OpenCode V2 regresses badly for mid-size models, reports long-time user
GodComplecs · reddit · 2026-11-03
A Reddit user reports that upgrading OpenCode V1→V2 significantly degraded his workflow. On V1 he worked 16 hours straight with smooth changes (strata with official qwen flash settings); on V2, even minor changes fail to land in one pass, and face/website-design edits consistently fall flat, requiring heavy follow-up work.
Since his planning setup (Qwen 3.8 MAX) stayed the same, he blames the harness itself, concluding OpenCode now optimizes only for bigger paid models despite claiming to be "lighter and faster." He asks for alternative harnesses for projects where even SOTA models struggle with planning and reasoning, and says DSH disappointed even after a major update.
More from coding & agent
- Pen ships a canvas plugin for Codex, adding design powers to ChatGPT — paw_lean · 2026-11-03
- Elicit engineers all use Ramp's Inspect — a credit card company builds a top coding agent harness — jungofthewon · 2026-11-03
- VibeArmory launches macOS private beta: monitoring and firewalls for AI coding agents — agf2007 · 2026-11-03
- Agentic real2sim: Claude Fable + GPT-6 Astra rebuild scenes in RoboSimStudio in under 2 hours — berkeley_ai · 2026-11-03
- QuixiAI shows a delightfully terse FizzBuzz using tuple pattern matching — QuixiAI · 2026-11-03
- GitHub Copilot app adds Projects: give Copilot an ongoing goal across repos and sessions — 0xkarasy · 2026-11-03