DeepSeek V4.1 Flash 1017 ties GLM-5.3 Flash on ToneBench writing
On ToneBench writing, 1017 ties GLM-5.3 Flash at #23/24 but costs $0.014 vs $0.022 per script, uses 17K vs 37K thinking tokens, and takes 1.3 vs 5.2 minutes.
DeepSeek V4.1 Flash 1017 now ties GLM-5.3 Flash on ToneBench, our writing benchmark! Against the September V4.1 Flash, at the same list price: - On high: +97 Elo and 21% cheaper. - On high: matches the old version on max, for $0.0095 a script instead of $0.017. - On max: same quality within noise, 18% cheaper. It's cheaper because it uses fewer thinking tokens: 17K thinking tokens per script on max instead of 22K for scripts of the same length. Against GLM-5.3 Flash (max) by @Zai_org: a tie, #23 vs #24 of 134 configs, with overlapping Elo ranges. DeepSeek costs $0.014 a script instead of $0.022 and thinks less than half as long (17K vs 37K thinking tokens). In our runs, a script took 1.3 minutes instead of 5.2. So @deepseek_ai's is clearly improving the Pareto curve on writing. If you're on the September version, switch. If you took my GLM-5.3 Flash pick from the AI Engineer talk, I'd now take 1017 on max! Full board: http://benchmark.towardsai.com
This post quotes another post: x.com/deepseek_ai/status/2111375640717033777