Gemini 4 Neon benchmarks mostly hold up, but several "first place" claims are ties and one ranking has already slipped. On the public boards we checked on November 19, 2026, Neon's 61 on the Artificial Analysis Intelligence Index (third by model), 70.87% on the Vals Index (second), first place on Vals Legal Research and Tax Agent, 56.3% on Zapier's AutomationBench and 46.2% on FrontierCode 2.0 all match. Its Agent Arena claim does not: Google quoted 1336 and second place, while the board now shows Neon third at 1,333. Its Terminal-Bench 4.1 score has no row on the official board, and "best in class" on HLE-Diamond only holds if you leave out Claude Fable 5.2 and GPT-6.2 Astra.
Google released Neon on November 18 as the middle model of the Gemini 4 family, adapted from Gemini 4 Carbon (launch post). This guide does not review the model or its price. It takes each number in the launch post and the model card, finds who ran it, and checks it against the board that owns the test.
Gemini 4 Neon benchmarks at a glance
| Benchmark | Google's claim | Who ran Neon | Public board on Nov 19 |
|---|---|---|---|
| AA Intelligence Index v5.0.1 | 61, third overall | Artificial Analysis | Matches (61.4, third by model) |
| Vals Index v2.2 | 70.9%, second | Vals AI | Matches, 0.02 points ahead of third |
| Vals Legal Research Bench | First | Vals AI | Matches: 56.25%, first |
| Vals Tax Agent Bench | First | Vals AI | Matches: 54.89%, first |
| Agent Arena Index | 1336, second | Arena | Does not match: 1,333, third |
| AA Coding Agent Index v2.1 | 58.6, first | Artificial Analysis | Matches, 0.17 points ahead |
| Terminal-Bench 4.1 | 65.2% | Artificial Analysis | Matches AA data; no official row |
| FrontierCode 2.0 | 46.2% | Cognition | Matches, third |
| AutomationBench | 56.3%, first | Zapier | Matches |
| Vending-Bench 2 | $16,843, first | Andon Labs | Matches, within one standard error of second |
| HLE-Diamond | 71.3%, best in class | Artificial Analysis | Score matches; two models score higher |
| RiemannBench | 84.0%, best in class | Surge AI | Matches, 0.8 points ahead |
| SWE-sweep, Agents' Last Exam, GraphWalks, SEC-bench Pro | various | Self computed |
Where the launch numbers came from
Google's evaluation methodology page for Neon sets the rules. Neon scores are pass@1 at the highest thinking setting unless noted. Scores for the two comparison models, Claude Opus 5.5 and GPT-6.2 Sol, are taken from what each provider reported itself, unless a row says otherwise.
That comparison set is the first thing to notice. The model card table compares Neon with Opus 5.5 and GPT-6.2 Sol only. It leaves out Claude Fable 5.2 and GPT-6.2 Astra, the two models that beat Neon on several of the same boards. Every "best in class" line in the launch post depends on that choice of class.
The second thing is the source column. Many rows are borrowed from independent boards: Artificial Analysis for the Intelligence Index, HLE-Diamond, MKA and Terminal-Bench 4.1; Vals AI for the Vals Index; official public leaderboards for FrontierCode, AutomationBench and RiemannBench. Others are Google's own runs: SWE-sweep, Agents' Last Exam v1.1, Terminal-Bench-Science, GraphWalks, LVBench, SEC-bench Pro and WebExploitBench. A borrowed number was produced by someone with no stake in the launch. A self computed number used Google's harness and Google's choice of run.
Gemini 4 Neon on Artificial Analysis
The Neon model page lists 61 at xhigh, 60 at the default high setting, 55 at medium and 44 at low. In the leaderboard data the xhigh score is 61.4. Counting each model once at its best setting, only Claude Fable 5.2 (64.6 at max) and GPT-6.2 Astra (62.1 at max) score higher, so "third overall" is accurate. Opus 5.5 at 59 and GPT-6.2 Sol at 57 also match Google's table. For what the nine evaluations inside the index measure, see our Intelligence Index v5 explainer.
The same page adds context the launch post skipped. Neon at xhigh wrote 283 million output tokens to finish the index, against a median of 65.1 million for models in its price tier. It costs $6.44 and takes 21.5 minutes per index task. A high score that needs four times the median token count costs more in practice than its $3 and $15 per million token price suggests.
HLE-Diamond. Neon's 71.3% matches the Artificial Analysis run exactly, as do Opus 5.5 at 67.9% and Sol at 64.0%. On the same board, though, Fable 5.2 scores 80.2% at max and GPT-6.2 Astra 72.2%. "Best in class" is true only against the two models Google picked. Our HLE guide explains what Humanity's Last Exam and its Diamond subset test.
MKA. Neon's 56.5% on the MKA multimodal agent benchmark matches Artificial Analysis (56.48%). It is second there, 0.08 points behind Fable 5.2. Google's 52.7% for GPT-6.2 Sol is labeled as coming from the same board, but we found no Sol row in the MKA data.
The Vals Index, Legal Research and Tax Agent
The Vals Index, updated November 18, has Fable 5.2 first at 71.31%, Neon second at 70.87% and Claude Sonnet 5.5 third at 70.85%. Second place is real, but the gap to third is 0.02 points. Neon's cost per test ($7.40) is less than half of Fable 5.2's ($17.94) or Sonnet 5.5's ($16.18), which is the better argument for it.
The index is a GDP weighted average of four sectors: finance (weight 8.0), coding (5.6), legal (1.2) and tax (0.5). We recomputed it from the published component scores and matched the board within 0.03 points for all six models we tried. The breakdown shows where Neon's score comes from. It is strong on finance (72.55 sector average) and legal, but its Code Migration score (75.8%) trails Sonnet 5.5 (87.4%) and Fable 5.2 (85.8%), so its coding sector is the weakest of the top five.
The two "first place" claims check out on Vals' own boards, both updated November 18. On Legal Research Bench Neon leads with 56.25%, ahead of Gemini 4 Carbon at 55.77%. On Tax Agent Bench it leads with 54.89%, ahead of Gemini 4 Hydrogen at 52.30%. The legal lead is half a point, small enough to flip on a rerun; the tax lead is larger.
Agent Arena: the claim that changed
Google's launch post says Neon debuted second worldwide on Arena's new Agent Arena Index with a score of 1336, based on more than 4,200 real agent sessions. The Agent Arena board, dated November 18 with 68 models, shows something else:
| Rank | Model | Index | Median cost per task |
|---|---|---|---|
| 1 | Claude Fable 5.2 (Max) | 1,448 | $9.15 |
| 2 | Claude Opus 5.5 (Max) | 1,336 | $5.93 |
| 3 | Gemini 4 Neon (High) | 1,333 | $2.60 |
| 4 | Gemini 4 Carbon (High) | 1,303 | $1.35 |
| 5 | Claude Sonnet 5.5 (Max) | 1,296 | $3.36 |
Neon is now third, three points behind Opus 5.5. The most likely reason is ordinary drift: Arena's changelog says it rebuilt the board on November 18, anchoring index scores to a fixed set of models and adding six focused indexes, and a new model's score moves as sessions accumulate. A three point gap on a preference style index is a tie in practice; our Agent Arena Index explainer shows how the score is built and why second to sixth place overlap. The cost column is the clearer story: Neon sits on Arena's cost frontier at less than half Opus 5.5's median cost per task. Google's separate claim of second place on steerability did not appear in the default board view we could read, so we could not confirm it. Our leaderboard comparison covers how Arena scores differ from fixed test sets.
Coding claims: Terminal-Bench, FrontierCode and the Coding Agent Index
Artificial Analysis Coding Agent Index v2.1. Antigravity CLI with Neon at xhigh scores 58.64, first. The next four are Claude Code with Fable 5.2 at xhigh (58.47), Devin Fusion CLI with Opus 5.5 plus GPT-6 Luna (58.40), Claude Code with Fable 5.2 at high (58.24) and Claude Code with Opus 5.5 at max (58.16). Five agents within half a point is a tie. Neon's run cost $6.86 per task against $15.47 for the Fable 5.2 xhigh run. Our coding agent benchmark rankings track this index over time.
Terminal-Bench 4.1: 65.2%. This number comes from Artificial Analysis, not Google, and the raw data backs it: 129 of 198 attempts passed (66 tasks, three attempts each), which is 65.15%. Averaging that with Neon's FrontierCode-AA (43.03%) and SWE-Atlas-QnA (67.74%) scores gives the 58.64 index score exactly. The catch is that the official Terminal-Bench 4.1 board has no Neon row, while Google took Opus 5.5 (65.8%) and GPT-6.2 Sol (61.5%) from that official board. So the table mixes two sources. The official board's leader is GitHub Copilot with Copilot Fusion at 71.2%, and Gemini 4 Carbon in Antigravity CLI sits at 64.5%. Each row carries about plus or minus 3 points of confidence interval. Our Terminal-Bench explainer covers what 4.1 changed and why Terminal-Bench scores swing with the agent harness.
FrontierCode 2.0: 46.2%. The Cognition leaderboard, which added Neon on November 18, lists it third at 46.2% behind Fable 5.2 (50.2%) and Opus 5.5 (47.3%). Google's own table shows Opus ahead, which is to its credit. Artificial Analysis's separate FrontierCode run at xhigh gives 43.0%, a reminder that harness and settings move this benchmark; see our FrontierCode guide.
SWE-sweep: 39.0%. Self computed. Google's methodology page notes that its Antigravity CLI harness has a built in subagent tool "particularly suited to bug finding tasks like SWE-sweep." The same model scores 12.8% in a neutral mini-swe-agent harness. Most of the gap is the harness, which our harness leaderboard guide shows is common.
Agent and science claims
AutomationBench: 56.3%, first. Zapier's report confirms 370 of 657 tasks at xhigh on version 1.0.6, run November 17 to 18, ahead of Gemini 4 Carbon (53.9%) and GPT-6.2 Astra (52.4%). Scoring is strict: every assertion about the simulated company must pass. Zapier notes Neon was evaluated before release and its costs are estimates.
Vending-Bench 2: $16,843, first. Andon Labs' leaderboard lists $16,843.41 plus or minus $972, ahead of Gemini 4 Carbon at $16,484.21 plus or minus $1,096. The gap is smaller than either standard error, so this is a tie with Carbon.
RiemannBench: 84.0%. Surge AI's board has Neon first at 84%, with Fable 5.2 and GPT-5.6 Sol at 83.2%. Matches, but by under one point.
Terminal-Bench-Science: 70.5%. Google ran this itself, averaging five attempts per task. Artificial Analysis's independent run puts Neon at 70.0%, close to Google's figure, but GPT-6.2 Astra scores 78.6% on the same run. Another "best in class" that depends on the class.
Agents' Last Exam v1.1: 75.8%. Self computed on 122 tasks with a five hour window. Rival scores come from a system card and a public board, so the comparison is not like for like.
How to read the Neon launch table
Four checks cover most of what we found, and they work for any model launch:
- Find the methodology page and sort rows into borrowed and self computed.
- Check who is missing from the comparison columns. Here it was Fable 5.2 and GPT-6.2 Astra.
- Open each board and look at the gap to second place and the error bars. Vals, Vending-Bench and the Coding Agent Index were all ties at the top.
- Recheck a day later. Arena moved within a day of launch.
Our Gemini 4 Hydrogen benchmarks check found the same pattern a week earlier, plus a Legal Research claim that did not match. This time the Vals claims held.
FAQ
Is Gemini 4 Neon better than Claude Fable 5.2 on benchmarks?
No, not on most independent boards. Fable 5.2 leads Neon on the Artificial Analysis Intelligence Index (64.6 against 61.4), the Vals Index, HLE-Diamond, FrontierCode 2.0 and Agent Arena. Neon leads on Vals Legal Research and Tax Agent, AutomationBench, Vending-Bench 2, RiemannBench and, narrowly, the Coding Agent Index, usually at a much lower cost per task.
What is Gemini 4 Neon's Artificial Analysis score?
61 at xhigh (61.4 in the data), 60 at the default high setting, 55 at medium and 44 at low on Intelligence Index v5.0.1. That ranks it third by model behind Claude Fable 5.2 and GPT-6.2 Astra.
Is Gemini 4 Neon second on Agent Arena?
It was quoted at 1336 and second at launch. The Agent Arena board dated November 18 lists Neon (High) third at 1,333, behind Claude Fable 5.2 (Max) at 1,448 and Claude Opus 5.5 (Max) at 1,336. Three points is within normal movement for a new model.
Did Google run Neon's Terminal-Bench 4.1 score itself?
No. The 65.2% comes from Artificial Analysis running Neon in the Antigravity CLI harness, and its data shows 129 of 198 attempts passing. The official Terminal-Bench 4.1 board has no Neon row yet, and the rival scores in Google's table come from that official board.
Which Gemini 4 Neon benchmarks are self reported?
SWE-sweep, Agents' Last Exam v1.1, Terminal-Bench-Science, GraphWalks, LVBench, SEC-bench Pro, WebExploitBench and the Chartography scores for Neon and GPT-6.2 Sol. Treat those as Google's claims until an independent board runs them.
