DeepSeek V4.1 Flash 1017 vs Claude Haiku 5.5 vs GPT-6 Luna vs Qwen3.8 Flash

DeepSeek V4.1 Flash 1017 is the October 17 update of V4.1 Flash, live on DeepSeek's website, app and API. DeepSeek describes it as a Pareto improvement: stronger benchmark results, faster task completion and shorter outputs than the September release. It was trained with reinforcement learning from self-feedback for end-to-end project work. It keeps V4.1's sparse mixture-of-experts architecture, native image input and 1M-token context, and the weights are open on Hugging Face.

A big jump for a point release (Artificial Analysis Intelligence Index):

• V4.1 Flash 1017: 45.1, up from 39.5 for the September V4.1 Flash
• That puts it ahead of every small model here: Claude Haiku 5.5 43.4 · Qwen3.8 Flash 39.8 · GPT-6 Luna 38.1
• For reference, it also beats GLM 5.3 (44.8) and Kimi K3 (43.6) and sits just behind Qwen3.8 Max 0902 (45.4)
• Output speed is 216 tokens/s, close to Haiku 5.5 (232) and well ahead of GPT-6 Luna (124) and Qwen3.8 Flash (53)

Pricing: DeepSeek's endpoint uses time-of-day rates. Off-peak it costs $0.15 / $0.60 per 1M tokens (cache hits $0.003). On weekdays during peak hours, 01:00–04:00 and 06:00–10:00 UTC, it costs $0.30 / $1.20 (cache hits $0.006). Haiku 5.5 and GPT-6 Luna list at a flat $0.10 / $0.50, and Qwen3.8 Flash at $0.15 / $0.47. Cache hits cost 50× less than uncached input at either rate, which makes 1017 hard to beat for agent loops that resend the same context.

Cost per task: on the Artificial Analysis index run, 1017 cost $0.26 per task, against $0.21 for Haiku 5.5, $0.37 for Qwen3.8 Flash and only $0.07 for GPT-6 Luna, which writes much shorter reasoning.

DeepSeek-V4.1-Flash-1017 release notes