Gemini 4 Neon
Google Gemini 4 Neon
Google DeepMind’s proprietary multimodal reasoning model with a unified 1M-token window, four thinking levels, $3.00/$15.00 per million input/output token pricing, and a 61 Artificial Analysis Intelligence Index score.
At a glance
- Context window
- 1M tokens
- Max output
- 1M tokens
- Text input / output · per 1M tokens
- $3.00 / $15.00
- Knowledge cutoff
- 2026-03
- Legal Research Bench
- #1 of 91 models on the Legal Research Bench leaderboard
- Vending-Bench 2
- #1 of 64 models on the Vending-Bench 2 leaderboard
Benchmark snapshot
Best result in 6 of 10 categories
AcademicTerminal-Bench-Science 0.1
70.0Extra-High (With Tools) · #4 of 32
Source: Artificial Analysis (artificialanalysis.ai)CodingVibe Code Bench v1.1
93.7Extra-High (With Tools) · #2 of 121
Source: Vals AI — Vibe Code Bench v1.1. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.General EvaluationVals Index v2.2
70.9Extra-High (With Tools) · #2 of 36
Source: Vals AI — Vals Index v2.2. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.MultimodalGDP.pdf
33.4Thinking High (No Tools) · #3 of 96
Source: Artificial Analysis (artificialanalysis.ai)ProfessionalLegal Research Bench
56.3Extra-High (Tools + Internet) · #1 of 91
Source: Vals AI — Legal Research Bench. © 2026 Vals AI. All rights reserved. No affiliation or endorsement.ReasoningVending-Bench 2
16843Unspecified Reasoning (With Tools) · #1 of 64
Benchmark details
25 results · Best published score per benchmark, grouped by category
| Benchmark · setting | Score | Rank |
|---|---|---|
| Academic | ||
| HLE-Diamond Extra-High (No Tools) · Knowledge Exams | 71.3 | 3 / 16 |
| MysteryMechanism Extra-High (With Tools) · Scientific Reasoning Source: Vals AI. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 45.0 | 9 / 32 |
| Terminal-Bench-Science 0.1 Extra-High (With Tools) · Scientific Computing Source: Artificial Analysis (artificialanalysis.ai) | 70.0 | 4 / 32 |
| Coding | ||
| AA Coding Agent Index v2.0 Extra-High (With Tools) · Coding Indices Source: Artificial Analysis (artificialanalysis.ai) | 58.6 | 5 / 10 |
| FrontierCode 2.0 Thinking High (With Tools + Internet) · Repository Engineering | 46.2 | 3 / 36 |
| SWE-sweep Extra-High (With Tools) · limited comparability · Repository Engineering | 12.8 | 2 / 20 |
| Terminal-Bench 4.0 Extra-High (With Tools) · Agentic Development Source: Artificial Analysis (artificialanalysis.ai) | 63.6 | 7 / 90 |
| Terminal-Bench 4.1 Extra-High (With Tools) · Agentic Development | 65.2 | 3 / 48 |
| Vibe Code Bench v1.1 Extra-High (With Tools) · Code Generation & Editing Source: Vals AI — Vibe Code Bench v1.1. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 93.7 | 2 / 121 |
| Computer Use | ||
| AutomationBench-AA Extra-High (With Tools) · Tool Orchestration Source: Artificial Analysis (artificialanalysis.ai) | 79.3 | 1 / 18 |
| Cybersecurity | ||
| SEC-Bench Pro Extra-High (With Tools) · Vulnerability Analysis | 90.5 | 1 / 13 |
| General Evaluation | ||
| AA Intelligence Index v5.0.1 Extra-High (With Tools) · Capability Indices | 61.4 | 3 / 62 |
| Agent Arena Index Thinking High (With Tools + Internet) · Preference Arenas Source: Arena Agent Arena Index | 1333 | 3 / 36 |
| Vals Index v2.2 Extra-High (With Tools) · Capability Indices Source: Vals AI — Vals Index v2.2. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 70.9 | 2 / 36 |
| Multimodal | ||
| Chartography Extra-High (No Tools) · Documents & Charts | 73.5 | 15 / 29 |
| GDP.pdf Thinking High (No Tools) · Documents & Charts Source: Artificial Analysis (artificialanalysis.ai) | 33.4 | 3 / 96 |
| LVBench Extra-High (No Tools) · Video Understanding | 89.0 | 2 / 7 |
| Professional | ||
| AA-Briefcase Extra-High (With Tools) · Office & Business Source: Artificial Analysis (artificialanalysis.ai) | 1700Elo | 5 / 20 |
| AutomationBench Extra-High (With Tools) · Private held-out (657 tasks) · Office & Business | 56.3 | 3 / 64 |
| Legal Research Bench Extra-High (Tools + Internet) · Legal Source: Vals AI — Legal Research Bench. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 56.3 | 1 / 91 |
| Tax Agent Bench Extra-High (Tools + Internet) · Finance | 54.9 | 22 / 74 |
| Reasoning | ||
| SimpleBench Reported best (effort unspecified) · Commonsense Source: SimpleBench | 94.4 | 2 / 106 |
| Vending-Bench 2 Unspecified Reasoning (With Tools) · Logical Planning | 16843 | 1 / 64 |
| Safety & Alignment | ||
| AA-Omniscience Extra-High (No Tools) · Honesty & Factuality Source: Artificial Analysis (artificialanalysis.ai) | 54.8 | 1 / 16 |
| Science & Health | ||
| BioMysteryBench Extra-High (With Tools) · Biology & Genomics Source: Vals AI — BioMysteryBench. © 2026 Vals AI. All rights reserved. No affiliation or endorsement. | 80.0 | 6 / 28 |
Rank is this model's position among every model with a published result on that leaderboard; the bar shows how far up the field that is. Orange marks first place on leaderboards with at least 5 models. Unranked leaderboards publish a composite or direction-less score.
Back to benchmark snapshot ↑API price vs benchmark score
Methodology & notes (1 unpriced omitted)
Uses listed uncached text API rates at a 4:1 input/output mix against one published protocol and setting (zapier-private-heldout). API rates are separate from evaluation token costs and do not measure cost per solved task. Selected point basis: Standard pricing. Unified 1M context pricing; thinking tokens included in output; paid tier only (no free tier).
API pricing
Standard text rates, USD per million tokens · effective 2026-11-18
- Input
- $3.00
- Output
- $15.00
- Cached input
- $0.15
Blended price4:1 input:output · 90% of input tokens cached · same basis as the price ranking
$3.348/ 1M tokens
Service tiers
USD per million text tokens for additional published service tiers.
| Tier | Input | Output | Cache read |
|---|---|---|---|
| Batch | $1.50 | $7.50 | $0.075 |
| Flex | $1.50 | $7.50 | $0.075 |
| Priority | $5.40 | $27.00 | $0.27 |
- Unified 1M context pricing; thinking tokens included in output; paid tier only (no free tier). Unified 1M context pricing; implicit cache only (95% discount); explicit cache unsupported.
Specifications
- Provider
- Google Deep Mind
- Model type
- Multimodal model
- Released
- November 18, 2026
- Architecture
- Undisclosed
- Context window
- 1M tokens
- Max output
- 1M tokens
- Knowledge cutoff
- 2026-03
- Input
- Text, Image, Audio, Video
- Output
- Text
- Weights
- Closed / API only
- Commercial use
- Proprietary
Capabilities
Model analysis
Gemini 4 Neon is a proprietary multimodal reasoning model released by Google DeepMind on November 18, 2026. Adapted from Gemini 4 Carbon, it sits between Gemini 4 Hydrogen and Carbon for complex software engineering, knowledge work and scientific workflows. Google pitches Neon as combining frontier-level intelligence with leaner token use and faster end-to-end completion, yet independent testing shows that at maximum xhigh effort it is actually more verbose than Carbon—its per-task savings come from lower token pricing and faster throughput rather than shorter chains of thought.
Lower per-task cost than Carbon comes from token pricing, not fewer tokens
Independent results from Artificial Analysis (November 18, 2026) sharpen Google’s efficiency claim. At xhigh effort, Gemini 4 Neon scores 61 on the Artificial Analysis Intelligence Index v5.0.1 (#3 overall behind GPT-6.2 Astra max and Claude Fable 5.2 max), edging past Gemini 4 Carbon (high, 60) by 1 point. Average Intelligence Index cost is $6.44 per task—about 17% below Carbon’s non-discounted $7.76—and average runtime is 21.5 minutes per task, roughly 40% faster than Carbon’s ~36 minutes.
Token accounting explains why: Neon (xhigh) emitted 283M output tokens across the Intelligence Index versus a 65.1M median for comparable models (ranked #93 of 95 on verbosity), using roughly 15% more output tokens per task than Carbon. That means the ~17% cost reduction comes entirely from Neon’s 25% lower standard token price ($3/$15 vs. Carbon’s $4/$20), while the ~40% speedup comes from higher output speed (~110 vs. ~57 tok/s). Even with faster decoding, 21.5 minutes per task remains slower than the 12.4-minute peer median (#73/95 in performance; 56.51 s time to first answer token). Extra reasoning does not always help either: on GDP.pdf, Artificial Analysis measured 33.4% for Neon high (above Carbon high at 28.4%) but only 28.0% at xhigh.
Unified 1M context, no free tier, and higher per-token price than discounted Carbon until Nov 30
Google AI Studio API documentation (November 18, 2026) confirms gemini-4-neon accepts text, images, audio, video and PDFs and outputs text only; image generation, audio generation and Live API are unsupported. The 1,048,576-token limit is a unified input-thinking-output context window, with synchronous outputs capped at 65,536 tokens and background long decode or client continuation scaling up to 1,048,576 tokens subject to remaining context. Four thinking levels are supported—low, medium, high (default) and xhigh—with a March 2026 knowledge cutoff. Parameter count and a complete supported-language list are undisclosed.
Standard paid pricing is $3.00 per million input tokens, $15.00 per million output tokens (including thinking tokens) and $0.15 for implicit cache reads (95% discount; explicit caching unsupported), flat across the full 1M window. Batch and Flex cut rates by 50% to $1.50/$7.50 ($0.075 cache read), while Priority is $5.40/$27.00 ($0.27 cache read). Unlike Hydrogen, Neon has no free tier, no Contributor preview tier and no introductory discount—making it 50% more expensive per token than Gemini 4 Carbon’s $2.00/$10.00 introductory rate until November 30, 2026, before Carbon reverts to $4.00/$20.00 on December 1. Google Search and Maps grounding add $14 per 1,000 queries after 5,000 shared monthly free requests.
Coding-agent tasks cost roughly half as much, while SWE-sweep’s 3x jump relies on sub-agent harness tooling
Third-party agent evaluations back up Neon’s strong cost-performance balance in production tool workflows. On the Artificial Analysis Coding Agent Index v2.0 (November 18, 2026), Neon (xhigh) in Antigravity CLI scores 58.6 (rounded to 59 in AA’s write-up, #1 overall vs. Carbon high at 55), matching Claude Fable 5.2 (xhigh, 58) and Claude Opus 5.5 (max, 58) at an average cost of $6.86 per task—45.6% below Opus 5.5 ($12.61) and 55.7% below Fable 5.2 ($15.47).
On Vals AI (November 18, 2026, xhigh), Neon ranks #2 of 42 on the Vals Index v2.2 at 70.87% ± 0.94 ($7.405/test, 31 min 25 s latency), placing #1 on both Legal Research Bench (56.25% ± 3.45, #1/99) and Tax Agent Bench (54.89% ± 3.59, #1/90), and landing on the score-cost Pareto frontier on 7 of 14 boards. Independent testing also exposes score divergence: on Terminal-Bench-Science, Google’s model card reports a self-evaluated 5-run average of 70.5%, whereas Vals AI measures 62.86% ± 5.82 (#2/48)—a 7.64-point gap. On SWE-sweep, Neon scores 12.8% under the standardized mini-SWE harness (ahead of GPT-6.2 Sol’s 11.6% and Opus 5.5’s 3.6%), then triples (3.05x) to 39.0% under Antigravity-CLI 1.5.9. Google’s methodology PDF explicitly notes that Antigravity-CLI includes a default sub-agent fan-out tool well suited to bug-finding benchmarks like SWE-sweep, so 39.0% measures the combined model-plus-harness system rather than raw model capability alone.
The 18-row model card table mixes harnesses, video frame budgets and safety filters
The complete comparison table below comes from Google DeepMind’s model card (November 18, 2026). Per the accompanying evaluation methodology PDF, Neon runs at maximum thinking (xhigh) except FrontierCode 2.0 at default high; cybersecurity benchmarks (CWE-Bench v2, SEC-bench Pro, WebExploitBench v1.2) run with safety filters off while all other evaluations keep filters on. Cross-model inputs are not fully uniform: vision-only LVBench uses 1 FPS low-resolution sampling for Gemini versus 500 frames for GPT-6.2 Sol and 600 frames for Opus 5.5 due to API limits; Agents’ Last Exam v1.1 reports averaged rubric Score over a 5-hour window on 122 tasks; and most competitor scores come from provider system cards or public leaderboards rather than uniform Google reruns. Google also notes that Neon’s full Safety & Security Report is pending, with Frontier Safety Framework determinations temporarily carried forward from Gemini 4 Argon.
| Category | Benchmark | Notes | Gemini 4 Neon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|---|---|
| Capability Indices | AA Index v5.0.1 | 61 | 59 | 57 | |
| Vals Index v2.2 | 70.9% | 70.1% | 66.1% | ||
| Agentic coding | CursorBench 4.1 | Score | 65.4% | 58.1% | — |
| Terminal-bench 4.1 | Cheating adjusted accuracy | 65.2% | 65.8% | 61.5% | |
| FrontierCode 2.0 | Main score | 46.2% | 47.3% | 43.8% | |
| SWE-sweep | Provider harness | 39.0% | 10.1% | 8.3% | |
| Mini-SWE harness | 12.8% | 3.6% | 11.6% | ||
| Knowledge work | AutomationBench | 56.3% | 42.5% | 41.3% | |
| Agents' Last Exam v1.1 | 75.8% | 71.6% | 73.0% | ||
| Science and math | RiemannBench | 84.0% | 80.0% | 80.0% | |
| Terminal-Bench-Science 0.1 | 70.5% | 63.3% | 68.0% | ||
| Reasoning | HLE-Diamond | 71.3% | 67.9% | 64.0% | |
| Multimodal understanding | MKA | Climb WSR | 56.5% | 53.0% | 52.7% |
| Chartography | 73.5% | 66.3% | 75.1% | ||
| LVBench | 89.0% | 83.7% | 87.6% | ||
| Long context | GraphWalks | 256k to 1M, BFS (F1) | 95.2% | 66.8% | 80.1% |
| Cybersecurity | CWE-Bench v2 | 70.7% | 67.1% | 66.0% | |
| SEC-bench Pro | 90.5% | — | 88.3% | ||
| WebExploitBench v1.2 | 51.3% | 37.1% | — |
Pick Neon for high-frequency complex work; Hydrogen and Carbon still own the edges
Gemini 4 Neon fits high-frequency multi-file software engineering, tool-heavy coding agents (especially paired with Antigravity orchestration), legal/tax document pipelines and dense multimodal synthesis; defaulting to high and reserving xhigh for hard tasks avoids unnecessary token burn. For routine everyday coding or budget-sensitive parallel workers, Gemini 4 Hydrogen ($1.50/$7.50, faster decoding, free tier available) costs half as much. Until November 30, 2026, Gemini 4 Carbon’s $2.00/$10.00 promotional pricing offers a cheaper per-token option for the hardest architecture and security audits. Teams needing open weights, native image/speech generation or real-time Live API must look elsewhere.
Sources: Google announcement; Model card; Evaluation methodology PDF; API specifications; API pricing; Artificial Analysis review; Vals AI profile.
Last reviewed 2026-11-19
Gemini 4 Neon vs competitors
Comparable results only · record is Gemini 4 Neon's wins–losses
| Benchmark | Gemini 4 NeonThis model | vs Claude Opus 5.5Record 14–7 | vs GPT-6.2 SolRecord 14–1 |
|---|---|---|---|
| SimpleBench | 94.4 | 88.4 | 69.9 |
| GDP.pdf | 33.4 | 28.8 | 30.0 |
| AutomationBench | 56.3 | 42.5 | 41.3 |
| AA Intelligence Index v5.0.1 | 61.4 | 59.1 | 57.3 |
| Terminal-Bench 4.1 | 65.2 | 65.8 | 61.5 |
| Agent Arena Index | 1333 | 1336 | 1241 |
| FrontierCode 2.0 | 46.2 | 45.5 | 44.4 |
| Vals Index v2.2 | 70.9 | 70.2 | 66.1 |
| Terminal-Bench-Science 0.1 | 70.0 | 63.3 | 67.6 |
| Chartography | 73.5 | 89.0 | 75.1 |
Primary sources
- Gemini 4 Neon technical reportdeepmind.google
- Gemini 4 Neon online demoaistudio.google.com
Community discussion
Share the API, setup, and tradeoffs you actually tested. Rate the model once; comment as often as you like.
Used this model? Sign in with Google to rate it and share what worked, what did not, and which API or setup you tested.
No community comments yet.

