DeepSeek ships V4.1-Flash-1017, claiming faster tasks and $0.2 coding runs

DeepSeek has shipped another quick iteration. On October 19 the Chinese AI company released V4.1-Flash-1017, an upgrade to the V4.1-Flash model it introduced a month earlier, pitching faster processing and a lower cost per task.

The release was reported by Maeil Business Newspaper's English edition on October 19, citing Global Times, Yicai and other outlets. According to the report, DeepSeek says the new model completes tasks faster while producing shorter answers — a deliberate shift away from the verbose output that inflates token bills in agent workloads.

On capability, DeepSeek says the upgrade improves reasoning, coding and AI agent functions, and extends to game development and 3D modelling. Those are precisely the areas where vendors have concentrated: agents demand reliable multi-step execution, coding demands accuracy over long context, and 3D and game assets test a model's control over structured output.

Price is where the claim gets sharpest. DeepSeek says a single coding task costs about $0.2, less than one-tenth of what competing U.S. models charge. Per million output tokens, the API is priced at $0.59 off-peak and $1.19 at peak; the report contrasts that with Anthropic's Claude Opus 4.5 at $25 and OpenAI's GPT-5.6 Sol at $20.

The performance comparison comes from DeepSeek's own published evaluation. On DeepSWE v1.1, a coding benchmark, the new model scored 77.0 against 74.0 for Claude Opus 4.5 and 73.0 for GPT-5.6 Sol. The report notes explicitly that these figures were released by DeepSeek and have not been independently reproduced.

Alongside the model, DeepSeek released an in-house API for search and content collection, charged at 5 yuan per 1,000 requests. Productising retrieval as a separate line item suggests the company wants to own the data pipeline that makes agent deployments slow and expensive, not just the inference bill.

A Chinese analyst quoted in the report said the upgrade shows competition in China's AI industry now centres on efficiency, specialisation and practical application. Bloomberg's reading, also cited, is that iterating only a month after the previous release signals an accelerating contest between U.S. and Chinese models.

Three things are worth watching: whether third parties can reproduce the 77.0 score; whether shorter answers translate into measurable savings inside real agent workflows; and how U.S. vendors respond to a price point an order of magnitude lower — with cuts of their own, or with claims about a capability gap.