I’ve seen opinions on Twitter that Kimi K2.5 used with OpenCode performs better than Claude Code paired with Opus 4.6.
I tried to cancel my Claude subscription (10) for a month to test that opinion out, and this experience proved that it is bullshit. I have not tried to use OpenCode with the Anthropic API key to pinpoint whether the problem is the harness or the model, but my “adventure” was so bad that I had to go back to Claude Code. For more context, the work involved writing TypeScript on the front-end and back-end, nothing too niche IMO.
I’ve read (I don’t have the source on me now, sorry…) that Claude tends to make certain assumptions about the request that were not specified, and this gap-filling makes the experience better. Maybe yes, maybe not, but at this point, the experience of coding with Claude Code remains undefeated.
On this topic, I also recommend this now: 5 Agent Skills I Use Every Day