DeepSeek shipped V4-Flash today, and the headline isn’t the price — it’s that the cheap “Flash” tier reportedly beats their own V4-Pro-Preview on agent benchmarks. That inverts the usual tiering, where a flash/mini model trades capability for latency. Here the release notes claim the opposite on exactly the tasks that matter for tool use.

• 🎯 82.7 on Terminal Bench 2.1 and 54.4 on DeepSWE — these are agentic, multi-step tool-use evals, not single-shot Q&A
• ⚡ Native Responses API support and explicit Codex adaptation — it’s built to drop into an existing agent harness, not just answer a chat prompt
• 🔍 Numbers ran on DeepSeek’s own “Harness minimal mode” at max effort, temp 1.0 — harness and sampling settings move these figures a lot, so cross-model comparison stays apples-to-oranges
• ⚠️ Agent Last Exam 25.2 and Automation Bench 25.1 are low in absolute terms — the genuinely hard agentic tasks are still mostly unsolved
• 📊 DSBench-FullStack (68.7) and DSBench-Hard (59.6) are internal test sets — unreproducible until they’re released

The pattern worth watching is harness-native model releases. A model “specifically adapted for Codex” is a bet that the agent scaffold, not the raw weights, is where the next capability jump lives — which lines up with what I see running agentic workflows in production, where the harness often decides more than the checkpoint does. The HN discussion is already picking apart whether “minimal mode, max effort” is a fair way to report a benchmark.

If a flash-tier model can top a pro tier on agent evals, is the pro/flash split about capability anymore — or just about who pays for the extra tokens the harness burns?