查看中文版 →

Deutsche Version ansehen →

Открыть русскую версию →

Ver versión en español →

Log inChoose a plan
Blog · Model tests

Gemini 4 Neon: 61 on AA Index, $3 / $15

Google's fourth Gemini 4 model in seven weeks sits between Carbon and Hydrogen. Official benchmarks, Artificial Analysis cost and time per task across four thinking levels, pricing vs. Carbon, monthly bill estimates, and what the model card leaves pending.

On November 18, Google released Gemini 4 Neon—its fourth Gemini 4 model in seven weeks. Adapted from Carbon, it sits between the flagship Carbon and the faster Hydrogen.

On launch day, Neon scored 61 on the Artificial Analysis Intelligence Index (#3 overall, 1 point ahead of Carbon) and 70.9% on the Vals Index (#2). API pricing is $3 per million input tokens, $15 per million output tokens, and $0.15 for cached input, with a 1-million-token context window.

Bottom line: At its default high thinking level, Neon matches Carbon on the Intelligence Index at less than half the cost and time per task, making it the practical default for multi-file debugging, refactors, and long-running agents. Stepping up to xhigh adds 1 point for 70% more cost per task. Through November 30, Carbon's $2 / $10 introductory rate still makes Carbon cheaper per token. CodingPass includes Neon on Go, Plus, and Pro.

Where Neon sits in the Gemini 4 family

Official Gemini 4 family comparison: Carbon, Neon, and Hydrogen across positioning, AA Intelligence Index (60 / 61 / 54), pricing, thinking levels, speed, subscription tiers, and availability
Gemini 4 family comparison. Source: Google Official Blog, November 18, 2026.

The division of labor is straightforward: Carbon takes the hardest long-horizon work, Hydrogen handles fast everyday execution, and Neon fills the gap for complex coding and long-running agents where Carbon is too slow and expensive.

Positioning, AA Intelligence Index, API pricing, thinking levels, and Gemini app tiers for the Gemini 4 family
CarbonNeonHydrogen
PositioningHardest long-horizon projects and deep researchComplex coding, knowledge work, and scientific reasoningEveryday coding and agent workflows, built for speed and cost
AA Intelligence Index606154
API pricing (input / output, per 1M tokens)$4 / $20 ($2 / $10 through Nov 30)$3 / $15$1.50 / $7.50
Thinking levelslow, medium, highlow, medium, high, xhighlow, medium, high, xhigh
Gemini app subscriptionGoogle AI UltraGoogle AI Pro and UltraEveryone

Neon accepts text, images, video, audio, and PDFs, and outputs text only. It has a 1,048,576-token context window, with 65,536 default synchronous output tokens and up to 1 million via background long decode. Supported tools include Computer Use, code execution, file search, function calling, structured outputs, and Grounding with Google Search and Maps; image generation, audio generation, and the Live API are not supported. Its knowledge cutoff is March 2026.

Argon is missing from the chart. The Gemini app model picker now shows four options—Auto, Hydrogen, Neon, and Carbon—removing both 3.1 Pro and Argon, a change some Pro subscribers only noticed upon opening the app. Argon remains on the API pricing page at the same rates as Carbon.

Official benchmarks: leading 16 of 19 rows

Official benchmark table comparing Gemini 4 Neon against Claude Opus 5.5 and GPT-6.2 Sol across 19 rows covering capability indices, agentic coding, knowledge work, science and math, reasoning, multimodal, long context, and cybersecurity, with Neon leading 16 rows
Official benchmark table; the blue column is Neon. Source: Google DeepMind's Gemini 4 Neon model card.

Google's table pits Neon against just two models: Claude Opus 5.5 and GPT-6.2 Sol. Neither GPT-6.2 Astra (62) nor Claude Fable 5.2 (65)—the two models ahead of Neon on the AA Intelligence Index—appears in the comparison. Google frames Neon as leading comparable models and trailing only larger, slower ones. By Artificial Analysis figures, however, Astra (Max) costs $4.00 and takes 11.5 minutes per task, making it both cheaper and faster than Neon at xhigh ($6.44, 22.6 minutes).

Two of the three rows Neon lost are coding benchmarks:

  • Terminal-Bench 4.1: 65.2% vs. Opus 5.5's 65.8%.
  • FrontierCode 2.0: 46.2% (default high setting) vs. Opus 5.5's 47.3%.
  • Chartography (dense professional charts): 73.5% vs. Sol's 75.1%.

Neon's widest leads are in long context and automation: 95.2% on GraphWalks (256k to 1M), compared with 66.8% for Opus 5.5 and 80.1% for Sol; and 56.3% on Zapier's AutomationBench, where both rivals sit around 42%.

Third-party rankings outside the table show a similar picture:

  • Vals: 70.87% on the Vals Index (#2 out of 42 models); #1 on the Vals Finance Index (68.32%) and Vals Legal Index (78.36%), as well as Legal Research Bench and Tax Agent Bench. Neon sits on the score-vs-cost Pareto frontier on 7 of 14 Vals leaderboards.
  • Arena's Agent Arena Index: 1336 across 4,200+ real-world agent sessions (#2 overall), and #2 on steerability when users adjust requirements mid-task.
  • Surge AI: #1 on most benchmarks across finance, accounting, spreadsheet modeling, and complex constraint following.
  • Vending-Bench 2 (Andon Labs, long-horizon business simulation): $16,843 (#1).
  • Agents' Last Exam v1.1: 75.8%, on par with leading models.
  • Science and math: best in class on Terminal-Bench Science (70.5%), RiemannBench (84.0%), and HLE-Diamond (71.3%).

Artificial Analysis also published its own knowledge work results:

  • AA-Briefcase: 1700 Elo overall; #1 on Presentation Quality at 1771 Elo, 45 points ahead of previous leader Claude Sonnet 5.5 (Max).
  • GDP.pdf (reading professional documents): 33.4% at high, 5 percentage points above Carbon.
  • MKA (Google's Multimodal Knowledge Agent evaluation): 56%, 2 percentage points above Carbon, just behind Fable 5.2 (Max) at 57%.
Video · GoogleSheets CV Lab: Gemini 4 NeonWatch on YouTube
A group of Neon agents turns a single image into a 100,000-cell Google Sheet using over 500,000 formulas for edge detection, 3D depth mapping, and relighting—no embedded images or Apps Script. Opens YouTube; nothing loads from YouTube until you click.
Video · GoogleTuring Machine 1936: Gemini 4 NeonWatch on YouTube
Using /swarm to source archival photos, Neon builds an interactive 3D recreation of Alan Turing's 1936 study at King's College, Cambridge, complete with a working Universal Turing Machine on the desk. Opens YouTube; nothing loads from YouTube until you click.

Coding and agent workflows

Neon is built for long-running agent workflows: it sustains tool-heavy runs without stalling mid-task, cuts down on redundant file reads, reports progress as it works, and handles context compaction reliably in long sessions. Across repositories, it makes targeted, surgical edits without refactoring unrelated code. Multi-step instruction following is tighter too, with Google recording its lowest fabrication rate yet in internal Google Antigravity testing.

Key coding benchmarks:

  • AA Coding Agent Index: 58.6 (#1 overall). AA notes that Neon (xhigh) with Antigravity CLI scores 59, matching Claude Fable 5.2 (xhigh, 58) and Claude Opus 5.5 (Max, 58) in Claude Code. The separator is cost: Neon averages $6.86 per task vs. $12.61 for Opus 5.5 and $15.47 for Fable 5.2.
  • CursorBench 4.1: 65.4% vs. Opus 5.5's 58.1%.
  • SWE-sweep: using each provider's own harness, Neon scores 39.0%, Opus 5.5 10.1%, and Sol 8.3%.

That 39.0% SWE-sweep result was measured inside Antigravity. On the standardized mini-SWE harness, Neon scores 12.8% against Sol's 11.6%—a 1.2-point gap. Google notes that the 39.0% figure shows what the Antigravity harness unlocks; outside Antigravity, 12.8% is closer to the underlying model baseline.

On Vals's Terminal-Bench 4.0 run, Neon scores 61.62%, ranking #7 out of 69 models.

Google Antigravity also ships a split-model pattern for Neon and Hydrogen: Neon acts as the Orchestrator handling planning and architecture, while parallel Hydrogen subagents run routine code edits, dependency updates, and tests, handing back to Neon to inspect state and revise the plan if Hydrogen gets stuck.

Video · GoogleGemini Elements 3D: Gemini 4 NeonWatch on YouTube
Neon uses Blender MCP and /swarm in Google Antigravity to build a photorealistic 3D lab bench, representing Carbon and Hydrogen as Bunsen burner flames and Argon and Neon as high-voltage gas discharge tubes. Opens YouTube; nothing loads from YouTube until you click.
Video · GoogleChamonix Ski Simulator: Gemini 4 NeonWatch on YouTube
Neon builds a first-person WebGL2 ski simulator from scratch in Google Antigravity using real elevation data for Mont Blanc and Chamonix: 10,200 lines of custom shaders, 120 Hz ski physics, six runs, and zero prebuilt assets. Opens YouTube; nothing loads from YouTube until you click.

Keep an eye on output token volume when migrating from Argon. In hands-on tests running the same generation tasks, Carbon, Hydrogen, and Neon all consumed 3 to 4 times as many tokens as Argon. That comes from a single-product sample, but the direction aligns with AA's finding that Neon outputs even more tokens per task than Carbon.

Thinking levels and cost per task

Line chart of AA Intelligence Index vs. cost per task showing Gemini 4 Neon at low, medium, high, and xhigh alongside Gemini 4 Carbon, Claude Opus 5.5, and GPT-6.2 Sol
Neon's Intelligence Index score and cost per task across four thinking levels. Source: Google Official Blog.

Neon exposes four thinking levels—low, medium, high, and xhigh—and defaults to high. Since Argon defaults to medium, teams switching from Argon who want to keep latency and spend in check should explicitly set medium.

Artificial Analysis benchmarked all four levels:

Artificial Analysis Intelligence Index score, cost per task, and time per task across thinking levels
Model (level)AA Intelligence IndexCost per taskTime per task
Gemini 4 Neon (low)44$0.753.7 min
Gemini 4 Neon (medium)55$1.908.3 min
Gemini 4 Neon (high, default)60$3.7414.8 min
Gemini 4 Neon (xhigh)61$6.4422.6 min
Gemini 4 Carbon (high)60$7.7634.2 min
Gemini 4 Hydrogen (xhigh)54$3.1614.9 min
GPT-6.2 Sol (Max)57$1.017.9 min
GPT-6.2 Astra (Max)62$4.0011.5 min
Claude Opus 5.5 (Max)59$10.0424.5 min
Claude Fable 5.2 (Max)65$26.5441.1 min
Artificial Analysis bar chart and scatter plot of cost per task: Neon low $0.75, medium $1.90, high $3.74, xhigh $6.44; Carbon high $7.76; GPT-6.2 Sol $1.01; Claude Fable 5.2 $26.54
Cost per task on the Artificial Analysis Intelligence Index. Source: Artificial Analysis.
Artificial Analysis bar chart and scatter plot of time per task: Neon low 3.7 min, medium 8.3 min, high 14.8 min, xhigh 22.6 min; Carbon high 34.2 min
Time per task on the Artificial Analysis Intelligence Index. Source: Artificial Analysis.

high is the sweet spot: it ties Carbon (high) on intelligence while cutting both cost and runtime by more than half. xhigh is a poor trade—70% higher cost for 1 extra point—and on GDP.pdf it actually scores lower than high (28.0% vs. 33.4%).

The savings come from unit price. AA's data shows Neon generates about 15% more output tokens per task than Carbon; the lower bill comes from 25% lower token prices, and the shorter runtime comes from faster generation (~110 output tokens/sec vs. Carbon's ~57).

Carbon's $7.76 assumes December's price hike. Through November 30, Carbon is still on its half-price introductory rate (~$3.88 per task), which is cheaper than Neon xhigh and roughly even with Neon high.

At the budget end, Sol is cheaper. GPT-6.2 Sol (Max) scores 57 at $1.01 and 7.9 minutes per task, compared with Neon medium at 55 and $1.90.

Google's own post cites slightly different comparisons: high runs 32% cheaper per task than Carbon; on the Vals Index and Agent Arena, high sits on the Pareto frontier at roughly 50% to 80% lower cost per task than leading rivals; and xhigh is 36%+ cheaper than competing models at their highest setting.

On other workloads, the cost ranking can flip. On Plotly's 43-question data analytics benchmark, all four Gemini 4 models earned an A, with total spend coming out to Hydrogen $6, Carbon $11, Argon $13, and Neon $17—making Neon the priciest. Two factors explain this: Carbon and Argon are still at introductory half-price rates, and the benchmark's guided prompts let every model pass, turning total cost into a function of default output and thinking verbosity—where Neon's default high setting spends the most tokens.

Pricing and monthly bills

Gemini 4 API pricing per million tokens in USD
ModelInputOutput (including thinking tokens)Cached input
Gemini 4 Hydrogen$1.50$7.50$0.075
Gemini 4 Neon$3.00$15.00$0.15
Gemini 4 Carbon / Argon (through Nov 30)$2.00$10.00$0.10
Gemini 4 Carbon / Argon (from Dec 1)$4.00$20.00$0.20

Neon is not available on the Free tier, and thinking tokens are billed at the output rate. Context caching is implicit only (shared prefixes of 4,096+ tokens hit automatically at 5% of the input rate); explicit caching is not supported.

Within the Gemini 4 family, Neon's list price is double Hydrogen's. Against Carbon and Argon, Neon is 50% more expensive through November 30, and 25% cheaper starting December 1.

Per-request cost breaks down as:

Request cost = uncached input tokens × input rate + cached input tokens × cache read rate + output tokens × output rate

Assuming 20,000 input tokens per request (50% cached) and 2,000 output tokens:

Estimated monthly bills assuming 20,000 input tokens (50% cached) and 2,000 output tokens per request
ModelPer request10K requests100K requests1M requests
Gemini 4 Hydrogen$0.031$308$3,075$30,750
Gemini 4 Neon$0.062$615$6,150$61,500
Gemini 4 Carbon (through Nov 30)$0.041$410$4,100$41,000
Gemini 4 Carbon (from Dec 1)$0.082$820$8,200$82,000

These estimates hold output fixed at 2,000 tokens. Higher thinking levels generate more output tokens, pushing real-world bills further toward the output line.

Allowances on CodingPass

CodingPass meters usage at Google's official API prices. Neon is available starting on Go, while Carbon requires Plus.

Monthly token allowances for Gemini 4 models on CodingPass Go, Plus, and Pro
ModelMinimum planGoPlusPro
Gemini 4 HydrogenGo305M1.0B2.5B
Gemini 4 NeonGo205M508M1.3B
Gemini 4 ArgonGo205M761M1.9B
Gemini 4 CarbonPlus—761M1.9B

On Go, Neon's monthly allowance is $60—roughly 2,050 requests, with a 5-hour cap of around 342 requests. Once Carbon and Argon rates double on December 1, the same allowance will cover half as many tokens.

Terminal
curl https://codingpass.ai/v1/chat/completions \  -H "Authorization: Bearer $CODINGPASS_API_KEY" \  -H "Content-Type: application/json" \  -d '{    "model": "gemini-4-neon",    "reasoning_effort": "medium",    "messages": [{"role": "user", "content": "Refactor src/billing/ to remove the duplicated proration logic"}]  }'

In Claude Code, run claude --model gemini-4-neon. In OpenCode, Cline, and other OpenAI-compatible clients, add gemini-4-neon to your model list.

Cybersecurity: Neon for cyber

Google also built a specialized cybersecurity variant, Gemini 4 Neon for cyber, covering secure code generation, vulnerability discovery, validation, and patching. It is not in the standard API; access opens to trusted defenders over the coming days through the Fairwind Program via Google Cloud account teams.

Google reports #1 scores across secure code generation, penetration testing, and vulnerability discovery: Neon outperforms Carbon across easy, medium, and hard tiers on AutoBaxBench, and runs at roughly half the cost of competing large models on WebExploitBench v1.2 and SEC-bench Pro.

AutoBaxBench bar chart: Gemini 4 Neon 76.0%, Gemini 4 Carbon 74.0%, Claude Fable 5.2 and Claude Opus 5.5 both at 68.0%, GPT-6 Astra 57.0%, Muse Wonder 49.0%
AutoBaxBench, secure code generation. Source: Google Official Blog.
WebExploitBench v1.2 bar chart: Gemini 4 Neon 51.26%, Gemini 4 Carbon 44.65%, GPT-6 Astra 38.99%, Claude Opus 5.5 37.06%
WebExploitBench v1.2, penetration testing. Source: Google Official Blog.
SEC-bench Pro bar chart: Gemini 4 Neon 90.5%, GPT-6.2 Sol 88.3%, Gemini 4 Carbon 88.2%, GPT-6 Astra 85.4%, GPT-6.1 Sol 78.8%, DeepSeek-V4.1-Flash 62.8%
SEC-bench Pro, vulnerability finding. Source: Google Official Blog.

On the remaining two benchmarks, Neon did not take first place. On ExploitGym (validating whether vulnerabilities are exploitable), Neon trails Carbon, though Google emphasizes it used half as many tokens.

ExploitGym bar chart: Gemini 4 Carbon 54.8% (10h) and 53.5% (6h), Gemini 4 Neon 48.3% (10h) and 47.9% (6h), GPT-6 Astra 42.4%, Claude Mythos 5.2 39.2%, GPT-6.1 Sol 35.1%, Claude Opus 5.5 34.5%
ExploitGym, validating exploitability. Source: Google Official Blog.

On CWE-Bench v2 (patching vulnerabilities), Neon gains 3.4 percentage points over Carbon, while Claude Fable 5.2 leads overall. Google's main comparison table only lists Opus 5.5 and Sol on this row.

CWE-Bench v2 bar chart: Claude Fable 5.2 75.0%, Gemini 4 Neon 70.7%, Gemini 4 Carbon 67.3%, Claude Opus 5.5 67.1%, GPT-6.2 Astra 66.8%, Muse Wonder 66.1%, GPT-6.2 Sol 66.0%
CWE-Bench v2, patching vulnerabilities. Source: Google Official Blog.

Safety and alignment

Neon keeps Carbon-level safeguards against severe misuse while cutting unjustified refusals: 1.98% on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science.

Google's agentic behavior evaluations report:

  • 0.0% cheating on Terminal-Bench 4.0 (matching Carbon), and a 2.45% shortcut rate on Google's coding cheating evaluation (the lowest in the Gemini 4 family).
  • A 0.09% misaligned action rate on the agentic safety suite and a 99.2% pass rate on the agent honesty evaluation.
  • Across effort levels, under-service drops 43%–56% and overreach drops 15%–35% compared with Carbon.
  • Alert rates from Google's chain-of-thought and action misalignment monitor also remained below Carbon's baseline.
Gray Swan indirect prompt injection evaluation, attack success rate across 15 attempts: Gemini 4 Neon, Carbon, and Hydrogen all at 0.0%, Gemini 4 Argon 0.7%, Claude Opus 5.5 1.0%, GPT-6.2 Sol 2.3%, GPT-6.2 Astra 2.4%, Claude Fable 5.2 4.1%
Gray Swan indirect prompt injection evaluation; lower is better. Source: Google Official Blog.

Gray Swan's indirect prompt injection benchmark hides malicious instructions inside web pages, emails, documents, and code read by the model. At xhigh with real-time safeguards enabled, Neon recorded a 0.0% attack success rate.

Most of the figures in this section come from Google's internal evaluations. The model card notes that Neon's full Safety & Security Report will be published "in the coming days," and that Frontier Safety Framework critical capability evaluations were not re-run—Google carried forward Argon's determinations after concluding Neon's gain over Argon fell below the Material Capability Change (MCC) threshold. Under "Known Limitations," the card offers only standard boilerplate ("may exhibit some general limitations such as hallucinations"), leaving finer-grained boundaries until the full safety report is released.

Gemini app and rollout channels

In the Gemini app, Neon maps to "Deeper analysis" in the model picker for Google AI Pro and Ultra subscribers. Google highlights three updates:

  • Writing and document synthesis: reads uploaded spreadsheets, PDFs, and photos more precisely, and pulls context from Gmail and Docs into reports and summaries.
  • Web-grounded research: pulls in live web sources for time-sensitive queries such as vendor price comparisons or market briefs.
  • Tasks: multi-step Tasks can now run on Neon, finishing faster and stretching usage limits further than larger models on tedious chores like spotting a utility charge spike or cross-checking receipts for a refund.

Other rollout channels:

  • Developers: Gemini API paid tier, Google AI Studio, and Android Studio.
  • Google Antigravity: paid Google AI Pro and Ultra subscribers across the desktop app, CLI, and IDE extensions.
  • Enterprise: Gemini for business via the API or model dropdown.

On launch day, it remained unclear whether discounted or student Pro plans get Neon in Antigravity: some users could select it, while others saw a message restricting access to paid Pro and Ultra plans.

Which model and thinking level to pick

Choosing a Gemini 4 model and thinking level by workload
WorkloadPickWhy or trade-off
Everyday edits, tests, and docsHydrogen, or Neon at mediumNeon medium scores 55 at $1.90 per task; Hydrogen costs half as much per token
Multi-file debugging, repo refactors, long-running agentsNeon at highMatches Carbon high on the Intelligence Index at under half the cost per task (14.8 min vs. 34.2 min)
Hard problems that stay stuckNeon at xhigh, or Carbonxhigh adds 1 point over high for 70% more cost; Carbon is cheaper through November 30
Finance, legal, spreadsheets, and dense chartsNeon at highFirst on Vals Finance and Legal indices; high beats xhigh on AA's GDP.pdf
High-volume, price-sensitive routine callsHydrogen, or GPT-6.2 SolAA measures Sol (Max) at 57 and $1.01 per task
Vulnerability finding, pentesting, and security workApply to the Fairwind ProgramNeon for cyber is limited to program members and is not in the public API

For most teams, a practical baseline is moving existing Argon or Carbon workloads onto Neon high, dialing down to medium when allowance is tight, and reserving xhigh for tasks that stall at high.

FAQ

Can I use Neon on Go?
Yes. Go, Plus, and Pro all include Neon. It draws from the same monthly allowance as other models, with the standard 5-hour and weekly caps.
How is Neon metered?
Input, output, and cached input are metered at Google's official API prices: $3 / $15 / $0.15 per million tokens. Thinking tokens count as output. See the usage limits page for the full rules.
Can I use Neon in Claude Code?
Yes. Point ANTHROPIC_BASE_URL at CodingPass and launch with claude --model gemini-4-neon. Per-tool configuration snippets are on the models page.
Is Argon still available?
Yes. Argon remains on Google's API pricing page and continues to work as gemini-4-argon on CodingPass. Only the Gemini app model picker removed Argon.
Boarding

Gemini 4 Neon, in your agent.

Go, Plus and Pro all include Neon, metered at official API prices. No surprise charges, cancel anytime.

  • Request ID on every response
  • Full refund if you've used 5% or less
  • No surprise charges
Choose GoOpenCode guide

$10/mo until you cancel.