DeepSeek V4: The Open-Source Challenger Disrupting AI Pricing
In short: As of September 5, 2026, DeepSeek V4 Flash is my choice for high-volume AI tasks, while V4 Pro is better suited to agents and difficult problems. Both models offer a one-million-token context and open weights. Their prices are aggressive, but results published by DeepSeek still show gaps against the best closed models.
The DeepSeek V4 preview was officially announced on April 24, 2026. Since then, V4 Pro has reached general availability and the lineup has gained new reasoning modes as well as a new pricing schedule. This update changes the initial verdict: DeepSeek remains very competitive, but does not dominate every use case.
What does DeepSeek V4 actually change?
DeepSeek V4 is a Mixture of Experts (MoE) family composed of two variants:
- V4 Pro: 1.6 trillion total parameters, with 49 billion activated per token
- V4 Flash: 284 billion total parameters, with 13 billion activated per token
These characteristics come from the DeepSeek V4 technical report published in April 2026. Weights for both models are distributed under the MIT license according to the official V4 Pro model card.
The main evolution concerns long-sequence processing. DeepSeek combines:
- Compressed Sparse Attention (CSA), which compresses the KV cache along the sequence dimension before applying sparse attention;
- Heavily Compressed Attention (HCA), which compresses the cache more aggressively while keeping dense attention;
- Manifold-Constrained Hyper-Connections (mHC), intended to stabilize signal propagation;
- the Muon optimizer, used to accelerate convergence and stabilize training.
According to the report’s estimates, with a one-million-token context, V4 Pro requires only 27% of inference FLOPs per token and 10% of the KV cache of V3.2. These are vendor measurements rather than an independent benchmark, but they explain why DeepSeek can offer such a long window at a low price.
Is the one-million-token context really useful?
Yes, but its size alone does not guarantee quality. The official pricing documentation consulted on September 5, 2026 indicates a one-million-token context and a maximum output of 384,000 tokens for both V4 Pro and V4 Flash.
This capacity is particularly useful for:
- analyzing several large documents in a single request;
- retaining more files from a software repository;
- extending an agentic workflow without summarizing its history as often;
- exploring several hypotheses with a high reasoning budget.
It does not turn the model into a perfect search engine. In results published by DeepSeek, V4 Pro Max reaches 83.5 on MRCR 1M, compared with 92.9 for Claude Opus 4.6 Max. On CorpusQA 1M, DeepSeek reports 62.0 for V4 Pro Max, compared with 71.7 for Opus 4.6 Max. For critical RAG, therefore, test retrieval on your own documents instead of relying solely on the advertised length.
What do DeepSeek V4 benchmarks really say?
The official results are strong in code, but they do not show an overall victory.
| Benchmark | V4 Pro Max | Best competitor shown in the report |
|---|---|---|
| LiveCodeBench | 93.5 | Gemini 3.1 Pro High: 91.7 |
| Codeforces | 3,206 | GPT-5.4 xHigh: 3,168 |
| SWE Verified | 80.6 | Claude Opus 4.6 Max: 80.8 |
| Terminal Bench 2.0 | 67.9 | GPT-5.4 xHigh: 75.1 |
| SimpleQA-Verified | 57.9 | Gemini 3.1 Pro High: 75.6 |
All these figures come from DeepSeek’s model card and report. Protocols, reasoning budgets, and environments can vary across models. They should therefore be read as results reported by the vendor, useful for directing a trial, and not as a guarantee for your project.
Where is V4 Pro convincing?
- Code generation and problem-solving: its LiveCodeBench and Codeforces scores are at the front of the published table.
- Software agents: its SWE Verified score remains close to the best result shown.
- Long reasoning: the Max mode increases effort on difficult problems.
Where does it still lag?
- Retrieval in a very long context: Opus 4.6 Max retains a clear lead on MRCR 1M and CorpusQA 1M.
- Factual knowledge: Gemini 3.1 Pro High is well ahead of V4 Pro Max on SimpleQA-Verified.
- Terminal use: GPT-5.4 xHigh achieves a better result in the Terminal Bench 2.0 table.
The report itself estimates that V4’s reasoning capabilities trail the leading closed models by roughly three to six months. That frankness is useful: V4 primarily aims for a compromise between quality, openness, context, and cost.
V4 Pro or V4 Flash: which should you choose?
Both models share the same context length and Non-think, High, and Max modes. Their difference mainly concerns capacity, speed, and price.
| Need | Recommended choice | Why |
|---|---|---|
| High-volume extraction, classification, summarization | V4 Flash | Lower cost and sufficient capacity for routine tasks |
| Demanding development agent | V4 Pro | Better official results on complex workflows |
| Sensitive factual research | V4 Pro with external verification | Flash declines further on SimpleQA-Verified |
| Experimental local deployment | V4 Flash | A much smaller model, but still very heavy |
| Critical decision | Comparison on your data | No public benchmark exactly represents your risk |
A good strategy is to start with Flash, then route requests that need more reasoning or reliability to Pro. This avoids paying for the heaviest model for every operation.
How much does the DeepSeek V4 API cost?
Since the update announced with the general availability of V4 Pro in August 2026, DeepSeek distinguishes between peak and off-peak hours. The prices below come from the official documentation consulted on September 5, 2026, in dollars per million tokens.
| Model | Uncached input, off-peak | Uncached input, peak | Output, off-peak | Output, peak |
|---|---|---|---|---|
| V4 Flash | $0.22 | $0.44 | $0.66 | $1.32 |
| V4 Pro | $0.66 | $1.32 | $1.98 | $3.96 |
The peak hours listed by DeepSeek are 01:00–04:00 and 06:00–10:00 UTC, Monday through Friday. All other periods receive the reduced rate. Cached input costs even less, but your bill will mainly depend on the input-output ratio, cache hit rate, and reasoning volume.
V4 Flash costs exactly one-third of V4 Pro in this schedule. For deferrable workloads, scheduling batches during off-peak hours also halves the displayed price. These prices can change: always check the official page before setting a budget.
Which limitations should you plan for in production?
A long context can consume many tokens. An available window does not need to be filled. Select useful files and passages before each call.
Maximum reasoning increases latency and output. Reserve it for problems that justify the cost. Flash without reasoning is better suited to simple transformations.
Open weights do not make self-hosting easy. Even Flash has 284 billion parameters. Downloading is possible, but performant operation requires specialized infrastructure.
The benchmarks are DeepSeek’s. They make it possible to compare variants under the published conditions. To choose a production model, build an evaluation set from your repositories, documents, tool calls, and costly errors.
What is my verdict on DeepSeek V4?
DeepSeek V4 is a credible advance, especially for long-context efficiency and API cost. I recommend V4 Flash as a starting point for high-volume processing, then V4 Pro for coding agents and complex requests.
However, I would not automatically replace a closed model on a critical task. V4 Pro matches or exceeds some competitors in the published coding results, but remains behind on several factuality, terminal, and long-retrieval tests. The right choice is therefore not the model with the best overall score: it is the one that succeeds on your difficult cases at an acceptable cost.
Sources
Frequently asked questions
Is DeepSeek V4 open source?
DeepSeek V4 weights are released under the MIT license. The model can therefore be downloaded and deployed, subject to having infrastructure suited to its size.
What is the difference between DeepSeek V4 Pro and Flash?
V4 Pro targets the most demanding knowledge, reasoning, and agent tasks. V4 Flash is smaller, faster, and three times cheaper on the official API, but loses more precision on complex tasks.
Does DeepSeek V4 support one million tokens?
Yes, DeepSeek advertises a one-million-token context window for V4 Pro and V4 Flash. That capacity does not, however, guarantee perfect retrieval of every piece of information in such a long document.
Should I choose DeepSeek V4 for coding?
Yes for experimentation, high-volume work, or reducing a development agent's costs. For a critical repository, compare it with your current models on your own tasks, because vendor benchmarks do not replace real-world testing.