DeepSeek tops Claude Opus 5 on coding and agent benchmarks, but two numbers don't settle who leads

DeepSeek scored higher than Anthropic's Claude Opus 5 on two tests, according to the short video note we analysed. On a software engineering test it scored 77%, against 74% for Opus 5. On Automation Bench, which measures how well a model completes tasks on its own as an autonomous agent, the score was 60.5 against 50.3. The note says DeepSeek stood out on those agent tasks.

The second result matters more. A 3-point gap on a coding test can come down to how the test was run. A gap of roughly 10 points on an agent benchmark is harder to dismiss. Agents are where the market is moving: models that run multi-step tasks without supervision. The note does not say which DeepSeek model was tested or who ran the tests. It also does not say whether the results have been independently reproduced. Those details decide how much the numbers are worth.

We have been tracking this for months. In our own index, DeepSeek V4-Pro sits well behind Anthropic's and OpenAI's frontier models. The strongest Chinese labs, led by GLM and Qwen, are close but not level. Results like this one show the gap shrinking fastest in open-weight models, whose weights are published for anyone to run. That is good news for cost, control and access. It also follows the usual pattern. One lab chooses the tests that suit it, and the headline turns into "it beat Claude."

**Our reading:** this is a real data point, not a verdict. We would need three things before calling it a lead: the exact model and test version, independent replication, and results on harder tests that haven't been learned or solved yet. Even so, the direction is clear, and we welcome it in the long run. Competitive, cheap agents that anyone can run bring the abundance we expect closer, because automating work stops being something only a few providers control. In the short term, cheaper autonomous agents also mean office jobs face this sooner, and the tools to govern these agents are behind. Whoever wins the benchmark, that is the issue that needs attention.