Analysts: open models like DeepSeek structurally challenge premium closed-lab pricing
International Business Times ● Covered by 2 sources
DeepSeek is undercutting OpenAI and Google on price, and sometimes beating them on coding. But outages, chip limits and weaker reasoning still leave it behind the top U.S. labs.
Based on reporting by International Business Times — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
DeepSeek has gone from market shock to real pressure point. Nearly two years after its cheap chatbot rattled global markets, the Hangzhou company is now being treated less like a curiosity and more like a competitor that can force the big U.S. labs to defend their prices. That said, the gap between “serious challenger” and “full-scale rival” still shows up in the numbers, in the hardware, and in the reliability of the service itself.
The strongest case for DeepSeek is coding. Its flagship V4-Pro, released as a preview on April 24, 2026 and made generally available on Aug. 13, is a 1.6 trillion parameter model released under the MIT open-source license. On DeepSeek’s own benchmarks, it outscored OpenAI, Google and Anthropic on Codeforces and LiveCodeBench, and it tied Gemini 3.1 Pro on SWE-bench Verified, the closely watched software engineering test.
But the same scorecards also show where it falls short. On Humanity’s Last Exam, a hard reasoning benchmark, V4-Pro trailed Claude, GPT-5.4 and Gemini 3.1 Pro. It also came in behind Gemini on GPQA Diamond, behind GPT-5.5 on Terminal-Bench 2.0, and behind Claude on a long-context retrieval test. Analysts have also pointed out that some of DeepSeek’s comparisons were made against older rival models, not the newest ones.
Independent checks have been less flattering, too. One coding benchmark that tested 24 models on the same project put Claude Opus 4.7 well ahead, while DeepSeek V4-Pro reached 89 only when run through a third-party tool. Smaller DeepSeek models have also landed behind rivals in a recent benchmark ranking. And while U.S. labs keep shipping new systems through 2026, DeepSeek is chasing targets that keep moving.
Price is where the company really bites. V4-Pro has been priced at about $3.48 per million tokens, versus about $25 for Claude’s top model, and analysts say DeepSeek charges roughly 11 times less than Gemini on input pricing while matching it on SWE-bench Verified. Combined with open weights that customers can run on their own hardware, that makes DeepSeek attractive to companies that want capable AI without paying premium closed-lab rates.
The bigger problem is compute. U.S. export controls have cut Chinese access to Nvidia’s most advanced chips, and DeepSeek says it plans to lean on Huawei hardware instead. Liang Wenfeng has said a model on OpenAI’s largest scale would need either about 50,000 Nvidia GB300 chips or roughly 200,000 Huawei Ascend 950 chips. The company is preparing to deploy at least 160,000 Huawei Ascend accelerators at a new data center in Inner Mongolia, which tells you how much the chip question matters.
Then there’s the messier stuff: 12 incidents in the past 30 days, a partial outage this week, and ongoing government concerns about data security. DeepSeek can clearly pressure the market on price, and sometimes on coding performance, but Western buyers still have good reasons to hesitate. The open model play is doing its job; the infrastructure around it is doing a much worse one.
My take — AI-written commentary, not fact-checked reporting
This is what closed-model pricing always feared: not a miracle, just a cheaper model that is good enough more often than the sales pitch can survive. The real moat was never “best model forever”; it was billing by the ounce and hoping nobody noticed. DeepSeek is making that harder to sell, which is exactly why the premium crowd looks so nervous.
Read more about this at: International Business Times
