DeepSeek just dropped two new models, V4 Pro and V4 Flash, on April 24, 2026. Same day OpenAI pushed GPT-5.5. The timing is not a coincidence, and the numbers explain why.
V4 Pro is already sitting at #2 on the Arena.ai text leaderboard, right behind GLM 5.1. It runs with a 1-million-token context window. The weights are open. And the pricing, well, the pricing is the story: $1.74 per million input tokens versus $5 for Claude Opus 4.7, and $3.48 on output versus $25. On output alone, that works out to roughly seven times cheaper. On blended workloads where output dominates, the gap can reach 15x.
For anyone currently burning $300 to $700 a month on premium API calls, this changes the math. Maybe not for every task. But for a lot of them.
What’s actually in the box
V4 Pro is the heavyweight: 1,600 billion parameters total, with 49 billion activated per query. V4 Flash is the fast variant, 284 billion total parameters and 13 billion active. Both ship with a 1-million-token context window. Both are open source and free to download and deploy commercially.
The activation math is what keeps inference cost down. When you ask V4 Pro a question, it does not fire all 1.6 trillion parameters. It routes the query through a subset of specialized pathways, roughly 49 billion worth. Think of it as the model quietly picking which experts show up for work on a given prompt.
Compared to V3.2 from December 2025, the size jumped by more than 2x. But the efficiency gains matter more. Reading a 500-page document now costs about 10 times less compute than with V3.2, and memory usage dropped from around 50 GB to roughly 5 GB. That is a genuine architectural leap on long contexts, which was a weak spot for DeepSeek before.
The real story here is long-context efficiency. Inference on V4 Pro got roughly 10x cheaper in compute and memory compared to V3.2. That is what takes a 1M-token window from spec sheet curiosity into actual production use.
Where V4 Pro actually wins
On LiveCodeBench, V4 Pro is currently on top. It beats Claude Opus 4.7 on paper, and it leads on real-world bug resolution too. That is a surprising result for a model this cheap, and it has to be tested in actual production before anyone should rearrange their stack. But the numbers are there.
On factual accuracy, where the question is whether a model will tell you the truth instead of making something up, V4 Pro is well ahead of the three American flagships. It is also visibly less filtered. Users will notice it answers things Claude and GPT-5 sidestep.
Where it still loses
No model is best at everything, and V4 Pro has visible weak spots. On Humanities Last Exam, a brutal abstract reasoning benchmark, it scores 37.7 against 44.4 for Gemini 3.1 Pro. Empirically that matches the feel of it: for the hardest, most abstract reasoning problems, Claude Opus 4.7 is still the better call.
Precise retrieval from very long documents is the other gap. On the 500-page retrieval test, Claude Opus 4.7 scores 92.9 versus 83.5 for V4 Pro. The gap narrowed from V3.2, but it is still a gap.
| Feature | DeepSeek V4 Pro | Claude Opus 4.7 |
|---|---|---|
| Input price per 1M tokens | $1.74 | $5.00 |
| Output price per 1M tokens | $3.48 | $25.00 |
| Context window | 1M tokens | 1M tokens |
| Open weights | Yes | No |
| LiveCodeBench | Leader | Behind |
| 500-page retrieval | 83.5 | 92.9 |
| Humanities Last Exam | 37.7 | Behind Gemini 3.1 Pro at 44.4 |
The swarm agents math is where it gets interesting
Here is the scenario that rewires everything: you pay Claude Opus 4.7 $25 for a million output tokens. You pay V4 Pro $3.48 for the same. That $22 difference is not just a saving. It is a budget unlock. You can run six V4 Pro instances in parallel on the same prompt, keep the best answer, and still pay less than a single Opus call.
Several 2025 papers showed this statistically outperforms a single call to a more powerful model on a wide range of tasks. Cheap models voting in parallel beat expensive models reasoning alone. V3.2 could already do this, but the economics were tighter. With V4 Pro the math gets aggressive.
If your monthly API bill sits above $200 and most of the workload is code generation, batching, or multi-agent orchestration, it is worth migrating part of the stack to V4 Pro via Open Router. Keep Claude Opus 4.7 for the hardest reasoning, long-document retrieval, and high-stakes strategic work. On everything else, the cost reduction can hit 10x.
The takeaway
V4 Pro is not going to replace Claude Opus 4.7 for every use case. It will not replace it for most users on complex reasoning or hundred-page retrieval. But for coding, for batching, for swarm agents, and for anyone who wants open weights they can run themselves, the value is hard to argue with.
The real test starts now. On paper V4 Pro looks like a coding leader at a fraction of the price. Whether it holds up under three weeks of real work, on real codebases with real architecture, is the question everyone will be answering over the next month.
One thing is already clear: running premium American models at $25 per million output tokens now has to be justified. Default-to-Claude is no longer the obvious play. For a lot of workloads, the smart move is to deploy V4 Pro in production and keep Claude Opus 4.7 on the bench for when the task genuinely needs it.
