DeepSeek Flash undercuts closed labs on price and search
DeepSeek-v4.1-Flash-1017 got the top comment slot with people doubting the benchmarks. The release is billed as "a minor update" but shows a substantial improvement, and it sits above GLM 5.3. The related thread, DeepSeek search and fetch at $0.75 per 1000 successful requests, drew a sharper reaction. One user said they use it daily and that it is "neck and neck with SOTA models," with search as the differentiator. Another noted Exa charges around $4 for the same thing.
The pattern here: the pricing pressure has moved from raw tokens to the tools around the model. Search, fetch and browsing are becoming bundled commodities. The comment that it "came at a bad time before the Anthropic IPO" captures the mood. Closed labs now have to defend their margins against an open-weight competitor that is both good and cheap.
The open question is reliability. Someone asked how the fetch API handles Cloudflare checkboxes and Anubis, and nobody had an answer. Cheap only matters if it works on the real, defended web.
So what?
If you pay per call for search or agent tooling, price a DeepSeek-backed alternative this week. If you resell a frontier model with a thin wrapper, assume your input costs will keep falling and your customers will notice.