Same weights, 3x score: Gemini 4 Neon jumps 12.8%→39% on SWE-sweep by swapping the harness
Delicious_Character6 · reddit · 2026-11-19
Google DeepMind released Gemini 4 Neon ($3/$15 per 1M tokens, 1M context, 127 tok/s on OpenRouter, #1 on the Coding Agent Index), its fourth release in two months. Two details deserve scrutiny: (1) promotional $2/$10 pricing on Carbon and Argon temporarily makes mid-tier Neon cost more per raw token than its larger siblings, even though it cuts cost per task by 32%; (2) on the minimal mini-SWE harness Neon scores 12.8% on SWE-sweep, but the same checkpoint hits 39.0% inside Google Antigravity—a 3x jump from harness alone. Practitioners report Neon still stumbles on messy multi-file refactors unless the loop tightly manages context compaction, tool routing, and file-read budgets. How much frontier-model hype is just harness engineering?
Related event: Gemini 4 Neon's SWE-sweep Score Challenged as Harness-Dependent(2 posts)→
More from coding & agent
- Garry Tan's "Personal AGI" is a folder of markdown files — 220,000 pages, 400x his 2013 output — jasonkneen · 2026-11-19
- ESP-Skainet: Open-Source Framework for Fully Offline Voice AI on ESP32 — alexcovo_eth · 2026-11-19
- Ploten ships an MCP server that live-edits open Power BI Desktop reports — plotenhq · 2026-11-19
- Testing gemini-4-neon with a DIY PowerShell agent harness: "benchmarks are for tourists" — dfinke · 2026-11-19
- Merge's Agent Handler adds credential passthrough: run tools without storing user tokens — shensi · 2026-11-19
- Gemini 4 Neon called best everyday coding model, beating Opus 5.5 — inductionheads · 2026-11-19