ModelBrief · Gemini 4 Neon · Topic Page
Last updated Nov 20, 2026

Gemini 4 Neon Benchmarks, Fully Mapped: AA 61 Across Four Tiers, the 18-Test Model Card Table, and the Independent Indexes

Published Nov 19, 2026 · Last updated Nov 19, 2026

One day after launch, Neon already has plenty of scores — but they split into two kinds: AA, Vals and Agent Arena are independent indexes, while DeepMind's model card table of 18 tests is Google's own runs. The sentence to remember: the headline 61 needs the most expensive xhigh tier; at the default high tier it scores 60, the same as Carbon; and at the middle and low tiers Neon actually trails Carbon. Here is the breakdown.

The AA Index: four tiers and their cost

The Artificial Analysis Intelligence Index v5.0.1 combines 9 evaluations (AA-Briefcase, AutomationBench-AA, Terminal-Bench 4.0, FrontierCode-AA 2.0, AA-Omniscience, GDP.pdf, HLE-Diamond, CritPt v2, Terminal-Bench-Science 0.1). Neon's four tiers, with their costs (AA model page, read Nov 19, 2026):

Thinking tierAA IndexTime per index taskTime to first answer tokenOutput tokens per index run
xhigh (top tier)61, #321.5 min56.51s283M
high (default)6014.8 min40.64s123M
medium558.3 min20.14s62.9M
low443.7 min13.93s28.3M

Three things stand out. First, xhigh beats high by 1 point while generating 2.3x the output tokens and taking 45% longer — the worst value per point in the whole table. Second, AA notes that at xhigh Neon is "among the slowest compared to other models in a similar price tier" (medians: 12.4 minutes per task, 13.08 seconds to first answer token), while Neon takes 21.5 minutes and 56.51 seconds — "fast task completion" is Google's claim relative to Carbon, not absolute speed. Third, the middle and low tiers flip the first impression: Carbon scores 58.3 at medium and 54 at low, both above Neon's 55 and 44. Neon's lead exists only at the top two tiers; users running medium or low get higher scores from Carbon.

Independent indexes: three No. 1 / No. 2 finishes

Beyond the composite index, Neon has placements on three independent boards. The AA Coding Agent Index at 58.6, No. 1 — Neon's hardest credential, and the one Google's blog pushes hardest; note that this index's composition was not separately verified against AA's page in this round, so cite it as Google's blog figure. The Vals Index at 70.9%, No. 2, on the Pareto frontier, with the Legal Research Bench and Tax Agent Bench both at No. 1 — enterprise knowledge work is Neon's strong suit. The Agent Arena Index at 1336, No. 2 (4,200+ real agent sessions), with steerability (adapting cleanly to mid-task user corrections) also No. 2. The LMArena Text snapshot (via kingy.ai, Nov 18): Neon High at 1518 ±10, third, on 3,885 votes with a rank range of 1–5; compare Carbon High at 1526 and Argon High at 1521 — few votes so far, and the score will move. Surge AI's first-place finishes across most finance, accounting, spreadsheet and constraint-following benchmarks are Google's blog summary of Surge's results; the original boards were not read directly in this round.

The model card's 18-test table: Google-reported, 15 wins and 3 losses

DeepMind's model card table pits Neon against Claude Opus 5.5 and GPT-6.2 Sol; Neon leads in 15 of 18 tests. The biggest leads: GraphWalks 256K–1M (95.2% vs Opus's 66.8%), SWE-sweep (provider harness 39.0% vs 10.1%), AutomationBench (56.3% vs 42.5%), CursorBench 4.1 (65.4% vs 58.1%). But Neon loses 3: Terminal-Bench 4.1 (65.2% vs Opus's 65.8%), FrontierCode 2.0 (46.2% vs 47.3%) and Chartography (73.5% vs Sol's 75.1%) — and two of those losses are in the coding-agent benchmarks that matter most (Terminal-Bench and FrontierCode), where Neon trails Opus 5.5 by a hair. That belongs next to the "Coding Agent Index No. 1" headline to keep the picture complete.

The biggest caveat is the two SWE-sweep harnesses: 39.0% on the provider harness (inside Google Antigravity) vs 12.8% on the third-party mini-SWE harness. Same model, same benchmark, a 3x gap depending on the harness. Agentic Commons flagged exactly this on launch day: "the harness matters enormously". Cite SWE-sweep only with the harness named, or it misleads. And remember: the model card table is Google's own runs, with rival figures taken from their reports rather than run side by side.

Safety and honesty numbers: all Google internal measurements

The model card also publishes safety and behavior figures, all from Google's internal evaluations (not independently verified): the "lowest fabrication rate yet" in internal Antigravity testing; 0.0% cheating on Terminal-Bench 4.0 (matching Carbon); a 2.45% shortcut rate on the coding cheating evaluation (lowest of any Gemini 4 model); a 0.09% misaligned action rate; 99.2% agent honesty; a 0.0% attack success rate on Gray Swan's indirect prompt injection benchmark at xhigh; and refusal rates of 1.98% of tasks on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science. All point the right way, but all come from Google; the full Neon Safety & Security Report is not yet published (the model card says "in the coming days"), and capability determinations currently carry forward from Argon.

Scores that do not exist yet

LiveBench 2.0 has no Neon entry; LMArena's vote counts are still thin; there is no independent hands-on test anywhere (no Quantslant- or Huxiu-style third-party review yet); and mainland Chinese first-tier tech media have not published depth coverage. In other words, Neon's score sheet currently consists of 3 independent indexes plus 1 Google-reported table, with the third-party testing column empty. This site's 8-hour tracking loop will watch for the first LiveBench entry, stabilized LMArena readings and the first independent tests.

Our take: Neon's scores are real, but the headline needs unpacking: 61 is the xhigh score, the default tier ties Carbon at 60, and the middle and low tiers trail Carbon; the Coding Agent Index No. 1 is a genuine credential, yet Neon loses Terminal-Bench 4.1 and FrontierCode 2.0 to Opus 5.5 by a hair; and the SWE-sweep 39.0% must be named as the Antigravity-harness figure — the third-party harness gives 12.8%. Check the tier first, then the harness, then the ranking — that is the right order for reading Neon's scores.

Reference sources: this article is an original rewrite produced by this site; facts and figures were checked against Artificial Analysis' model page (read Nov 19, 2026), DeepMind's model card (Nov 18, 2026), Google's official blog (Nov 18, 2026) and kingy.ai's LMArena snapshot (Nov 18, 2026)

← Back to the Gemini 4 Neon topic page