DeepSeek V4.1-Flash-1017: Open Weights and $0.75 Search API
The short answer
DeepSeek V4.1-Flash-1017 is the October 17, 2026 refresh of DeepSeek’s cheap open-weight flagship: stronger at coding and agent tasks, shorter outputs, same price, and MIT weights on day one. It arrived with a second product that matters to agent builders: DeepSeek’s own Search and Fetch APIs at $0.75 per 1,000 requests each.
- Model name:
deepseek-flash(legacydeepseek-v4-flashnames now route to 1017). - Price (unchanged): off-peak $0.15 input / $0.003 cached / $0.60 output per million tokens; peak 2x. Details in our current API prices.
- Limits: 1M context, up to 384K output, thinking on by default, tool calls, JSON output, Responses API and Anthropic-format endpoints, vision.
- Weights: MIT licence, Hugging Face and ModelScope.
What changed in the model
DeepSeek’s model card compares 1017 with the September V4.1 Flash. All scores are vendor-reported:
| Benchmark | V4.1-Flash-1017 | V4.1 Flash (Sep) | Claude Opus 5 (DeepSeek’s run) |
|---|---|---|---|
| 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 |
| Automation-Bench v1.0.6 | 60.5 | 54.8 | 50.3 |
| ProgramBench | 29.1 | 20.3 | 37.0 |
DeepSeek used the Claude Code harness for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE. Note the comparison is against Opus 5, not the current Claude Opus 5.5. DeepSeek also says the update reduces task time and output length, which lowers cost per task at an unchanged token price — treat that as a claim to test on your own workload.
What the Search and Fetch APIs are
- Search: web results from DeepSeek’s in-house engine, which DeepSeek says serves hundreds of millions of requests a day.
- Fetch: retrieves a page’s content for the model to read.
- Price: $0.75 per 1,000 successful requests, per endpoint.
- Free window: server-side tool calls through DeepSeek Harness, Claude Code and Codex on the DeepSeek API are free until November 1, 2026, 00:00 Beijing time (16:00 UTC, October 31).
For an agent that searches and then reads two pages, that is about $2.25 per 1,000 tasks in tool fees. Compare that with the incumbents in best web search APIs for AI agents, and weigh data-residency: requests go to DeepSeek’s service.
Who should care
- Teams already on deepseek-flash: nothing to change; re-run your evals, because output lengths shift.
- Self-hosters: the 1017 weights replace the September checkpoint under the same MIT licence. There is no Jinja template — use DeepSeek’s
encodingreference ordeepseek-recipe. Serving engines are compared in best LLM inference engines. - Cost-sensitive coding agents: schedule batch work off-peak; peak (01:00–04:00 and 06:00–10:00 UTC on weekdays) doubles every line of the bill. Weekends and Chinese public holidays are off-peak all day.
How it compares with other low-cost models is covered in DeepSeek V4.1 Flash vs V4 Pro vs Gemini 3.8 Flash.
Last verified: October 20, 2026.