DeepSeek V4.1 Flash hands-on tests and comparisons

A Japanese poster suggests putting Grok Bot at the top, letting Claude Code split tasks per project, and delegating execution to Devin SWE-2 and free OpenCode models.

Despite being labeled a minor update, V4.1-Flash-1017 reportedly outperforms GPT-5.6 Sol and Opus 5 on several published agent and coding benchmarks with shorter outputs and faster completion.

Hands-on speed reaches 244 tok/s and nearly 300 at peak, faster than GPT-6.1 Sol's Ultrafast mode with intelligence near GPT-5.6 Sol; it adds built-in web research, is free on the official harness until Oct 31, and shows clear coding and 3D gains.

The author made V4.1-Flash-1017 the default in Hermes for all-day recurring tasks and research, citing off-peak pricing of $0.15 input / $0.60 output per million tokens, and plans to report actual token usage and cost per task.

The poster says the model appears to have shown up on the API platform before any announcement and asks whether anyone has called it and how speed, price and coding perform; no test data yet.

AIHubMix ran both versions on the same prompt at max reasoning: the older build took ~20 min and $0.50, the 1017 build ~17 min and $0.46.

A user reports the overseas V4.1-Flash gives roughly 100M free tokens per day through month-end, and notes DeepSeek opened its own search API at about ¥5 per 1,000 queries.

The author focuses on agent cost and latency: DeepSeek reports shorter outputs, faster task completion and unchanged API pricing, with Terminal-Bench 4.0 at 42.3 and DeepSWE v1.1 at 77.0, though scores depend on the harness.

Highlights from the release: DeepSWE improves from 74.2 to 77.0, outputs are shorter and less verbose, and official weight downloads are available.

The post notes that DeepSeek-V4.1-flash-1017 continues to ship open source, suggesting new regulations do not restrict open-weight models.

The post cites 77% on DeepSWE versus Opus 5's 74 and GPT-5.6-Sol's 73, Terminal-Bench jumping 31 to 42, and about $0.20 per task against roughly $12 for Opus 5 and $22 for Fable 5.

The official model card shows leads on DeepSWE (77.0 vs 74.0) and Agents' Last Exam (35.4 vs 28.6) but deficits on Terminal-Bench (42.3 vs 51.8) and HLE (42.6 vs 56.3); shorter outputs cut per-task cost.