DeepSeek V4.1 Flash 1017 jumps 72.9→81.9 on BuildingBench at 1/15 of Sonnet's cost
ZhitingHu · x · 2026-10-18
EnactraAI benchmarked 3D building reconstruction from real photos across three models: DeepSeek V4.1 Flash 1017 (the new open model) scores 81.9 overall on BuildingBench, up 9 points from 72.9, improving on 11 of 12 buildings and entering the top five. It slightly beats Fable 5.1's 81.4, roughly matches Sonnet 5.5's 82.1, and closes in on GPT-6 Astra's 84.3. The cost gap is striking: median $3.01/task, about 1/11 of Fable 5.1 and 1/15 of Sonnet 5.5 (Astra: $7.05/task). Open models are getting seriously competitive at spatial reasoning and world-building.
Related event: DeepSeek V4.1 Flash Cracks BuildingBench Top 5 at 1/15 the Cost(2 posts)→
More from Models
- Mocked benchmark shows DeepSeek's document chart grounding badly misaligned — burkov · 2026-10-18
- Dev complains Claude's habit of saying "not done" is his biggest pain point — jasonkneen · 2026-10-18
- Open-weight TYGER speech decision model makes 80ms turn-taking calls for full-duplex voice agents — solyarisoftware · 2026-10-18
- Paying Google AI subscriber blocked: custom-domain Workspace accounts can't access AI Studio — ChrisRackauckas · 2026-10-18
- Running Ling 3.1 Flash (560B MoE, 25B Active) Across 4 DGX Sparks via llama.cpp RPC — BerryBlush_Rogers691 · 2026-10-18
- Users press OpenAI on long-delayed Adult Mode a year after it was promised — ___Patrice___ · 2026-10-18