Open-source harness Intern matches Opus performance at ~half cost on Terminal-Bench
m4rkGS · reddit · 2026-11-20
A new open-source coding harness Intern (GitHub) runs every task across three model tiers: a frontier model (Opus/Hy4) plans the hard parts, an efficient model (Hy4 Flash/DeepSeek Flash/Sonnet, or even local) does the rest, and a small decision model handles hand-off timing, retrieval reranking, and context pruning.
Key benchmarks (Terminal-Bench 2.1, 87 tasks):
- Opus 5.5 → Hy4 Flash: 92.4% at $0.127/task, vs. Opus alone at 91–94% costing $0.28–0.39 in Pi/Hermes/OpenCode — same performance at roughly half the cost, with Hy4 Flash handling 2/3 of turns for 1% of the cost.
- Open weights (FrontierHarness v1.1, 62 tasks): DeepSeek V4.1 Flash → self-hosted Qwen3.8 27B scores 65.3% ($0.092) vs. 69.4% ($0.182) for DeepSeek alone.
- Single OpenAI vendor: gpt-6.1-sol → gpt-6-luna costs $0.143 vs. $0.233 solo (same 44/62 pass), and beats Codex CLI's $0.274.
Install via uv tool install "intern-agent @ git+https://github.com/nace-ai/intern" and run intern.
Related event: Open-Source Coding Harness Intern Matches Opus at a Fraction of the Cost(2 posts)→
More from coding & agent
- Gemini 4 Neon agent swarm builds two playable WebGPU pinball games from scratch — AI_Andrew · 2026-11-20
- Claude Fable 5.2 draws and folds a paper plane into its own painted world — flyable in browser — prasenx · 2026-11-20
- Beaker autonomously un-bloats agent wikis by tracing production errors back to bad knowledge — VictorSanh · 2026-11-20
- Vibe Coding Means Enterprises Build the Same Internal App Three Times Over — shensi · 2026-11-20
- Joseph Suarez open-sources his 'skill issues.md' — called a $100M file — yacineMTB · 2026-11-20
- Airbnb Cut Policy Classification Costs 99% With Gemma 4 E2B at 35ms Latency — petrusenko_max · 2026-11-20