Gemini 4 Neon xhigh costs 8.6x more but adds just 1-4 points on most benchmarks
jeremyphoward · x · 2026-11-23
A widely shared chart breaks down Gemini 4 Neon's benchmark scores across thinking levels (low/medium/high/xhigh) alongside cost: xhigh runs 8.6x the per-task cost of low (medium 2.5x, high 5x).
Key gains from stepping up to xhigh:
- Terminal-Bench Science +4.3, HLE-Diamond +2.9, Terminal-Bench 4 +2.0, AA Intelligence Index +1.8, FrontierCode 2.0 +1.1, CritPt +0.9
- AutomationBench barely moves (+0.1), and GDP.pdf actually drops 5.4 points
The takeaway: xhigh — the setting behind Google's headline scores — yields marginal improvements on most tasks and even regresses on one, raising questions about its cost-effectiveness. Data from Artificial Analysis and benchmark boards via Priors, Nov 18, 2026.
More from Models
- 55-Page Riemann Hypothesis Proof Guide, De-Formalized from Lean with Gemini and GPT — chaumian · 2026-11-23
- Sarvam open-sources its latest 105B model; founder says scaling agentic RL 'is not for the faint hearted' — itsOmSarraf_ · 2026-11-23
- Poland's 9B-parameter Thalamus Reasoning tops LongevityBench, beating 29 models — Dr_Singularity · 2026-11-23
- Dev: Gemini 4 Neon rewrote my training kernels and supervised runs; rumors point to bigger models — xennygrimmato_ · 2026-11-23
- Ant Group open-sources LingSante-3.1-air-VL, a 124B MoE medical multimodal model — jacek2023 · 2026-11-23
- Inkling 2 Mini: a 26B local model that runs on laptops, and someone rebuilt Apple Notes around it — simonguozirui · 2026-11-23