DeepSeek V4.1 Flash 1017 vs GPT-6 Astra: Same Score for 9 Cents a Test
DeepSeek updated V4.1 Flash today (snapshot 1017), so I put it through my VisualForge benchmark: the same interactive 3D web pages I give every model, each one opened, clicked and scored in a real browser. In this video: • what DeepSeek released: the 1017 update of V4.1 Flash, same API name (deepseek-flash), same price, already in the DeepSeek app and on chat.deepseek.com • what's inside: a 552B mixture of experts with about 8B active parameters reading and 16B writing, 1M context, native vision, open weights under MIT • DeepSeek's own benchmark table, row by row, and why its comparison columns are Opus 5 and GPT-5.6 Sol, not the current models • the price: $0.15 in / $0.60 out per million tokens off-peak, double in the weekday peak windows • how to use it: the DeepSeek app, the API, OpenRouter (plus the privacy setting that stopped my first run), and the open weights • my tests in 4K: Aero Lab, Jet Odyssey and Alex Optics, then Blast Island, Launch Day, Afterburner and Rotor Rampage, each against GPT-6 Astra, Opus 5.5 and Sonnet 5.5, plus GPT-6 Luna and Haiku 5.5 in the same price bracket • my results: 7.50 out of 10, second of 18 models, 0.01 behind GPT-6 Astra and ahead of Opus 5.5, Sonnet 5.5 and Gemini 4 Argon • the bill: 71 cents for all eight tests, about 9 cents a test, against $3.77 for GPT-6 Astra and $5.18 for Opus 5.5 on the same tests Settings: OpenRouter (deepseek/deepseek-v4.1-flash-1017, DeepSeek's own endpoint) at high effort, the same effort as Opus 5.5, GPT-6 Astra and Sonnet 5.5 on my board. One run, on release day, on a Saturday, so all of it was at the off-peak price, with about 99% of input tokens served from the cache. DeepSeek's cost is my OpenRouter bill; the other models' costs are estimates from each run's token counts at list prices. DeepSeek's numbers are labelled as DeepSeek's claims. Its Hangar Planner page saves to browser storage, which my public site's sandbox blocks, so it doesn't load on the site and isn't shown in the video. Every page DeepSeek built, with its scores, tokens and cost: https://visualforgebench.com/run/?id=20261017T091535_deepseek-deepseek-v4.1-flash-1017_8287 All results: https://visualforgebench.com Chapters 0:00 Intro 0:29 What DeepSeek released 1:25 What's inside 1:59 DeepSeek's benchmarks 3:18 The price 4:06 How to use it 4:54 My benchmark 5:28 Aero Lab 6:03 Jet Odyssey 6:38 Alex Optics 7:13 Quick round 8:12 Same price bracket 8:46 My leaderboard 9:54 The bill 10:43 Verdict Sources (read 17 October 2026) - DeepSeek API change log: https://api-docs.deepseek.com/updates - DeepSeek API pricing: https://api-docs.deepseek.com/quick_start/pricing - Model card, DeepSeek-V4.1-Flash-1017: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 - Model card, DeepSeek-V4.1-Flash: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash - OpenRouter: https://openrouter.ai/deepseek/deepseek-v4.1-flash-1017