Российский стартап Kodacode обогнал Claude Code в бенчмарке харнессов для DeepSeek

⚡️ Russian startup Kodacode overtook Claude Code in the Harness benchmark for DeepSeek The same DeepSeek V4 Flash was launched with the same level of reasoning in different agent shells. The guys from Kodacode compared five agent shells on one DeepSeek V4 Flash model. At the Max level, Kodacode harvested first place with a score of 52.2%. Claude Code scored 48.2%, OpenCode 47.6%. The test included 261 practical tasks, from fixing bugs to creating entire projects. It turned out that DeepSeek was not enough to just launch. It still needs to be cooked correctly. By the way, the guys from Kodacode are one of the few who have free plans, without VPN and with payment from Russia.