DeepSeek V4.1 Flash: One Model to Replace Them All, With Prices Back Near Pre-Hike Levels

DeepSeek just shipped V4.1 Flash, and the implications are bigger than the modest version bump suggests. The new model does not merely iterate on its predecessor. It absorbs three separate product lines into one architecture, outperforms the company's former flagship on every published benchmark, and slashes prices back to roughly where they were before DeepSeek's cost hike a few weeks ago.

From Three Models to One

Previously, developers choosing DeepSeek's API had to pick between V4 Flash (fast, cheap, text-only), V4 Flash Vision Exp (same speed with image understanding bolted on), and V4 Pro (the slower, more capable option). V4.1 Flash eliminates that fragmentation entirely.

The new model ships with native multimodal support built into its base weights. A DeepSeek-ViT encoder trained from scratch processes images at arbitrary resolutions up to roughly 1344×1344 pixels, and text and image inputs are handled jointly from the start of pre-training. The old V4-Flash-Vision-Exp approach, which layered a separate vision encoder on top of a text model, is gone.

For developers, this means one API call instead of two. No encoding bridge overhead, no pipeline juggling, and no latency tax for mixing text and images. DeepSeek also removed the previous V4 Flash and V4 Flash Vision Exp endpoints entirely. Requests to those legacy names now route to V4.1 Flash automatically.

V4 Pro Is Retiring Because Flash Is Better

The more surprising move is what happens to V4 Pro. Starting September 14, all requests to deepseek-v4-pro will silently route to V4.1 Flash and bill at Flash prices. DeepSeek's stated reason is blunt: third-party tests show V4.1 Flash ahead of V4 Pro on performance, speed, and total runtime, so continuing to charge more for the weaker model "is no longer appropriate."

Community benchmarks from the September 8 beta window put V4.1 Flash at 355–427 tokens per second with a 178ms time-to-first-token, compared to V4 Pro's 63 tokens per second and 766ms TTFT. That is a roughly 5.7x throughput improvement and a 77% latency reduction. For agent pipelines where latency compounds across tool calls, the difference is not cosmetic.

The move has drawn criticism from some developers who argue that validated production workflows should not have models swapped out from under them without explicit consent. DeepSeek gave roughly 48 hours of beta testing before the switch, announced through a community group post rather than a formal deprecation timeline. The concern is not that V4.1 Flash is worse, but that model behavior shifts in ways that aggregate benchmarks miss.

How They Made It Cheaper and Faster

The price cuts are not a promotional discount. They follow from real architectural changes:

Causal Encoder-Decoder (CED): The 40-layer network splits into a 20-layer encoder and a 20-layer decoder. Instead of recomputing KV cache per decoder layer, it projects decoder KV from the encoder's final hidden states. This roughly halves prefill computation.

Compressed Sparse Attention 2 (CSA2): Each layer runs in one of three modes (Full, Reindex, or Reuse), borrowing KV and attention indices across layers. Combined with FP4 quantization, the global KV cache shrinks to about 890 bytes per token, roughly one quarter of V4 Flash. Compared to DeepSeek V1, the per-token cache is 437 times smaller.

DSpark Speculative Decoding: A lightweight 3-block drafter generates 5 draft tokens per forward pass, replacing the MTP module from V3.

The result is a 748B parameter model (552B backbone + 196B Engram memory module) that activates only 8B parameters per token during prefill and 16B during decode, keeping inference costs low despite the large total count.

Pricing Back Near Pre-Increase Levels

The new Flash pricing, effective September 10, runs:

• Off-peak: $0.003 per million cached input tokens, $0.15 per million uncached input tokens, $0.60 per million output tokens
• Peak hours (01:00–04:00 and 06:00–10:00 UTC weekdays): Double the off-peak rates

Compared to V4 Pro, that is roughly one-third the cost at equivalent hours. The Flash series itself also saw cuts of roughly 11–57 percent depending on token type, meaning prices are now close to where they sat before DeepSeek's recent pricing adjustment. High-volume workloads that previously ran about $87 per million tokens now land closer to $60, and they run nearly six times faster.

What This Means

DeepSeek's move is a bet that raw capability per dollar matters more than product segmentation. Where competitors maintain a ladder of models from cheap/small to expensive/powerful, DeepSeek is collapsing the middle and top rungs into a single fast option. The fact that V4.1 Flash is MIT-licensed and ships with open weights and a full technical report only sharpens the pressure on closed providers.

For developers, the practical upshot is simpler: one model name, one price tier, text and images handled natively, and costs back near pre-hike levels. The tradeoff is the usual one with fast-moving AI providers: your validated workflows may drift slightly as the model under the endpoint changes. DeepSeek is not the first company to route legacy endpoints to newer models, but the speed of this transition, and the fact that the replacement beats the flagship on every metric, makes it notable.

If you are currently on V4 Pro, no code changes are needed for the September 14 switch. Your requests will simply land on a faster, cheaper model. Whether that is a generous upgrade or an unplanned regression test depends on whether you ran evals during the 48-hour beta window. For everyone else, the new model name is deepseek-flash, and it handles both text and images out of the box.