DeepSeek updates its Flash model and opens its search engine

Answers are shorter now, and DeepSeek says the model also finishes work sooner and beats the last Flash on its benchmarks. Shorter answers mean a smaller bill. On DeepSWE, a coding benchmark, DeepSeek’s chart puts the new model at 77.0 per cent of tasks solved for just under $0.20 each, The Decoder reports. That compares with 74.2 per cent for the earlier version, at roughly $0.50 a task. DeepSeek’s own tables rank it above bigger rivals such as Claude Opus 5 and GPT 5.6-Sol in some coding and agent categories. Independent scores have not been published.

The lab also opened its in-house search engine to every developer. Two endpoints, Search and Fetch, cost $0.75 per 1,000 calls each. DeepSeek API users can make those calls free through DeepSeek Harness, Claude Code and Codex until 16:00 UTC on 31 October 2026, the release note says.

If your team runs coding agents at volume, an open model at a fraction of the per-task cost is one more option to price against the US labs.