# Research|LLM: DeepSeek V4.1-Flash-1017 Reaches the Top Tier of Open-Weight Models, Still Behind Proprietary Frontrunners > Source: > Published: 2026-10-19 12:25:20+00:00 DeepSeek released V4.1-Flash-1017 on October 17, an update to the V4.1 Flash model launched on September 10. The API model name is unchanged. Pricing remains $0.30 per million input tokens and $1.20 per million output tokens, with a 50% off-peak discount. The weights are available under the MIT license. The new version outperforms its predecessor on all 11 benchmarks reported by DeepSeek: its Terminal-Bench 4.0 score increased from 31.2 to 42.3 and its DeepSWE score from 74.2 to 77.0. In our view, it now offers the best balance of capability and cost among open-weight models for coding and agent tasks. Since the release, some articles have claimed that 1017 beats GPT-6, mainly on the strength of its DeepSWE score and cost per task. **Our preliminary view is that 1017 has joined Kimi K3 and Xiaomi’s MiMo-V2.6 Pro in the top tier of open-weight models.** DeepSeek has made clear progress, but this update has not brought the open-weight frontier any closer to the leading proprietary models. Our channel checks indicate that data from the previous generation of proprietary models remains accessible. Distilling from newer models, particularly Fable 5.1 and Opus 5.5, has become more difficult. Opus 5.5 retains a clear lead, and 1017 has not narrowed the gap. OpenAI also has an even stronger internal model, Bel. ### **Joins the top open-weight tier, but trails proprietary leaders overall** DeepSeek’s model card compares 1017 with Kimi K3, GLM-5.3, GPT-5.6 Sol and Opus 5. It does not include Opus 5.5, GPT-6, Fable 5.1 or Gemini 4. **DeepSeek’s published benchmark comparison for V4.1-Flash-1017** According to DeepSeek’s results, 1017 outperforms Kimi K3 and GLM-5.3 on almost every benchmark and beats GPT-5.6 Sol on 9 of 11, trailing only on HLE and ZeroBench. It leads Opus 5 on four benchmarks and trails on seven. On HLE and Terminal-Bench 4.0, it trails by 13.7 and 9.5 percentage points, respectively. The claim that 1017 beats GPT-6 rests mainly on DeepSWE. DeepSeek reports 77.0 for 1017, while Datacurve’s official leaderboard shows 74.1 for GPT-6 Astra. The two scores come from different sources and do not establish a lead. In Mercor’s independent testing, the September version, Astra, GPT-6.1 Sol and Opus 5.5 all scored around 72, with very little between them. Public Terminal-Bench 4.0 results suggest 1017 is still well behind the leading proprietary models. DeepSeek reports 42.3 for 1017, while Google’s launch page lists 58.2 for Astra and 66.4 for Opus 5.5. Artificial Analysis (AA) measured 59% for Astra and 60% for Opus 5.5 in independent testing but has not yet tested 1017, so the size of the gap still needs to be confirmed in a like-for-like test. As of October 19, AA had not published an Intelligence Index score for 1017; the September version scored 39. The leading open-weight models are clustered between 44 and 46, with MiMo-V2.6 Pro at 46, GLM-5.3 at 45 and Kimi K3 at 44. Our preliminary view is that 1017 has moved into this group, pending independent testing. For comparison, Opus 5.5 scores 58, five points above GPT-6 Astra, Gemini 4 Argon and Fable 5.1 and 12 points above the best open-weight score to date. Haiku 5.5, Anthropic’s small model released in October, offers another useful comparison, with an AA Intelligence Index score of 43. V4.1 Flash scored 69% on AutomationBench-AA versus 35% for Haiku 5.5. Haiku 5.5 performed better on Terminal-Bench 4.0, HLE and factual accuracy. This comparison uses the September release; 1017 has yet to be tested independently. **Compared with proprietary flagships, 1017’s main advantage is cost.** DeepSeek’s DeepSWE cost chart puts the new version at under $0.20 per task at its highest reasoning setting. Separately, AA measured a weighted average cost of $0.27 per Intelligence Index task for the September version. In the same AA tests, Astra cost roughly 12 times as much, and Opus 5.5 at its highest reasoning setting roughly 22 times as much. The DeepSWE and AA figures are not directly comparable because the tasks and cost calculations differ. Enterprises also need to factor in retries and human review: a low cost per attempt does not necessarily translate into a low cost per successfully completed task. Published results for 1017, GPT-6 and Opus 5.5, with sources