V4.1 Flash 1017 跑分汇总：Terminal-Bench 超 GPT 5.6 Sol，成本约 1/17
汇总帖称 V4.1 Flash 1017 在 Terminal-Bench 以约 1/17 成本胜过 GPT 5.6 Sol，DeepSWE 1.1 得 77%、每任务约 0.2 美元，100 个漏洞查出 54 个共花 7.53 美元，比 Gemini 4 Argon 便宜 85% 且快一倍，MIT 开源权重。
The whale never disappoints 🐋 DeepSeek V4.1 Flash 1017 dropped a few hours ago, and the open-weight model is going toe to toe with US frontier models: > Beats GPT 5.6 Sol on Terminal-Bench at ~17x lower cost, jumping 11 points over the previous Flash in a single update > 77% on DeepSWE 1.1, ahead of GPT 5.6 Sol, Opus 5, GLM-5.3 and Kimi K3, for about 20 cents per completed coding task > Leads GPT 5.6 Sol and Opus 5 on Automation-Bench and Agents' Last Exam, the multi-step tool-use tests where agents usually fall apart > Found 54 of 100 known vulnerabilities in pull requests for $7.53 in total, cheap enough to run a security pass on every PR > Ran 85% cheaper and twice as fast as Gemini 4 Argon in a head-to-head test, while landing close on quality > Rebuilt Portal 2 in Godot, a playable racing game and a 3D Forbidden City, each from a short prompt > Open weights under MIT, released the same day, so anyone can download the exact checkpoint No new architecture, just RL training on coding agent harnesses with the model learning from its own feedback. Every benchmark DeepSeek published went up, and answers got shorter, so every task uses fewer tokens. Frontier-level coding at a fraction of the price of US models. Live on SPRO now: > $0.15 in / $0.60 out per 1M tokens > Cached input at $0.003 per 1M tokens for long agent runs > 1M context, up to 393K output tokens > Model ID: deepseek-v4.1-flash-1017 Same API, with guardrails, caching and spend limits on every call. https://spro.hridaai.com/explore/models/deepseek/deepseek-v4.1-flash-1017 #DeepSeek #LLM #AIGateway
这条是引用着另一条写的：x.com/deepseek_ai/status/2111375640717033777