DeepSeek V4.1 Flash 1017

DeepSeek V4.1 Flash 1017, which DeepSeek released on October 17, 2026, is a minor update to September's V4.1 Flash that scores higher, finishes tasks faster and writes shorter answers at the same price, $0.15 / $0.60 per million tokens off-peak, with MIT-licensed weights on Hugging Face.

The benchmarks

DeepSeek compared the update with September's V4.1 Flash, Kimi K3, GLM-5.3, GPT-5.6 Sol and Claude Opus 5. Its table names no effort level, but its DeepSWE cost chart puts the 77% score at max effort (max on DeepSWE v1.1). Terminal-Bench 4.0 ran in Claude Code, DeepSWE in mini-SWE-agent v2 and ZeroBench as pass@5. The Humanity's Last Exam row is the full set (44.0 on the text-only subset), so GLM-5.3, which only has a text-only score, is left out, and blank cells in DeepSeek's table are left out too.

DeepSeek's numbers

• Humanity's Last Exam (no tools) (expert questions, full set): 42.6%. +5.8 over DeepSeek V4.1 Flash (36.8%), −0.9 under Kimi K3 (43.5%), −1.9 under GPT-5.6 Sol (44.5%), −13.7 under Claude Opus 5 (56.3%).

• Terminal-Bench 4.0 (multi-hour terminal work): 42.3%. +11.1 over DeepSeek V4.1 Flash (31.2%), +29.7 over Kimi K3 (12.6%), +4.4 over GLM-5.3 (37.9%), +2.4 over GPT-5.6 Sol (39.9%), −9.5 under Claude Opus 5 (51.8%).

• DeepSWE v1.1 (fixing real repositories): 77%. +2.8 over DeepSeek V4.1 Flash (74.2%), +9.5 over Kimi K3 (67.5%), +10.1 over GLM-5.3 (66.9%), +4.0 over GPT-5.6 Sol (73%), +3.0 over Claude Opus 5 (74%).

• ProgramBench (writing whole programs): 29.1%. +8.8 over DeepSeek V4.1 Flash (20.3%), +11.6 over Kimi K3 (17.5%), +10.1 over GLM-5.3 (19%), +6.1 over GPT-5.6 Sol (23%), −7.9 under Claude Opus 5 (37%).

• NL2Repo (building a repo from a spec): 68.2%. +2.8 over DeepSeek V4.1 Flash (65.4%), +10.2 over Kimi K3 (58%), +10.2 over GLM-5.3 (58%), +11.4 over GPT-5.6 Sol (56.8%), −7.1 under Claude Opus 5 (75.3%).

• AutomationBench v1.0.6 (business automations): 60.5%. +5.7 over DeepSeek V4.1 Flash (54.8%), +13.8 over Kimi K3 (46.7%), +11.7 over GLM-5.3 (48.8%), +14.7 over GPT-5.6 Sol (45.8%), +10.2 over Claude Opus 5 (50.3%).

• Agents' Last Exam (real professional tasks): 35.4%. +3.6 over DeepSeek V4.1 Flash (31.8%), +7.8 over Kimi K3 (27.6%), +6.9 over GLM-5.3 (28.5%), +8.7 over GPT-5.6 Sol (26.7%), +6.8 over Claude Opus 5 (28.6%).

• Apex Bench v1.1 (agent tasks): 72.1%. +7.5 over DeepSeek V4.1 Flash (64.6%), +24.3 over Kimi K3 (47.8%), +15.5 over GLM-5.3 (56.6%), +17.8 over GPT-5.6 Sol (54.3%), +6.3 over Claude Opus 5 (65.8%).

• Chartography (with tools) (reading charts): 83%. +4.1 over DeepSeek V4.1 Flash (78.9%), +14.9 over Kimi K3 (68.1%), +3.1 over GPT-5.6 Sol (79.9%), −1.0 under Claude Opus 5 (84%).

• BabyVision (with tools) (basic visual puzzles): 92.5%. +2.9 over DeepSeek V4.1 Flash (89.6%), +6.8 over Kimi K3 (85.7%), +3.6 over GPT-5.6 Sol (88.9%), −1.6 under Claude Opus 5 (94.1%).

• ZeroBench (with tools) (hard visual reasoning): 51%. +2.0 over DeepSeek V4.1 Flash (49%), +10.0 over Kimi K3 (41%), −2.0 under GPT-5.6 Sol (53%), −1.0 under Claude Opus 5 (52%).

Independent numbers

• Artificial Analysis Intelligence Index (current v4.3.2): 45.1. +5.6 over DeepSeek V4.1 Flash (39.5), +2.4 over MiniMax M3.1 Flash (42.7), +1.5 over Kimi K3 (43.6), −1.2 under MiMo-V2.6-Pro (46.3), −5.7 under Claude Opus 5 (50.8).

First impressions

The update is three days old. Reactions come from X and the Hacker News thread on its Artificial Analysis results.

The positives

Arena (@arena), which ranks models on blind human votes, put it at the top of the open models in Code Arena's web development board.

That's people picking its websites over Opus 5 (High) without knowing which model made them, which is a better check than DeepSeek's own table.

techjamie on Hacker News spent an afternoon building with it in the Pi coding agent.

I was doing some fairly heavy development on Flash 1017 yesterday for around ~3 hours on pi, and all told that session ran about $0.40. (comment)

They added that Claude might have needed less hand-holding, but forty cents for three hours of agent work is the whole pitch.

VulKan (@VulKan42069), a general X user, ran the same one-shot build on it and on Claude Opus 5.5.

One test, but the gap is the story: less than half the time and about 2% of the bill, even though it used more tokens.

The negatives

Lisan al Gaib (@scaling01), who made LisanBench, says DeepSeek trained on his benchmark.

If one public test leaked into training, it's fair to ask about the others, which is why the independent scores matter more than DeepSeek's table.

Tiberium on Hacker News read the Artificial Analysis numbers less kindly.

Somehow GPT 6.1 Sol (medium) is cheaper per task ($0.21 vs $0.26) and faster per task (2.9 vs 4.9 minutes) vs v4.1 Flash 1017 (max). (comment)

Cheap tokens don't guarantee a cheap task when max effort thinks for a long time, and other commenters said lower effort settings save most of the cost for little loss.

Alexandru Indrei (@aindrei), a general X user, found it worse at reading images than the September version.

DeepSeek's own vision scores went up, so this is one person's experience, but test it on your own images before switching.