DeepSeek V4 Pro Challenges Claude With Agent Gains, Pricier API

DeepSeek's flagship leaves preview

DeepSeek has released the finished version of its top model. On August 13, the Hangzhou company made V4-Pro generally available in its app, on the web and through its API, after the model had been in preview since April1. The release build is called DeepSeek-V4-Pro-0813, and DeepSeek built it around agent work: the model calls tools, runs code and carries out multi-step jobs with little human supervision13. DeepSeek's own announcement highlighted bigger agent improvements, adjustable reasoning effort and native support for OpenAI's Responses API, with one-click setup for Codex5.

The headline numbers are large. V4-Pro is a sparse mixture-of-experts model with 1.6 trillion parameters, of which 49 billion are active for each token. It accepts up to 1 million tokens of context and can return as many as 384,000 tokens in one response2. DeepSeek reported scores of 87.9 on Terminal Bench 2.1, 62.7 on DeepSWE and 61.5 on NL2Repo1. The DeepSWE result stands out because the April preview scored only 12.8 on the same software-engineering test3. The weights for the 0813 build are now on Hugging Face under the MIT license2.

DeepSeek is openly setting the model against Anthropic. One report says the company pointed to Anthropic's Claude Fable 5 as only about 5% ahead on key benchmarks while costing roughly 45 times more per token24. That claim is the center of the launch, and it holds up only in part.

A four-month road to general availability

The V4 family first appeared on April 24 as a preview of two models: V4-Pro and a smaller V4-Flash. Flash has 284 billion total parameters, with 13 billion active per token1215. DeepSeek said at the time that Pro could rival the best closed-source models. Its own technical report was more cautious, saying V4 fell slightly short of GPT-5.4 and Gemini 3.1 Pro and trailed the frontier by roughly three to six months12. Several outlets noted that the preview landed only hours after OpenAI released GPT-5.517.

The model's efficiency comes from a hybrid attention design that pairs Compressed Sparse Attention with Heavily Compressed Attention. DeepSeek says this cuts memory needs by a factor of 9.5 to 13.7 compared with V3.215. The company also published a 58-page technical paper describing other changes, including manifold-constrained hyper-connections and the Muon optimizer17.

The path to the August release was uneven. V4-Flash got its official open-weight release on July 31, and independent tests found it beating the April Pro preview in several cases. That was awkward for a product sold as the premium tier14. The finished Pro build fixed that: Artificial Analysis gave the reasoning version of V4-Pro a 53 on its Intelligence Index, compared with 40 for Flash4.

The accounts of the timeline don't all match. One explainer says the V4 family reached general availability on July 208. Reuters-sourced coverage, DeepSeek's own post and Chinese state media all put the Pro launch in mid-August135. The mid-August date has the stronger support.

How it stacks up against Claude

Comparisons with Anthropic have followed V4 since April, and they keep landing in the same place. DeepSeek wins on price, speed and openness. Claude wins on hard, repository-scale engineering and general reasoning.

On coding, the results depend on the benchmark. Head-to-head tables report V4-Pro scoring 80.6% on SWE-bench Verified and 55.4% on SWE-bench Pro, behind Claude Opus 4.8's figures in the high 80s and mid-60s2526. V4-Pro leads on LiveCodeBench, with 93.5 against an estimated 88.8 for Opus, and has a Codeforces rating of 3,206, a number Anthropic does not publish27. An April comparison with Opus 4.7 found Terminal-Bench 2.0 close, at 67.9% for DeepSeek and 69.4% for Anthropic. DeepSeek led agentic web search with 83.4% on BrowseComp against 79.3%21. In practice, developers are routing work by task type: Claude for large refactors across many files, DeepSeek for single-file and algorithmic work where its cost edge matters most25.

The general-intelligence picture is less favorable, and here the sources disagree more sharply. The roughly-5% claim aimed at Fable 5 sits awkwardly next to the same report's index scores of about 59.9 for Fable and 44.3 for V4-Pro24. A late-September comparison, using Artificial Analysis Index v4.3.2, scored V4-Pro 0813 at 36.0 and Anthropic's newer Claude Opus 5.5 at 57.6, a gap of more than 21 points28. The same analysis noted that DeepSeek's technical report compares V4-Pro with Claude Opus 4.6, a model Anthropic has since replaced several times28.

The likely reading is that DeepSeek's "near-parity" framing depends on choosing benchmarks and comparing against older rivals. On broad independent indices the gap to Anthropic's current flagship is large, and it widened when Opus 5.5 arrived on September 2228. One independent 38-task evaluation from June did place V4-Pro in the same analytical tier as Opus 4.7, scoring 8.27 against 8.7222. So the model is competitive. It is just not equal.

Pricing: cheap, but no longer rock-bottom

Price has always been DeepSeek's main argument, and the Pro launch changed it in a way the company's marketing played down. At the April preview, V4-Pro cost $3.48 per million output tokens, compared with $25 for Anthropic's Opus and $30 for OpenAI's GPT-5.41216. In May, DeepSeek made a 75% discount permanent1418. That brought Pro down to about $0.435 for input and $0.87 for output9.

With the general-availability release, prices went back up. New rates took effect at 16:00 UTC on August 16, along with peak and off-peak billing, where off-peak costs half the peak rate15. Artificial Analysis lists the 0813 build at $1.32 per million input tokens and $3.96 per million output tokens. That is about 9 times Flash's input price and 14 times its output price4. Off-peak, that works out to $0.66 and $1.9827. Chinese state media described the pricing as still relatively low, quoting yuan rates3. Reuters-sourced coverage called it a reversal of DeepSeek's earlier strategy1.

Both descriptions hold up. Even at peak rates, V4-Pro's output costs a fraction of the $25 per million Anthropic charges for Opus27. Measured per task of an intelligence benchmark, one analysis put V4-Pro at $0.67 and Opus 5.5 at $5.9828. But the price rise shows that DeepSeek now wants Pro to earn a premium and is no longer pricing it to lose money for market share.

The chip story underneath

The pricing makes more sense alongside DeepSeek's hardware. The company explicitly linked V4-Pro's economics to Huawei's Ascend processors. It warned at launch that Pro's throughput was limited by scarce high-end compute and said prices should fall once Ascend 950 supernodes ship at scale in the second half of 20261719. Huawei announced "day zero" support across its Ascend SuperNode line13. Reuters reported that ByteDance, Tencent and Alibaba then approached Huawei about Ascend 950 orders, partly because the 950PR is the only domestic chip that supports the compressed number format V4 needs to run efficiently19.

Coverage has been loose about which models were trained on which chips. Fortune reported that DeepSeek said it used Huawei's Ascend processors to train the new model12. The V4 paper describes validating its expert-parallel scheme on both Nvidia GPUs and Ascend NPUs15. The best-supported position is that Ascend matters most for inference, and that a full break from Nvidia is not yet proven.

Geopolitics have also shadowed the launch. The April preview arrived on the same day that Reuters reported a U.S. State Department cable warning foreign governments about alleged intellectual-property theft by Chinese AI firms. China dismissed the accusations as groundless16.

What it means for the model race

Two points stand out. First, V4-Pro shows that an open-weight model can now handle production agent work. Its API accepts OpenAI ChatCompletions, Anthropic Messages and DeepSeek's own Responses format, so teams can switch by changing a base URL2. Developers have plugged it into Claude Code since the April preview29.

Second, DeepSeek's own product line moves faster than its flagship. V4.1 Flash replaced V4 Flash on the API on September 102, and one benchmark shows it scoring 39.5 against V4-Pro's 36.0 at about a quarter of the cost per task28. A V4.1 Pro has been named but not yet announced2. V4-Pro's challenge to Anthropic is therefore partly undercut by DeepSeek's own cheaper model.

The conclusion: V4-Pro does not overtake Claude. It makes Anthropic's premium harder to justify for coding, search and high-volume agent work, while Anthropic stays clearly ahead on the hardest reasoning and large-scale engineering tasks.