DeepSeek-V4.1-Flash-1017 is being presented as more capable at coding and project-building, while costing less to use on some tasks. A new report also describes Search and Fetch interfaces for developers, giving AI agents a way to find web pages and pull their contents.
DeepSeek Watch analysis
What happened
In a report published on 19 October, 36Kr says DeepSeek’s updated model improves on several evaluations and can complete some tasks with fewer tokens. It cites Terminal-Bench 4.0 rising from 31.2 to 42.3, and estimates that the average cost of a DeepSWE evaluation task fell from about $0.49 to $0.20. The model’s API token prices are said to be unchanged, so the reported saving comes from using fewer tokens to finish a task.
The report also describes two developer-facing interfaces: Search finds web pages and returns titles, links and summaries; Fetch retrieves a selected page and organises its text for a model. It gives a price of RMB 5 per 1,000 successful requests for each, with failed requests, empty results and crawl timeouts free. 36Kr also presents hands-on examples of the model building interactive projects and audiovisual web experiences. Read the 36Kr report.
Why it matters
For developers building agents, fewer tokens can mean a lower bill across repeated tool calls, not just a shorter final answer. Search and Fetch could also make it easier to give an agent current web material without building those retrieval steps from scratch. The numbers and demonstrations are reported results, not a guarantee that every workload will see the same improvement.
Our read
The useful story is not simply that a model has a new suffix. It is the combination of reported task-cost savings and separately priced tools for finding and reading web material. Developers should compare the model on their own jobs and count the whole bill, including tool use. An impressive demo is a starting point; a lower invoice is the better encore.
What to watch
- Whether DeepSeek publishes further detail on the evaluation methods and cost estimates.
- How Search and Fetch perform on developers’ own sites, languages and agent workflows.
- Whether the reported gains hold up across ordinary projects, not only selected demonstrations.
Discussion spark: For an AI agent, what matters more in practice: a stronger coding score, a lower cost per task, or reliable built-in access to web information?
Sources and evidence
- DeepSeek No Minor Updates: Hands-On Test of New Model Reveals Revolutionary "Wish-Based Programming" Feature – 36Kr (19 October 2026, 04:33 UTC)
not affiliated with or endorsed by DeepSeek