V4.1 Flash 1017 cost comparison: output roughly 1/33 of Opus 5.5, DeepSWE 74.2 vs 74.0
The author crunches numbers: off-peak output at $0.60 per million tokens vs Opus 5.5's $20, near-tie on DeepSWE v1.1, a 4.5-point Automation-Bench lead, but nearly 20 points behind on HLE.
DeepSeek dropped V4.1-Flash-1017 this morning and the thing that stands out isn't the benchmark table, it's the price tag sitting under it. Off peak, output runs $0.60 per million tokens. Claude Opus 5.5 is $20. That's roughly 33x. Even DeepSeek's own peak-hour pricing, $1.20, still undercuts Opus by more than 16x. And on DeepSWE v1.1 the new Flash model is essentially tied with Opus 5, 74.2 to 74.0, with Automation-Bench actually favoring DeepSeek by four and a half points. Where it still loses is knowledge-heavy work. HLE has Opus ahead by close to twenty points, and Terminal-Bench 3.0 still favors the bigger frontier models. So this isn't "DeepSeek beats Opus," it's "DeepSeek beats Opus on the tasks that are mostly mechanical, and loses on the tasks that need breadth." Zoom out and the whole frontier is restructuring around that split. Google just pushed Gemini 4 Argon live through Antigravity, its multi-agent command center, pricing it at $2 in / $10 out during the introductory window, jumping to $4/$20 after. It leads on long-context tasks like GraphWalks at 84.2%, well ahead of GPT-6 Astra and Opus 5.5 in the 65-67% range. Meanwhile the open-weight tier, Kimi K3 at $0.90/$4.50, Qwen3.7 Max around $0.885/$2.655, GLM-5.2 at $0.18/$0.60, is pricing itself for high-volume agent loops rather than one-off chat. My read: we're watching three separate markets form under one "frontier model" label. One for reasoning depth, where Opus and GPT still lead. One for long context and multi-agent orchestration, where Argon is making a real claim. And one for cheap, fast, good-enough coding loops, where DeepSeek and the open-weight pack are racing each other to the floor on price. Picking a single default model for an agent stack in 2026 means picking which of those three jobs matters most to you, not which leaderboard looks best. https://x.com/deepseek_ai/status/2111375640717033777
This post quotes another post: x.com/deepseek_ai/status/2111375640717033777