DeepSeek has released V4.1-Flash-1017, an update to the Flash line, and the more interesting half of the announcement is not the model.
Search and Fetch
DeepSeek now offers two endpoints alongside the model. Search returns ranked web results, 1 to 50 per query. Fetch returns a single page as Markdown. Both cost $0.75 per 1,000 successful requests, and empty results or failures are not charged. Rate limit is 5 QPS per account on each.
Server-side calls made through DeepSeek Harness, Claude Code or Codex are free until 1 November 2026. If you are building a research agent, that is a window worth using.
Not charging for failed requests is a small thing that matters more than it sounds. Search APIs that bill on empty results make retry logic expensive and push you toward caching results you should not cache.
The model
552B backbone plus 196B Engram parameters, with 8B active per token during prefill and 16B during decode. Multimodal mixture-of-experts, 40 layers split into a 20-layer encoder and 20-layer decoder, 1M context with up to 384K output tokens. Text and image in, text out.
| Benchmark | 1017 | Previous | Claude Opus 5 |
|---|
| DeepSWE v1.1 | 77.0 | 74.2 | 74.0 |
| Terminal-Bench 4.0 | 42.3 | 31.2 | 51.8 |
| HLE (no tools) | 42.6 | 36.8 | 56.3 |
It improves on its predecessor across all eleven reported benchmarks by 2.0 to 11.1 points. Note the comparison column: these are DeepSeek's own figures, and on two of three Claude Opus 5 is ahead. Terminal-Bench 4.0 is a harder benchmark than v2.1 and the numbers do not transfer between versions.
API pricing is unchanged at $0.30 per million cache-miss input tokens and $1.20 output at peak, halving off-peak.
One caveat: the September V4.1-Flash weights were MIT-licensed, but the licence for the 1017 checkpoint has not been stated. Check before you build on it commercially.
Sources: MarkTechPost