DeepSeek's price hike hits right as GLM-5.3 pulls ahead

DeepSeek just warned that V4 prices are going up a lot. That timing could not be worse: Zhipu AI's GLM-5.3 has pulled within a yuan of DeepSeek on peak-hour pricing — and it's winning most of the head-to-head benchmarks.

Assuming GLM-5.3 keeps GLM-5.2's API pricing, the peak-hour race is basically a tie. GLM-5.3 charges 8 yuan per million input tokens and 28 yuan for output; V4-Pro charges 9 and 27. GLM is a yuan cheaper on the way in, DeepSeek a yuan cheaper on the way out.

Once prices equalize, raw capability matters more. On Zhipu's apples-to-apples comparison table, there are nine directly comparable benchmarks, and GLM-5.3 takes seven. DeepSWE: 66.9 vs 62.7. Agents' Last Exam: 28.5 vs 25.7. HLE with Tools: 62.5 vs 60.0. CyberGym: 84.5 vs 83.3. DeepSeek leads only on NL2Repo and Toolathlon Verified. Terminal Bench 2.1 is essentially a tie: 88.2 to 87.9.

So after this price hike, DeepSeek's peak-hour edge for heavy coding agents gets awkward. V4-Pro barely costs less anymore, and GLM-5.3 wins most of the same-measure benchmarks. For continuous code editing, tool calls, and long-horizon tasks, GLM-5.3 has the better pitch.

DeepSeek's real bargain is off-peak and cache. V4-Pro drops to 4.5 yuan input and 13.5 yuan output outside peak, literally half price. Cache hits fall to 0.15 yuan (0.3 at peak), while GLM sits at 2 yuan. If you're running tough agents during the day, GLM-5.3 is the value pick. If you can shift to off-peak or lean heavily on cached context, DeepSeek is still a lot cheaper.