Argon lands on VoxelBench: uneven builds, no overlap checks, and 2-minute attempts
legit_api · x · 2026-10-15
The new Argon model is out on VoxelBench, and third-party testing finds very uneven performance: sometimes clean, detailed builds, but often much smaller and less detailed output. The reviewer is underwhelmed overall and hopes the team closes the gap fast.
Notable details:
- Like Fable 5.0, Argon prefers smaller builds — but unlike Fable, it makes no effort to verify blocks aren't overlapping or clipping
- Argon averaged just 2 minutes per VoxelBench prompt, far below the time spent by Opus, Fable, Astra, or Sol
Speed but sloppiness is the current core weakness.
Related event: Gemini 4 Argon Debuts on VoxelBench to Underwhelming Results(2 posts)→
More from Models
- Google tops Signal65 PINNACLE agentic leaderboard: Gemini 4 Argon at $0.87 per correct task beats Sonnet and GPT — ryanshrout · 2026-10-15
- DeepSeek V4.1 Flash tops cost-efficiency at $0.134 per task, leaderboard shows — bindureddy · 2026-10-15
- Whatever Happened to Test-Time Scaling? Problems Needing pass@8192 Now Solve One-Shot — marc_lelarge · 2026-10-15
- cantrell on Gemini 4 Argon: heavy thinker, building long decode to push past 64K token limit — cantrell · 2026-10-15
- Opus 5.5 diagnoses fake OneDrive 'desktop takeover' in one prompt, maps 3 autostart paths — rehtorical · 2026-10-15
- GPT-5.5 discontinued: users mourn the loss of their favorite model for cyber work — Top-Support5831 · 2026-10-15