-peak rates of $0.007, $0.22, and $0.66, the reductions are roughly 57 percent, 32 percent, and 9 percent. The real beneficiary of this price change is the Pro user, not the Flash user. For someone already on Flash, where output falls only 9 percent, the headline of this notice is not price but capability and the 1M-token context.

The savings originate in a KV cache that shrank to 890 bytes per token

The card describes a Causal Encoder-Decoder built from a 20-layer causal encoder followed by a 20-layer decoder. The decoder's KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states, and that design is what allows 8B active parameters during prefill and 16B during decode. Each MoE layer holds 1 shared expert and 384 routed experts, activating 6 routed experts per token. Separately, an Engram conditional memory of 196B parameters is accessed sparsely through token-based lookup.

Cache compression comes from two mechanisms. One is FP4 main KV caching in E2M1 format with one E4M3 scale per 16 channels. The other is Compressed Sparse Attention 2, which assigns each attention layer one of three static modes named Full, Reindex, and Reuse so that main KV and indexer K are shared across layers. Together they bring the global KV cache footprint to roughly 890 bytes per token, about one quarter of DeepSeek-V4-Flash. DeepSeek's notice states the same result as one quarter of the HBM and one eighth of the SSD storage previously required.

Multiplying that figure by context length shows the scale. At 890 bytes per token, one full 1M-token sequence holds about 890 MB of KV cache. The same arithmetic on the prior generation exceeds 3.5 GB. For a 1M-token window to be a servable number rather than a spec-sheet ornament, roughly a fourfold reduction was required, and a fourfold reduction is exactly what this release delivers.

Giving up 5.9 points of knowledge buys a third of the price and more concurrency

From here the reading is ASAP's own. What DeepSeek executed is not an upgrade but a trade. It surrendered 5.9 points on HLE and 1.5 on GPQA, both measured without tools, and collected gains in code, math, agentic execution, and price. Some of the knowledge held in a 1.6T backbone did not survive the move into 552B, and DeepSeek chose to accept that loss.

Whether the trade favors a given deployment depends on which side that deployment stands. Systems that attach retrieval and documents rarely expose a drop in parametric knowledge, because the context carries the answer, and those same systems collect the 1M-token window and the lower cache-miss rate as direct gains. Systems that let the model answer from what it knows, such as specialist question answering that leans on parametric recall, are the ones exposed to quality regression after September 14, because an identical prompt now reaches a different model under an identical name.

The asymmetry in activated parameters deserves attention as well. Prefill at 8B and decode at 16B means reading long input costs half of what producing short output costs. The fourfold gap between $0.15 for cache-miss input and $0.6 for output writes that structure directly into the price sheet. Designs that push long documents in and ask for short answers out are favored by both architecture and pricing, which reads as an explicit signal that DeepSeek has aimed this generation at long-document summarization, codebase queries, and retrieval-augmented work.

What a team actually meets at 13:00 KST on September 14 is a silent swap

The operational core of this notice is that version pinning is unavailable. DeepSeek does not decommission deepseek-v4-pro; it points the same name at a different model. A team that manages versions by model string crosses 13:00 KST on September 14, 2026 without touching code, and every response after that moment comes from a different architecture. Because no deployment occurs, there is no natural trigger for a regression run either.

The practical preparation is to capture a baseline before the switch. Running the current prompt set against V4-Pro and storing the responses gives something to place beside the same inputs afterward. Three categories deserve priority: queries that lean on parametric knowledge, areas where training coverage may be thin such as Korean proper nouns and domestic regulations, and prompts that demand strict output formatting.

Budget preparation runs the opposite direction. Because rates fall, the question shifts from controlling spend to reallocating it. When output costs a third as much, the same budget buys three times the output tokens, and retry or self-verification stages that were cut for cost become available again. Peak hours of 01:00 to 04:00 and 06:00 to 10:00 UTC map to 10:00 to 13:00 and 15:00 to 19:00 in Korea, so overnight batches stay inside the half-price window without being moved.

Every number in the table still comes from DeepSeek's own harness

Three verification limits remain on the DeepSeek-V4.1-Flash model card, covering self-reported benchmarks, absent speed measurements, and the storage cost of 196B Engram parameters. First, the comparison values are self-reported measurements taken in DeepSeek's own environment rather than third-party reproductions. Competitor entries such as Opus-5.0 at 93.4 on GPQA Diamond and 56.3 on HLE, GPT-5.6 Sol at 94.1 on GPQA, and K3 at 92.9 are not documented as having been measured under identical conditions. Second, although DeepSeek names speed and total time among the metrics it claims to lead on, no measured figures such as tokens per second or time to first token appear in the public documentation.

Third, the license is MIT and the weights are published at deepseek-ai/DeepSeek-V4.1-Flash on Hugging Face, but the self-hosting threshold is not computed from the backbone alone. The 552B backbone is joined by 196B of Engram parameters, and describing Engram as sparsely accessed reduces computation without reducing storage. The volume that must actually be downloaded and loaded sits closer to the sum of the two.

The swap therefore reduces to one sentence. DeepSeek withdrew the premise that a bigger model is a better model from its own price sheet, paid for that withdrawal with several points on knowledge benchmarks, and left users no model name through which to decline the exchange. Response quality after September 14 decides whether the trade pays, and that verdict exists only inside each team's own prompts.

Source: DeepSeek API documentation notice of September 10, 2026; DeepSeek API pricing documentation; Hugging Face model card for deepseek-ai/DeepSeek-V4.1-Flash