Opus 4.7 Rebounds

We added coding evaluations, accumulating 145 results across the new and old version leaderboards.

The models tested in this pass include Kimi K2.6-Code-Preview, Opus 4.7, GLM 5.1, and Minimax M2.7, among others.

Key impressions are as follows.

• Opus 4.7 showed a substantial improvement, which is rare among recent upgrades.
• Kimi K2.6-Code-Preview doesn't yet feel significantly better, so judgment was reserved pending more use in other agentic environments.
• GLM 5.1 was quite good, and the author believes some open-weight models still don't reach Opus/GPT tier, contrary to exaggerated claims.
• In the top tier are Kimi K2.5 and GLM 5.1, which were assessed as possibly approaching Gemini/Sonnet level.
• In the mid tier, Minimax M2.7 and Qwen 3.6 Plus were named, and were considered still attractive in terms of price or local-run feasibility.

Another point of note is ForgeCode.

• It recorded the highest score with Minimax M2.7.
• However, its UX/DX differs greatly from OpenCode, making it hard to recommend.
• It comes in the form of a Zsh plugin, which may suit users who prefer that format, but at the time it was summarized as too buggy for immediate use, with breakage occurring across most other models/providers.