# DeepSeek Ships V4 Pro, Rivaling Claude and GPT-5 for a Fraction of the Cost > Source: > Published: 2026-08-13 12:49:49+00:00 *DeepSeek V4 Pro is real, cheap, and strong on coding, but the clean story is not a secret August launch. It is a price-pressure story built on open weights, a 1 million token window, and benchmark claims that closed labs now have to answer.* DeepSeek already made its big V4 move on April 24, when it released the V4 family as a preview and put the weights for V4 Pro on Hugging Face. No mystery is needed here. The company shipped a 1.6 trillion parameter mixture-of-experts model, with 49 billion parameters active for each token, and gave it a 1 million token context window. If you're building coding agents, that combination is the part you should care about. The price gap is the story. DeepSeek's own API pricing page lists V4 Pro at $0.435 per million input tokens on a cache miss and $0.87 per million output tokens, with a far lower cache-hit price. Anthropic's pricing page lists Claude Opus 4.6 and Opus 4.7 at $5 per million input tokens and $25 per million output tokens. The math is blunt. On output tokens alone, DeepSeek is charging less than four cents on the dollar compared with Opus. That doesn't mean V4 Pro is the better model for every job, and frankly, any team that treats one benchmark table as a buying decision deserves the bill it gets. But it does mean the default question changes. You can't justify a frontier-model invoice by habit once an open-weight rival is close enough on the work your agents actually do. ## The benchmarks are strong, but not magic DeepSeek's Hugging Face model card reports V4 Pro at 80.6% on SWE-bench Verified, 93.5 on LiveCodeBench and a 3,206 Codeforces rating. Those are DeepSeek-reported results, so you should test them on your own task set before moving production traffic. Still, the comparison is not trivial. The same model card places Claude Opus 4.6 at 80.8% on SWE-bench Verified, Gemini 3.1 Pro at 80.6%, and GPT-5.4 at a 3,168 Codeforces rating. That's enough to make procurement uncomfortable. A model does not have to beat Claude or GPT across the board to hurt their pricing power. It only has to take the routine runs: repo scanning, test-log triage, issue reproduction, code review drafts, dependency upgrades. Those jobs burn tokens all day. DeepSeek's architecture explains part of the bill. A mixture-of-experts system carries a huge pool of weights but routes each token through only a slice of them. V4 Pro's 49 billion active parameters are still large, but they are not 1.6 trillion active parameters. That is the difference between a headline number and an operating cost. ## The cheap lane may not stay this cheap There is a catch. According to Reuters, DeepSeek said in May that it would make a 75% cut to V4 Pro API pricing permanent after the promotion ended on May 31, bringing prices down to a quarter of the launch rate. South China Morning Post later reported that DeepSeek introduced peak-hour surcharges for V4 API access, doubling prices during busy Beijing-time windows to manage resources and improve service stability. Even after that, the gap is large. A doubled V4 Pro output price is still nowhere near $25 per million tokens. For teams running long agent loops, the question is not whether DeepSeek has ended the price debate. It has started one that incumbents cannot ignore. Open weights also change the negotiation. Any team with enough hardware can self-host V4 Pro from Hugging Face and keep traffic away from DeepSeek's hosted API. That option is not free, because serving a 1.6 trillion parameter model is serious infrastructure work. But the existence of that option gives larger customers a lever they do not have with closed models. The useful way to read DeepSeek V4 Pro is simple: not as a Claude killer, not as an OpenAI killer, and not as a miracle model. It is a cheap, open-weight coding model with enough benchmark strength to force a harder conversation about where premium tokens are actually worth spending. If your agent fails only when the task is genuinely hard, pay for the best model. If it is just chewing through context, logs and retries, you now have less excuse to pay top dollar for every pass. **Also read:** [Elon Musk Says SpaceX AI Revenue Will Overtake Rockets by September](https://startupfortune.com/elon-musk-says-spacex-ai-revenue-will-overtake-rockets-by-september/) • [India's Power Grid May Not Be Ready to Fuel the AI Data Center Boom](https://startupfortune.com/indias-power-grid-may-not-be-ready-to-fuel-the-ai-data-center-boom/) • [Cool CPI Data Pushes Tech Stocks to Records, Then Cisco's AI Guidance Rattles Investors](https://startupfortune.com/cool-cpi-data-pushes-tech-stocks-to-records-then-ciscos-ai-guidance-rattles-investors/)