Key takeaways
- Google's model card puts Gemini 4 Neon ahead of Claude Opus 5.5 and GPT-6.2 Sol on 16 of 19 benchmark rows, with Opus leading Terminal-Bench 4.1 and FrontierCode 2.0.
- Artificial Analysis scored Neon 61 on its Intelligence Index on November 18, 2026, one point above Gemini 4 Carbon and two above Opus 5.5.
- Neon costs $3 per million input tokens and $15 per million output tokens, below Carbon's $4 and $20 from December 1, 2026.
Gemini 4 Neon benchmarks, as Google published them on November 18, 2026, put the new mid-tier model ahead of Claude Opus 5.5 and GPT-6.2 Sol on 16 of the 19 rows in its model card, and level with or ahead of its bigger sibling Gemini 4 Carbon on most of the tests the two share. Neon costs $3 per million input tokens and $15 per million output tokens, with a 1M-token context window.
The first independent numbers agree on the broad picture. On the Artificial Analysis leaderboard, read on November 18, Neon at its highest thinking setting scores 61 on the Intelligence Index, one point above Carbon and two above Opus 5.5. The rest of the table below is Google's own testing, and most of the rival scores in it are numbers the rivals reported about themselves.
This comparison is based on the products' docs and published results. Nothing here was tested hands-on.
Gemini 4 Neon benchmarks against Opus 5.5 and GPT-6.2 Sol
Google's launch post and the model card carry the same table. It sets Neon against two rival flagships, not against Claude Sonnet 5.5 or Google's own models. According to Google's evaluation notes, Neon ran through the Gemini API at its highest thinking setting unless noted, and the Opus 5.5 and GPT-6.2 Sol scores are "sourced from providers' self-reported numbers unless otherwise mentioned".
| Benchmark (Google's table) | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|
| Artificial Analysis Index v5.0.1 | 61 | 59 | 57 |
| Vals Index v2.2 | 70.9% | 70.1% | 66.1% |
| CursorBench 4.1 | 65.4% | 58.1% | not listed |
| Terminal-Bench 4.1 (cheating-adjusted) | 65.2% | 65.8% | 61.5% |
| FrontierCode 2.0 | 46.2% | 47.3% | 43.8% |
| SWE-sweep, each lab's own harness | 39.0% | 10.1% | 8.3% |
| SWE-sweep, mini-SWE harness | 12.8% | 3.6% | 11.6% |
| AutomationBench | 56.3% | 42.5% | 41.3% |
| Agents' Last Exam v1.1 | 75.8% | 71.6% | 73.0% |
| RiemannBench (math) | 84.0% | 80.0% | 80.0% |
| Terminal-Bench-Science 0.1 | 70.5% | 63.3% | 68.0% |
| HLE-Diamond (reasoning) | 71.3% | 67.9% | 64.0% |
| MKA (multimodal) | 56.5% | 53.0% | 52.7% |
| Chartography (charts) | 73.5% | 66.3% | 75.1% |
| LVBench (video) | 89.0% | 83.7% | 87.6% |
| GraphWalks, 256K to 1M tokens | 95.2% | 66.8% | 80.1% |
| CWE-Bench v2 (secure code) | 70.7% | 67.1% | 66.0% |
| SEC-bench Pro | 90.5% | not listed | 88.3% |
| WebExploitBench v1.2 | 51.3% | 37.1% | not listed |
Bold marks the three rows where a rival leads. Opus 5.5 wins two of the coding-agent tests, Terminal-Bench 4.1 and FrontierCode 2.0, by about a point each. GPT-6.2 Sol wins Chartography by 1.6 points.
Two rows deserve a second look. The SWE-sweep gap is mostly a harness effect, meaning the software wrapped around the model. In each lab's own tool (Antigravity CLI for Neon, Claude Code for Opus, Codex for Sol) Neon scores 39.0% against 10.1% and 8.3%. In a neutral harness the gap with Sol shrinks to 12.8% against 11.6%. Google's notes say its CLI ships a sub-agent tool "particularly suited to bug finding tasks like SWE-sweep". The GraphWalks lead, 95.2% against 66.8% for Opus, is the clearest sign that Neon holds up when the prompt runs to hundreds of thousands of tokens, but Google computed it itself.
How Neon compares with its Gemini 4 siblings
Google's Neon pages don't put the siblings in the table. Their own model cards do, and several tests overlap. The Carbon and Argon numbers below come from the Gemini 4 Carbon model card, with results as of October 2026. The Hydrogen numbers come from the Gemini 4 Hydrogen model card, as of November 2026. Vending-Bench 2 for Neon is from Google's launch post.
| Benchmark | Neon | Carbon | Argon | Hydrogen |
|---|---|---|---|---|
| Vals Index v2.2 | 70.9% | 70.4% | 69.2% | not listed |
| Agents' Last Exam v1.1 | 75.8% | 77.4% | 69.0% | 71.6% |
| AutomationBench | 56.3% | 53.9% | 51.3% | not listed |
| Vending-Bench 2 (final balance) | $16,843 | $16,484 | $13,718 | $14,549 |
| RiemannBench | 84.0% | 80.0% | 76.0% | not listed |
| Chartography | 73.5% | 72.5% | 71.6% | 66.6% |
| GraphWalks, 256K to 1M tokens | 95.2% | 96.2% | 84.2% | not listed |
| SEC-bench Pro | 90.5% | 88.2% | 83.4% | not listed |
| HLE-Diamond | 71.3% | not listed | not listed | 58.8% |
| Terminal-Bench 4.1 | 65.2% | not listed | not listed | 57.0% |
| MKA | 56.5% | not listed | not listed | 53.4% |
On Google's numbers, Neon edges out Carbon on six of the eight shared rows and trails it on two, Agents' Last Exam and GraphWalks, by about a point each. Against Argon, Neon leads every shared row. Against Hydrogen, the fast and cheap model, it leads every row, by 12.5 points on HLE-Diamond and 8.2 on Terminal-Bench 4.1. Our Gemini 4 Carbon vs Argon comparison covers the older pair in more detail.
That Neon matches Carbon is the surprise. Google pitches Carbon as the model for the hardest, longest projects, and the Neon model card says Neon was adapted from Carbon. Google also says Neon at its high thinking setting costs 32% less per task than Carbon.
Compare rows only within one table
The rival scores move between Google's own tables. In the Carbon model card, dated October 2026, Opus 5.5 scores 69.6% on RiemannBench and 59.0% on Terminal-Bench-Science. In the Neon card a month later, the same model scores 80.0% and 63.3%. Benchmarks get new versions and leaderboards get new runs, so a score copied from one table into another can be out of date. The tables above keep each lab's numbers in the setting where Google published them.
Some of Neon's own rows are also measured differently from Carbon's. Neon's Terminal-Bench-Science score was computed by Google over five attempts per task, while Carbon's came from Artificial Analysis. That row is left out of the sibling table for that reason.
What independent tests show so far
Independent results lag a launch by days or weeks. Here is what the benchmark sites showed when checked on November 18, 2026:
- Artificial Analysis Intelligence Index. Neon at xhigh scores 61, third behind Claude Fable 5.2 at 65 and GPT-6.2 Astra at 62. Carbon at high scores 60, Opus 5.5 59, GPT-6.2 Sol 57, Hydrogen 54 and Argon 50. Its Neon page lists 110 output tokens per second, 56.5 seconds to the first answer token and $6.44 per Index task. Google's AA index row comes from this same source, so it is not a separate check.
- Vals AI. The Vals Index lists Neon second at 70.87%, matching Google's figure.
- Terminal-Bench. Neon is not on the official leaderboard yet. Opus 5.5 is there at 65.8% and GPT-6.2 Sol at 61.5%, the same numbers Google used.
- ARC Prize. Neon is not on the leaderboard yet.
Most of Google's other claims, including SWE-sweep, GraphWalks, LVBench and the cybersecurity rows, have no independent result yet.
Gemini 4 Neon pricing and context against the rest
Prices below are per million tokens, standard tier, as of November 18, 2026, from Google's Gemini API pricing page, Anthropic's pricing docs and OpenAI's pricing docs. Context figures are from OpenRouter's model list.
| Model | Input | Output | Cached input | Context |
|---|---|---|---|---|
| Gemini 4 Neon | $3.00 | $15.00 | $0.15 | 1,048,576 |
| Gemini 4 Carbon | $2.00, then $4.00 from Dec 1 | $10.00, then $20.00 | $0.10, then $0.20 | 1,048,576 |
| Gemini 4 Argon | $2.00, then $4.00 from Dec 1 | $10.00, then $20.00 | $0.10, then $0.20 | 1,048,576 |
| Gemini 4 Hydrogen | $1.50 | $7.50 | $0.075 | 1,048,576 |
| Claude Opus 5.5 | $4.00 | $20.00 | $0.20 | 1,000,000 |
| GPT-6.2 Sol | $2.00 (up to 272K), $4.00 above | $10.00, $15.00 above | $0.10, $0.20 above | 1,050,000 |
Google's pricing page lists Carbon and Argon at $2 and $10 through November 30, 2026, and at $4 and $20 from December 1, with no reason given. From that date Neon is the cheaper of the two strongest Gemini 4 models per token, not the dearer one. Our Gemini 4 Carbon price article has the earlier details.
Batch and Flex halve Neon's price to $1.50 and $7.50. Neon has no free tier in the API; Hydrogen does. In the Gemini app, Neon is for Google AI Pro and Ultra subscribers. GitHub made it generally available in Copilot Pro+, Max, Business and Enterprise on the same day. Our Gemini 4 Neon release article covers the launch itself.
Which model to pick
On Google's numbers and the first independent index, Neon is the strongest Gemini 4 model per dollar, and from December 1 it is cheaper per token than Carbon too. Pick Opus 5.5 if your work sits in a terminal or a coding agent, where it still leads on Terminal-Bench 4.1 and FrontierCode 2.0, at $4 and $20. Pick GPT-6.2 Sol if most of your prompts stay under 272K tokens and price matters most: at $2 and $10 it is the cheapest of the three flagships here, though it trails Neon on 16 of Google's 19 rows.
The numbers to watch are the independent ones still missing: Terminal-Bench, SWE-bench and ARC Prize runs of Neon, and Google's full safety and security report for Neon, which the model card says will follow within days.
Frequently asked questions
Is Gemini 4 Neon better than Gemini 4 Argon?
On Google's published numbers, yes. Neon scores higher than Argon on every benchmark the two model cards share, including Vals Index (70.9% against 69.2%) and SEC-bench Pro (90.5% against 83.4%). On Artificial Analysis, Neon scores 61 and Argon 50.
How does Gemini 4 Neon compare with Gemini 4 Hydrogen?
Neon leads Hydrogen on every shared benchmark in Google's model cards, by 12.5 points on HLE-Diamond and 8.2 points on Terminal-Bench 4.1. Hydrogen costs half as much, $1.50 and $7.50 per million tokens, and has a free API tier.
How much does Gemini 4 Neon cost?
As of November 18, 2026, the Gemini API charges $3 per million input tokens, $15 per million output tokens and $0.15 per million cached input tokens. Batch and Flex cost half that, and there is no free tier.
Where is the Gemini 4 Neon model card?
Google DeepMind publishes it as a PDF on its model cards site. Its results table compares Neon with Claude Opus 5.5 and GPT-6.2 Sol, and a separate evaluation document explains how each score was measured.
Sources
- Gemini 4 Neon model cardGoogle DeepMind
- Introducing Gemini 4 NeonGoogle
- Gemini 4 Neon model evaluation: approach, methodology and resultsGoogle DeepMind
- Gemini 4 Carbon model cardGoogle DeepMind
- Gemini 4 Hydrogen model cardGoogle DeepMind
- Gemini API pricingGoogle
- PricingAnthropic
- API pricingOpenAI
- Gemini 4 Neon in GitHub CopilotGitHub
- Google: Gemini 4 NeonOpenRouter
- LLM leaderboardArtificial Analysis
- Gemini 4 Neon: intelligence, performance and price analysisArtificial Analysis
- BenchmarksVals AI
- Terminal-Bench leaderboardTerminal-Bench



