DeepSeek V4.1 Flash Replaces V4 Pro: Pricing, Benchmarks, and What Changed Since V4 Flash

DeepSeek released DeepSeek-V4.1-Flash on September 10, 2026. Four days later, its own flagship stops serving its own traffic: from 12:00 Beijing time on September 14, which is 04:00 UTC, every request to deepseek-v4-pro is routed to V4.1 Flash and billed at Flash rates, until a V4.1 Pro is released. Anyone building on Pro gets that switch applied for them, and with it a drop of about 77 percent in cache-miss input cost and about 70 percent in output cost.

DeepSeek’s stated reason is that V4.1 Flash surpassed V4 Pro in performance, cost, speed and total time. Its own published numbers support that on agent work and complicate it on knowledge recall, which is the more useful story. Here is what changed since V4 Flash, what it costs, and the one table in this release that should change how you read any model comparison.

What Changed Since V4 Flash

V4.1 Flash is a generational replacement, not a point update. V4 Flash and the experimental V4 Flash Vision are retired; their old model names temporarily route to V4.1 Flash so existing code keeps running. All figures below come from DeepSeek’s official model card.

Two rows carry most of the weight. The model roughly doubled in backbone size while the parameters it activates to read your input fell from 13B to 8B, and Terminal-Bench 4.0, the hardest current agentic coding evaluation, went from 7.0 to 31.2. Our earlier write-up on DeepSeek V4 Flash is the baseline for that comparison, and the Kimi K3 versus DeepSeek V4 matchup reads differently now too.

Pricing and the V4 Pro Switch

DeepSeek prices per million tokens, with peak and off-peak rates. Off-peak is exactly half of peak, and peak is 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, so most of the clock is discounted.

What the switch means in practice

Read the last column against the middle one and the retirement explains itself. DeepSeek’s official pricing page confirms the routing date and rates, and the concurrency limit rises from 500 on V4 Pro to 2,500 on Flash. Three things are worth flagging for anyone running production traffic. The switch is automatic, so a workload tuned for Pro’s behavior will change model underneath it on September 14 without a code change. Cache-hit input is priced at $0.006 against $0.30 for a cache miss, a fiftyfold difference, so how your requests reuse context now matters more to the bill than the rate card does, a mechanic our explainer on inference cost covers. And if you are moving off V4 Flash rather than V4 Pro, note that the retired model’s rate card is no longer published and the reduction DeepSeek describes is measured against what Pro traffic was paying, so check your own per-token cost before assuming this release made your workload cheaper.

Where V4.1 Flash Beats V4 Pro, and Where It Does Not

Every number below is DeepSeek’s own run at maximum reasoning effort. Independent replication has not happened yet, so read these as the vendor’s evidence rather than as settled results, and see the next section for why small gaps between models deserve particular caution.

The split is legible once the V4 Pro column stays in view for both halves of the table. V4.1 Flash leads Pro clearly wherever the model has to act: run a terminal, drive a repository, hold a multi-step agent loop together. It is close to Opus 5 and GPT-5.6 Sol on the older agentic tests and far behind Opus 5 on the current Terminal-Bench 4.0 and on NL2Repo-Bench.

Where the smaller model loses to Pro

On the tests that ask a model to recall or reason unaided, the 1.6T-backbone Pro is still ahead: 92.4 against 90.9 on GPQA Diamond, and 42.7 against 39.1 on the text-only subset of Humanity’s Last Exam. The base-model evaluations show it more starkly, with V4 Pro at 55.2 on SimpleQA-Verified against 42.3 for V4.1 Flash Base, a gap in factual recall that a much smaller backbone would be expected to show. So DeepSeek’s phrase “comprehensively surpassed” is the company’s framing. The precise version, on its own data, is that the new Flash overtakes Pro across agentic work and gives up ground on unaided knowledge and reasoning. If tool use is central to your workload, the trade favors you. If accurate recall without retrieval is, it may not.

The Table That Should Change How You Read Model Comparisons

DeepSeek also published something that rarely appears in a launch: the same model, on the same benchmark, run through eight different agent scaffolds.

One model, one benchmark, an 8.7 point spread on DeepSWE and 6.5 on Terminal-Bench, produced entirely by the software wrapped around the model. Now compare that spread with the gap the launch is built on. On DeepSWE, V4.1 Flash, Opus 5 and GPT-5.6 Sol sit at 74.2, 74.0 and 73.0, a range of 1.2 points. The tooling moves the same model seven times further than that.

It is worth being careful about what this does and does not establish. It does not let you re-rank models by pairing one vendor’s weakest scaffold against another’s best, and it does not mean the published figures are wrong: every lab reports its strongest configuration, and DeepSeek’s 74.2 comes from mini-SWE. What it does establish is that a lead of one or two points across vendors, measured with different tooling, carries very little information about which model will do better in your stack.

That has a direct practical use. When you evaluate models at work, hold the harness fixed and change one variable at a time, because a model that scores worse inside your setup may simply be wrapped badly. Ask which scaffold produced any number you are shown. And treat the benchmark as a reason to run your own test rather than a substitute for one. Of everything in this release, this table is the part most likely to save you money.

Architecture: Where the Efficiency Comes From

The encoder-decoder split

V4.1 Flash uses a 40-layer Transformer arranged as a 20-layer causal encoder followed by a 20-layer decoder. The decoder’s global key-value cache is projected from the encoder’s final hidden states instead of being derived layer by layer, which is what allows 8B active parameters while reading and 16B while writing. DeepSeek presents this as a cost-efficiency gain for input-heavy agent workloads, and the direction is sound, since agents read files, logs and prior turns far more than they write. How much any given job actually costs still depends on tokens consumed end to end, which no parameter count settles.

KV cache compression

The key-value cache is the model’s working memory for a conversation, and in long-context work it is usually what runs out first. DeepSeek combines Compressed Sparse Attention 2, which assigns each attention layer one of three modes so layers can share cache and reuse sparse-attention indices, with FP4 quantization of the main cache. The reported result is 890 bytes per token, about a quarter of V4 Flash, and a separate technique called SWA Bounded Replay cuts the persistent footprint to roughly an eighth. Against DeepSeek-V1 the per-token cache is down by a factor of 437. DeepSeek’s framing is that cache-hit charges are a large share of agent costs, so the compression is aimed at the bill; whether it reaches yours depends on your rates and how much context you reuse.

Three smaller additions round it out. Engram is a 196B-parameter conditional memory held separately from the 552B backbone and reached by sparse token-based lookup. DSpark speculative decoding drafts several tokens ahead and verifies them, which is where a share of the throughput gain originates rather than in the cache work. And reasoning effort is now a continuous integer from 1 to 100 rather than a few named modes, which turns thinking budget into a per-task setting.

One pre-training detail is worth noting alongside the architecture, because it is what the one million token context rests on. DeepSeek trained the model from scratch on a 45T-token multimodal corpus, ran sparse attention at a 64K sequence length, and extended the context to 1M tokens only at the 34T-token mark. The long context is a late-stage extension of a model trained mostly at shorter lengths, which is the usual pattern and a fair reason to test long-context behavior yourself rather than assume it is uniform across the window.

Native Vision, Now in the Base Model

Previous DeepSeek vision support was a separate experimental model. In V4.1 Flash a purpose-built vision encoder and a two-layer projector feed image embeddings into the language model from the start of pre-training rather than being attached afterward. The published evidence is base-model scores of 95.6 on DocVQA and 56.5 on MMMU-Pro, which point at documents, charts and screenshots as the practical uses. On visual agent tasks with tools it still trails Opus 5, 89.6 against 94.1 on BabyVision. Our glossary entry on multimodal AI covers what native training means here.

What Is Confirmed for Non-Developers, and What Is Not

Worth being straight about this. DeepSeek’s announcement and changelog cover API availability and three partner integrations, WorkBuddy, CodeBuddy and OpenCode. The company has not published which model version serves its consumer chat app, whether the one million token context window is available there, or what the usage limits are. The app being free does not confirm that V4.1 Flash is behind it, so if that matters to you, treat it as unverified for now. Our guide to using DeepSeek for free covers the access routes that are documented.

For scale, one million tokens is roughly two thousand pages of text, which is the order of a full contract set or a quarter of technical documentation rather than a long email. Two things transfer regardless of which model an app runs. Compare assistants on your own work rather than on a chart, for the reason the scaffold table gives. And keep data questions separate from capability questions, which we cover in is DeepSeek safe; an MIT license means someone other than DeepSeek may be hosting the weights, and that changes the answer.

Reading Releases Like This One

Judging what a model is good for, asking which setup produced a number, and testing on your own work are transferable skills, and they outlast any single release. Coursiv teaches them through short daily lessons and hands-on practice with real AI tools, built for people without a technical background. Check the official site for current course details and pricing.

What to Watch

Watch September 14, when V4 Pro traffic moves across and the first large-scale production signal arrives. Watch for independent DeepSWE and Terminal-Bench runs with a stated scaffold, which is the only thing that will settle the near-tie with Opus 5. Watch for the expected V4.1 Pro, which DeepSeek references without a date and which our tracker on the DeepSeek V5 release date follows. And watch list prices, because a 552B open-weights model at $0.30 per million input tokens puts pressure on everything above it.

The Bottom Line

V4.1 Flash is an efficiency release, it is good enough at agent work that DeepSeek is parking its own flagship behind it, it gives up real ground on unaided knowledge, and the scaffold table it published is worth more than the leaderboard it tops.

FAQ