DeepSeek V4.1 Flash-1017 lifts Terminal-Bench 4.0 from 31.2 to 42.3 at the same $0.30 price
DeepSeek updates V4.1 Flash to DeepSeek-V4.1-Flash-1017, with MIT-licensed weights and big gains on its own agentic coding benchmarks at an unchanged $0.30 in and $1.20 out per million tokens, and opens its in-house web search and fetch API at $0.75 per 1,000 calls.
- DeepSeek released DeepSeek-V4.1-Flash-1017, an update to V4.1 Flash with MIT-licensed weights; the
deepseek-flashAPI now serves it at an unchanged $0.30 in and $1.20 out per million tokens at peak, half that off-peak. - On DeepSeek’s own numbers, Terminal-Bench 4.0 rises from 31.2 to 42.3 and ProgramBench from 20.3 to 29.1, and it finishes tasks faster with shorter outputs.
- DeepSeek’s table puts it ahead of GPT-5.6 Sol on 9 of 11 rows but behind Claude Opus 5 on 7, including Terminal-Bench, ProgramBench, NL2Repo and HLE, and it doesn’t compare against the newer Opus 5.5 or GPT-6.
- DeepSeek also opened its in-house web search and fetch API at $0.75 per 1,000 calls each, free through its own harness, Claude Code and Codex until 1 November, Beijing time.
DeepSeek’s cheap open model gets a free upgrade
DeepSeek released DeepSeek-V4.1-Flash-1017 on 17 October, a “minor update” to V4.1 Flash that it says improves reasoning, coding, agentic and visual benchmarks while cutting task time and output length. The weights are on Hugging Face under the MIT license, with a 1M-token context. The API change log says deepseek-flash now points to it and “the API calling method and pricing remain unchanged”: $0.30 in and $1.20 out per million tokens at peak (01:00–04:00 and 06:00–10:00 UTC on weekdays), and half that at other times.
Why it matters. If you already call deepseek-flash, you’re on the new model now, so re-run your evals. The old deepseek-v4-flash names also route to it. DeepSeek’s vendor numbers show the biggest gains on agentic coding: Terminal-Bench 4.0 rises from 31.2 to 42.3, ProgramBench from 20.3 to 29.1 and Apex Bench from 64.6 to 72.1. The release ships no Jinja chat template; DeepSeek points to a Python reference encoder and deepseek-recipe, Rust libraries that convert Chat Completions, Responses and Messages requests into its prompt format.
| DeepSeek’s table | V4.1 Flash-1017 | V4.1 Flash | GLM-5.3 | Claude Opus 5 |
|---|---|---|---|---|
| Terminal-Bench 4.0 | 42.3 | 31.2 | 37.9 | 51.8 |
| DeepSWE v1.1 | 77.0 | 74.2 | 66.9 | 74.0 |
| ProgramBench | 29.1 | 20.3 | 19.0 | 37.0 |
| HLE, no tools | 42.6 | 36.8 | 42.0 (text only) | 56.3 |
Scores from DeepSeek’s model card, not independently checked. Highlighted values are the best in each row.
The bar for cheap open models just moved again. DeepSeek’s table beats GLM-5.3 and Kimi K3 on every agentic row, at close to the price of yesterday’s Reflection Beam, which already trailed the old V4.1 Flash. The comparison stops at Opus 5 and GPT-5.6 Sol, not current flagships, so treat it as the best open coding model per dollar on DeepSeek’s word until independent scores land. Prices are on our models table.
Quick hits
- DeepSeek opens its in-house search API: web search and fetch backed by an index DeepSeek says covers “hundreds of billions of web pages”, at $0.75 per 1,000 successful calls each; server-side calls through DeepSeek Harness, Claude Code and Codex are free until billing starts on 1 November (Beijing time).