AI

Gemini 4 Neon: Independent Benchmarks, $3/$15 Token Pricing, and Production API Trade-Offs

Released on November 18, 2026, Gemini 4 Neon scores 61 (#3 overall) on the Artificial Analysis Intelligence Index at $3/$15 per million tokens and ~110 output tokens per second. While Google DeepMind highlights 40% faster end-to-end task completion than Gemini 4 Carbon, independent testing shows Neon uses ~15% more output tokens per task and temporarily costs more per token than Carbon's discounted rate through November 30. Here is the full benchmark, pricing, and routing breakdown.

By 0xScorPionSon

7 min read

Gemini 4 Neon: Independent Benchmarks, $3/$15 Token Pricing, and Production API Trade-Offs

Gemini 4 Neon Architecture showing 1M context core, four-tier reasoning router, and $3/$15 per 1M API pricing

Executive Summary

  • Model & Release: Gemini 4 Neon, Google DeepMind's fourth Gemini 4 family release in seven weeks, announced on November 18, 2026 (Google Blog, Artificial Analysis).
  • Core Capabilities & Specs: 1,048,576 (1M) token context window, multimodal input (text, image, video, speech) with text output, and four configurable thinking effort levels (low, medium, high, xhigh) (Artificial Analysis).
  • API Pricing: $3.00 per 1M input tokens, $0.15 per 1M cached input tokens (95% discount), and $15.00 per 1M output tokens (thinking tokens bill at the output rate; no free API tier) (Google Blog, NeuralTrust).
  • Primary Developer Takeaway: On independent Artificial Analysis testing, Neon (xhigh) scores 61 on the Intelligence Index (#3 overall) at ~110 output tokens/sec and $6.44 per task—~40% faster than Gemini 4 Carbon (high, ~36 min vs ~21 min). However, Neon emits ~15% more output tokens per task, and until Carbon's introductory $2/$10 discount expires on November 30, 2026, Carbon remains cheaper per token (Artificial Analysis).

Architecture & How It Works

Within the Gemini 4 lineup, Google DeepMind positions Gemini 4 Neon between Gemini 4 Hydrogen (fast everyday execution with free API tier access) and Gemini 4 Carbon (heaviest long-horizon projects) (Google Blog). Neon is engineered specifically for high-frequency agentic loops, multi-file repository edits, and visual/document synthesis across a unified 1M-token context window with flat token pricing (NeuralTrust).

Thinking Effort Routing & Orchestrator-Worker Pattern

Rather than a single fixed reasoning budget, Neon exposes four effort tiers (low, medium, high [API default], and xhigh) (Artificial Analysis). In Google Antigravity, Neon also operates as a fast Orchestrator that delegates parallel subtasks to smaller Hydrogen worker subagents while managing context compaction across long sessions (Google Blog).

[Multimodal Inputs: Text / Image / Video / Speech (up to 1M tokens)]
                                  │
                                  ▼
                   [Gemini 4 Neon API / Antigravity]
                     ├─ effort: low / medium ──▶ Fast classification & tool routing
                     ├─ effort: high (default) ─▶ $3.74/task avg (AA Index ~59.6)
                     └─ effort: xhigh ─────────▶ $6.44/task avg (AA Index 61, #3)
                                  │
                    ┌─────────────┴─────────────┐
                    ▼                           ▼
         [Direct Tool Execution]     [Parallel Hydrogen Subagents]

API Request Configuration Example

Because thinking tokens bill at the $15/1M output rate, production services should explicitly set the thinking level per route rather than leaving every endpoint on xhigh:

{
  "model": "gemini-4-neon",
  "generationConfig": {
    "thinkingConfig": {
      "thinkingLevel": "high"
    },
    "temperature": 0.2
  },
  "contents": [
    {
      "role": "user",
      "parts": [
        { "text": "Audit this repository diff for SQL injection and race conditions." }
      ]
    }
  ]
}

At its default high effort, Neon avoids the extra $2.70 per-task reasoning overhead of xhigh while retaining near-peak coding and analytical accuracy (dev.to / Studio1HQ).


Benchmarks, Cost & Latency Analysis

Separating Google DeepMind's self-reported evaluation suite from independent third-party telemetry highlights where Neon genuinely leads—and where harness or pricing nuances matter.

Independent & Vendor Benchmark Comparison

Benchmark / Metric Gemini 4 Neon Claude Opus 5.5 GPT-6.2 Sol Verification Source
AA Intelligence Index v5.0.1 61 (xhigh, #3 overall) 59 57 DeepMind, Artificial Analysis
Vals Index v2.2 70.9% (#2) 70.1% 66.1% DeepMind
AA Coding Agent Index 59 (xhigh, Antigravity) / 58.6 58 — Artificial Analysis, Google Blog
Terminal-Bench 4.1 65.2% 65.8% 61.5% DeepMind
FrontierCode 2.0 46.2% (high) 47.3% 43.8% DeepMind
SWE-sweep (Provider / Mini-SWE) 39.0% / 12.8% 10.1% / 3.6% 8.3% / 11.6% DeepMind
Zapier AutomationBench 56.3% (#1) 42.5% 41.3% DeepMind
GraphWalks (256k–1M BFS F1) 95.2% 66.8% 80.1% DeepMind
HLE-Diamond / RiemannBench 71.3% / 84.0% 67.9% / 80.0% 64.0% / 80.0% DeepMind
SEC-bench Pro / CWE-Bench v2 90.5% / 70.7% — / 67.1% 88.3% / 66.0% DeepMind

Cost-Per-Task, Speed & Benchmark Caveats

  1. The Harness Multiplier on SWE-sweep: Neon's SWE-sweep score jumps more than 3× from 12.8% on the generic mini-SWE harness to 39.0% inside Google Antigravity (Google Blog, DeepMind). Teams calling Neon over raw APIs without Antigravity's tool harness should benchmark against the 12.8% baseline rather than assuming 39.0%.
  2. Token Verbosity vs. Token Price: According to Artificial Analysis, Neon (xhigh) generates ~15% more output tokens per task than Gemini 4 Carbon (high). Its $6.44 average cost per task (~17% below Carbon's $7.76 standard rate) is driven entirely by its 25% lower list pricing ($3/$15 vs. $4/$20).
  3. Throughput Advantage: Neon streams at ~110 output tokens/second versus ~57 tokens/second for Carbon—cutting average end-to-end task wall time by ~40% (~21 minutes vs. ~36 minutes across the AA suite) (Artificial Analysis).
  4. Coding Agent Cost Comparison: On the AA Coding Agent Index, Neon in Antigravity averages $6.86 per task, nearly half of Claude Opus 5.5 (max, $12.61) and less than half of Claude Fable 5.2 (xhigh, $15.47) (Artificial Analysis).
  5. Multimodal & Document Synthesis: On AA-Briefcase, Neon (xhigh) leads Presentation Quality at 1771 Elo (45 points ahead of Claude Sonnet 5.5 max at 1726) and hits 56% Climb WSR on the Multimodal Knowledge Agent (MKA) evaluation (Artificial Analysis).

Production Deployment & Trade-Offs

When to Choose Gemini 4 Neon vs. Alternatives

  • Choose Gemini 4 Neon (high or xhigh) for latency-sensitive agentic coding, long-context retrieval (256k–1M tokens, where it scores 95.2% F1 on GraphWalks), and visual document pipelines (PDFs, charts, spreadsheets) where ~110 tok/s throughput and $0.15/1M cached input pricing materially cut runtime and cost (DeepMind, Artificial Analysis).
  • Stick with Gemini 4 Carbon (until November 30, 2026) if you are running batch workloads that are not latency-bound: Carbon's promotional $2/$10 pricing remains 33% cheaper per token than Neon until December 1, 2026 (Artificial Analysis, NeuralTrust).
  • Use Gemini 4 Hydrogen for high-volume worker subagents, simple classification, or projects requiring a free API tier—Neon has no free API tier (dev.to / Studio1HQ).
  • Evaluate Claude Opus 5.5 if your workload centers on raw CLI/container coding benchmarks like Terminal-Bench 4.1 (65.8% vs. 65.2%) or FrontierCode 2.0 (47.3% vs. 46.2%), where Opus 5.5 still edges Neon by roughly a point (DeepMind, dev.to / Studio1HQ).

Operational Gotchas

  • Thinking Token Billing: Because reasoning tokens are billed as $15/1M output tokens, verbose xhigh prompts without explicit thinkingLevel caps can inflate bills unexpectedly (NeuralTrust).
  • Security & Alignment Profile: Google reports 0.0% attack success on Gray Swan's indirect prompt injection (IPI) benchmark at xhigh with real-time safeguards enabled, alongside a 0.09% misaligned action rate and 1.98% false refusal rate on Terminal-Bench 4.0 (Google Blog). Keep real-time safeguards enabled when processing untrusted web pages or emails.

Frequently Asked Questions (FAQ)

How much does the Gemini 4 Neon API cost?

Gemini 4 Neon costs $3.00 per 1M input tokens, $0.15 per 1M cached input tokens (95% discount), and $15.00 per 1M output tokens—flat across the entire 1M-token context window (Google Blog, NeuralTrust). Thinking tokens are billed at the output rate, and there is no free API tier.

Should I use high or xhigh thinking effort in production?

Start with high (the API default). Independent testing shows high delivers a 59.6 AA Index score at an average of $3.74 per task, whereas xhigh adds roughly 1.8 index points (reaching 61) at $6.44 per task and higher latency (dev.to / Studio1HQ, Artificial Analysis). Reserve xhigh for hardest-tier engineering or scientific tasks.

How does Gemini 4 Neon compare to Gemini 4 Carbon?

At standard pricing, Neon ($3/$15) is 25% cheaper per token than Carbon ($4/$20), streams nearly twice as fast (~110 vs. ~57 output tok/s), and scores 1 point higher on the AA Intelligence Index (61 vs. 60) (Artificial Analysis). However, Neon emits ~15% more output tokens per task, and Carbon's discounted $2/$10 rate remains active through November 30, 2026 (Artificial Analysis).

Does Gemini 4 Neon beat Claude Opus 5.5 on coding?

It depends on the harness and benchmark. On SWE-sweep inside Google Antigravity, Neon scores 39.0% vs. 10.1% for Opus 5.5, and on the AA Coding Agent Index Neon averages $6.86 per task vs. $12.61 for Opus 5.5 (max) (DeepMind, Artificial Analysis). On Terminal-Bench 4.1 (65.8% vs. 65.2%) and FrontierCode 2.0 (47.3% vs. 46.2%), Claude Opus 5.5 holds a slight edge (DeepMind).

Where can developers access Gemini 4 Neon today?

Gemini 4 Neon is available in the Gemini API via Google AI Studio and Android Studio (paid tier), Gemini for business, Google Antigravity (Pro/Ultra subscribers), the Gemini app (Pro/Ultra), and for cybersecurity defenders via the Fairwind Program (Google Blog).


Sources

More on AI

AI

GPT-6.2 Astra vs Gemini 4 Carbon and Claude Opus 5.5: Token Cost Inversion, Code Review Noise, and Jailbreak Regressions

OpenAI launched GPT-6.2 Astra and Sol on November 17, 2026, alongside a new System Card and Codex sandbox architecture. While Astra leads Claude Opus 5.5 on Intelligence Index v5.0.1 (62.1 vs 59.1) and costs 2.5x less per task despite double the rate-card price, teams face a 272K long-prompt surcharge, noisier PR review comments, and a reported universal jailbreak regression. Here is how Astra, Sol, Gemini 4 Carbon, and Claude Opus 5.5 compare.

7 min read

AI

LiteLLM Proxy Flaw (CVE-2026-40217) Added to CISA KEV: How Guardrail Bytecode Rewriting Enables LLMjacking

On November 18, 2026, CISA added BerriAI LiteLLM CVE-2026-40217 (CVSS 3.1: 8.8) to its Known Exploited Vulnerabilities catalog with a December 2 deadline under BOD 26-04. Attackers are exploiting the /guardrails/test_custom_code sandbox escape in LLMjacking campaigns to steal provider API keys and run code as root in default Docker containers. Learn why you must target 1.83.11+ rather than 1.83.10 and how to block the endpoint at your gateway.

7 min read

AI

Microsoft agent-framework PyPI Hijack: How Two Malicious Releases Used .pth Hooks to Steal AI and Cloud Keys

On November 17, 2026, attackers published two malicious releases (1.26.1 and 1.26.2) of Microsoft's official agent-framework Python AI SDK on PyPI within 33 minutes. Using .pth startup hooks that fire on every Python invocation—even without importing the package—the payload harvested environment variables, /proc/*/environ secrets, and cloud/AI keys, and attempted to mint Azure DevOps, GitHub Actions OIDC, and Azure managed-identity tokens. This defensive incident breakdown covers the full timeline, SHA-256 IOCs, safe non-Python containment steps, and token revocation.

6 min read

Tell us what is not working

A thirty-minute call, no deck and no obligation. You will leave with a straight answer about whether this is worth building - including if that answer is no.

Start a project

Direct line

hello@codebumble.net
Dhaka, Bangladesh  :   local timeCalls in English or Bangla