Deepseek's V4-Pro hits 87.9 on Terminal Bench while cache prices rise sixfold

Deepseek shipped V4-Pro-0813, lifting Terminal Bench 2.1 from 72.1 to 87.9 and DeepSWE from 12.8 to 62.7, and released Deepseek Harness v0.1 under the MIT license — a plugin-based agent runtime pitched against Codex and Claude. The agent gains are real but narrow: Artificial Analysis moves V4-Pro from 45 to 53 on its Intelligence Index, still behind Claude Opus 5 at 63. The offset lands on August 16, when cache hits go from $0.003625 to $0.022 per million tokens off-peak — the steepest cost increase for agents that reread the same files, arriving as Deepseek raises capital ahead of an IPO.

Source: the-decoder.com ↗

Cache hits are seeing the steepest increase, going from $0.003625 to $0.022 off-peak and $0.044 at peak.

The Decoder

Why this matters

→ Cache prices spike sixfold, raising agent-building costs sharply.
→ V4-Pro gains narrow; still trails Claude Opus 5 in overall capability.
→ Open-source Harness lowers agent-building barrier but Deepseek's own benchmarks cherry-pick narrow wins.

Speed gains, price shock