DeepSeek V4 is here. Not a press release, not a carefully curated blog post with cherry-picked benchmarks, but a 58-page research paper with nothing held back. And the conclusion it forces is uncomfortable for anyone paying twenty dollars a month for a premium AI subscription: Claude is a waste of money.
This is not hyperbole. The numbers are public, and they are absurd. Depending on whether there is a discount or not, DeepSeek-V4 can be 30 times cheaper than Anthropic's Claude. Even without a discount, you are looking at 8 to 20 times less. That is not a small difference. That is the kind of difference that makes you stop and ask what exactly you are paying for.
Two Minute Papers has a nice breakdown in video format here.
The Benchmarks
If you were paying 30 times more for a clearly superior product, you could make a case for it, but you are not. Look at the numbers:
DeepSeek-V4-Pro on the maximum reasoning effort mode scores 90.2% on HLE, which is one of the hardest "trust me bro" benchmarks in existence. Claude Opus 4.6-Max? 89.1%. On Apex, DeepSeek hits 85.9% against Claude's 78.1%. On Codeforces rating, DeepSeek scores 3206, matching GPT-5.4 and leaving Claude in the dust. On long-context retrieval, DeepSeek's MMR score of 92.9 crushes Claude's 76.3. On CorpusQA accuracy at one million tokens, it is 71.7 against 53.8.
For the vast majority of real-world tasks, there is no meaningful gap. In several important categories, DeepSeek outright wins. Yet you pay a fraction of the cost, or nothing at all if you somehow manage to self-host all 1 trillion parameters. This is the first time an open model has matched or beaten closed frontier models at this breadth of tasks, and it happened faster than almost anyone predicted.
A Million Tokens for Free
A million-token context window used to be a flagship feature that justified enterprise pricing. Google made a whole product launch out of it with Gemini. I remember flipping out about it two years ago. Now DeepSeek hands it to you for free in open weights. Ask it to inhale 1,500 pages of dense documentation, and it will.
The engineering is worth understanding because it explains how this is even possible. The transformer's vanilla attention mechanism has quadratic complexity: double the context length, quadruple the compute. This is the bottleneck that makes long contexts ruinously expensive for most models. DeepSeek broke through it with a hybrid attention architecture combining Compressed Sparse Attention and Heavily Compressed Attention.
Think of it like reading a book. You cannot process every word simultaneously, so you summarise. DeepSeek does this at three levels:
Token-level compression: summarise each paragraph into a sentence. Keep the book. Search it faster.
Heavily compressed attention: look at the table of contents. If each chapter has a short name, you grasp the whole story at a glance. A 128-to-1 compression.
Compressed sparse attention: use an index. Searching for fights in a novel? The index gives you the top five pages. The model attends only to what matters.
Three layers: summaries, structure, index. Together they reduce KV-cache memory needs by approximately 90%. Squashing 100 words into a storage space of 10 without losing every piece of information. The benchmarks back it up: on MRCR, retrieval remains remarkably stable within 128K tokens and stays strong all the way to one million.
The Efficiency Leap
The previous DeepSeek-V3.2 was already efficient. DeepSeek-V4 is not an improvement on it. It is a different category of efficiency entirely. In a one-million-token context, DeepSeek-V4-Pro requires only 27% of the single-token FLOPs and 10% of the KV-cache compared with V3.2. The Flash model is more extreme still: 10% of the FLOPs and 7% of the KV-cache.
To translate: the Pro model needs about three times less computing power than the previous generation for the same output. Flash needs about ten times less. This is not incremental. This is a generational leap that reshapes the cost calculus of running large-scale AI inference. Any company relying on Claude or GPT for heavy inference workloads could achieve comparable results, results that match or beat the frontier, at a fraction of the operational cost, or by self-hosting entirely.
Two architectural innovations deserve specific mention because they are the kind of thing that makes you stop and reread the paper. Manifold-Constrained Hyper-Connections replace standard residual connections with something more mathematically principled: the residual mapping matrix is constrained to the manifold of doubly stochastic matrices, ensuring the spectral norm stays bounded by one. In plain terms, it stops signal from exploding or vanishing across deep stacks of layers. Muon, the new optimiser, replaces AdamW for most modules and converges faster with better stability. DeepSeek is not just scaling up, but are rethinking fundamentals.
What DeepSeek Lacks
It would be dishonest to pretend DeepSeek V4 has no weaknesses. It has two notable ones.
For one, it is unimodal: no images, no audio. It is blind and deaf, at least for now. If your workflow involves heavy image analysis or multimodal reasoning, Claude still has an edge, though it is hard to imagine that edge surviving the next twelve months. If you need that edge, maybe combine with Qwen.
The paper also admits something that is rare and genuinely refreshing: two techniques used to stabilise training, Anticipatory Routing and SwiGLU Clamping, work effectively, but the creators are not entirely sure why. This is not the polished corporate non-answer you get from most AI companies. This is a real research paper with real admissions of uncertainty. The transparency is admirable, but it does mean there are open questions about behaviour under certain edge cases.
Context window performance also degrades as you approach the limits. Models forget, drift, hallucinate. More text means less truth. This is not a DeepSeek problem, it is a universal one, but nevertheless worth knowing.
The Trajectory
The gap between open and closed models is not closing. It has closed. On knowledge benchmarks, DeepSeek-V4-Pro sets a new state-of-the-art for open models, surpassing all prior open-source baselines by a margin of 20 absolute percentage points on SimpleQA. On reasoning, it matches or beats the frontier. On code, it is competitive with GPT-5.4, which is the first open model to achieve this. On agents, it approaches Claude Opus 4.5. On long-context, it surpasses Gemini-3.1-Pro.
None of this was supposed to happen this fast. The conventional wisdom a year ago was that open models would trail the frontier by 12 to 18 months indefinitely. DeepSeek-V4 trails by perhaps 3 to 6 months in a few narrow categories and has pulled ahead in others. The proprietary incumbents are running out of road.
Conclusion
We are witnessing something unprecedented: frontier-level intelligence at a price that is approaching zero. A 1.6 trillion parameter model. A million tokens of context. Benchmark scores that match or beat the best proprietary systems. Open weights. Free to use, cheap to run, available to self-host. The phrase "too cheap to meter" used to be an aspiration. It is becoming a description.
If you are still paying for Claude without having tried DeepSeek-V4, you are not paying for quality. You are paying for inertia, for brand recognition, for American datacentres, and for the comfort of a familiar chat window. Those are not nothing, but they are not worth 30 times the price.
Top comments (0)