DeepSeek-V4 Claims Spark SoTA Debate

The “SoTA” argument is only true if you narrow the frame to specific agentic and coding tasks; across broader reasoning benchmarks, the picture is much less clean.

• –Official tables show V4-Pro-Max near the top on agent work: SWE Verified is 80.6 vs Opus 4.6 Max at 80.8, and MCPAtlas Public is 73.6 vs 73.8.
• –On general knowledge and hard reasoning, it trails in several spots: MMLU-Pro, GPQA, and HLE do not show a blanket win over the best closed models.
• –DeepSeek’s Chinese release notes explicitly say V4 is close to Opus 4.6 in non-thinking mode, but still behind Opus 4.6 thinking mode.
• –The bigger story is economics and usability: open weights, 1M context, and lower compute/memory cost are the real differentiators.
• –So the right read is “best open model in some important regimes,” not “new universal SoTA.”