SHENZHEN, Oct. 19 — A Chinese artificial-intelligence startup has released a new model it says leads Anthropic’s Claude Opus 5 on several agentic benchmarks while running at a fraction of the cost — the latest sign that the global large-model race is shifting from raw capability toward price-performance.
Asymmetric Architecture and Benchmark Victories DeepSeek said its DeepSeek-V4.1-Flash-1017, released on October 17, leads Claude Opus 5 on four of 11 published benchmarks, including DeepSWE v1.1 (77.0 versus 74.0), Automation-Bench (60.5 versus 50.3), Agents’ Last Exam (35.4 versus 28.6) and Apex Bench (72.1 versus 65.8). On several other benchmarks, including Terminal-Bench 4.0 and HLE, the Anthropic model retains an advantage. The model is a sparse mixture-of-experts system with a 552-billion-parameter backbone that activates about 8 billion parameters on input and 16 billion on output, an asymmetric design intended to keep per-token compute low.
The Economics of Inference DeepSeek said per-task running costs come in at a fraction of comparable U.S. products. Artificial Analysis data placed DeepSeek V4.1 Flash (Max) at $0.27 per task, against $1.10 for Claude Opus 5 (Low). The launch is a fine-tuned upgrade of the DeepSeek-V4.1-Flash model released in September, with API pricing unchanged.
From Capability Race to Profit Race The release lands as the industry pivots from a “capability race” to a “profit race,” with developers under mounting pressure to deliver models that are not only powerful but economically viable at scale. DeepSeek first drew international attention in early 2025 with a chatbot far cheaper to develop and operate than offerings from U.S. technology giants, unsettling assumptions about the cost curve of frontier AI.
Analysts note that the claim of surpassing a leading closed model on agentic tasks — where systems plan and act across multiple steps — has not yet been independently verified. DeepSeek’s benchmark figures are company-reported, and external evaluations are still pending. If confirmed, the result would mark a notable narrowing of the gap between low-cost and high-cost AI development.