OpenCode 2 jumps from 8th to tied-first across models, cost per solved task falls 87%
lvwerra · x · 2026-11-10
The Hugging Face Harness Leaderboard (comparing coding agent harnesses on the same models) shows OpenCode 2 far outscoring 1.19: from 8th of 9 harnesses to tied for first on all three models.
Key numbers:
- At high effort, task solve rate rose from 43.3% to 65.7% on GLM-5.3-Flash and from 57.0% to 69.4% on DeepSeek-V4.1-Flash
- On Flint, from 7.1% to 22.9%
- Cost per solved task fell from $3.35 to $0.45
Author joelniklaus explains part of the old gap was setup: Harbor's standard route for OpenCode 1.19 used OpenAI's Responses API, which he translated to Chat Completions for HF Inference Providers. It ran without errors but silently dropped the model's earlier reasoning — a bug spotted with help from the OpenCode team. OpenCode 2 now uses the Chat Completions API directly, making it more comparable to other harnesses.
More from coding & agent
- Muse Ember hits 46 on AA Coding Agent Index in Muse Code, beating Haiku 5.5 and GPT-6 Luna — ArtificialAnlys · 2026-11-10
- YC F26 startup Boxes launches cloud environments for running parallel Claude Code and Codex agents — ycombinator · 2026-11-10
- Nous Research launches Hermes Cloud: hosted personal agents running 24/7 in a browser tab — NousResearch · 2026-11-10
- Meta open-sources Astryx, an 8-year-in-the-making UI design system with 150+ components built for AI agents — udmrzn · 2026-11-10
- Local-first open-source GrokBot alternative runs qwen3.5:4b via Ollama — InternationalGap3698 · 2026-11-10
- SurrealDB Earth launches serverless database on Cloudflare: 3M free requests, 330+ regions — mattturck · 2026-11-10