Google Gemini 4 Neon Tops Coding Benchmarks: 58.6 on Artificial Analysis, Beats Opus 5.5 and GPT-6.2 Sol
Google released Gemini 4 Neon on November 18, achieving 58.6 on Artificial Analysis's coding agent index — ranking first — with detailed benchmarks showing it outperforms Opus 5.5 and GPT-6.2 Sol across terminal tasks, repository merging, science, math, reasoning, and long-context retrieval, while offering 100k token context and competitive pricing at $3/M input and $15/M output.
On November 18, Google released Gemini 4 Neon, positioning it as a model optimized for deep, frequently run engineering, knowledge, and science workloads — contrasted with Carbon for the hardest long-running projects and Hydrogen for everyday speed tasks.
Model Specifications
According to the model card, Neon is derived from Carbon. It accepts text, image, audio, and video inputs with a total context window of 100k tokens . Synchronous output defaults to 64k tokens , while asynchronous background generation can reach 100k tokens . The model offers four thinking tiers: low, medium, high, and ultra-high.
Coding Agent Benchmark (Artificial Analysis)
On the Artificial Analysis coding agent index, Neon scores 58.6 , placing first. The blog notes this is not the same leaderboard as the intelligence index shown in the comparison table.
Intelligence Index Comparison (Three-Tier Table)
The official three-model comparison table shows intelligence index scores at the high tier: Carbon 60 , Neon 61 , Hydrogen 54 . Carbon's column only extends to the high tier. Pricing footnotes indicate a Carbon promotion through November 30.
Access Requirements
In the App and Antigravity interfaces, Neon requires Pro or Ultra subscription. It is available via the AI Studio paid tier and in the enterprise model list. Hydrogen is marked as available to all users; Neon is not. The cybersecurity tier requires Fairwind access, rolling out to trusted defenders in coming days.
Terminal Benchmark (Post Decontamination)
Neon: 65.2%
Opus 5.5: 65.8%
GPT-6.2 Sol: 61.5%
Repository Merge Benchmark
Main-branch merge tasks at default high tier:
Neon: 46.2%
Opus 5.5: 47.3%
Sol: 43.8%
With Google's internal workflow applied to the same repo tasks, Neon reaches 39.0% ; without it, only 12.8% . Sol without workflow scores 11.6% , close to Neon's no-workflow result.
Ultra-High Tier Intelligence Index
At ultra-high tier, Neon scores 61 , described as third overall behind two larger, slower models (unnamed). The model card table lists Opus 5.5 at 59 and Sol at 57 for this tier.
Science, Math, Reasoning, and Professional Exams
Science terminal: Neon 70.5% , Opus 5.5 63.3% , Sol 68.0%
Math: Neon 84.0% , both competitors 80.0%
Reasoning: Neon 71.3% , Opus 5.5 67.9% , Sol 64.0%
Multi-step professional exams: Neon 75.8% , Opus 5.5 71.6% , Sol 73.0%
Vals index: Neon 70.9% , Opus 5.5 70.1% , Sol 66.1%
Chart Reading and Long Context
Dense chart reading: Sol 75.1% , Neon 73.5% (Sol leads)
Long-context retrieval: Neon 95.2% , Opus 5.5 66.8% , Sol 80.1%
Real-World Agent Interaction
Mid-task correction adherence ranks second (raw score unpublished). A separate real-world agent session index scores 1336 (global second, 4,200+ samples), distinct from the 58.6 coding agent index.
Pricing
Neon: $3.00/M input, $0.15/M cached input, $15.00/M output
Carbon (standard): $4.00/M input, $20.00/M output; promo through Nov 30: $2.00/$10.00
Hydrogen: $1.50/M input, $7.50/M output
The blog cites a 32% lower per-task cost at default high tier versus competitors (not simple price division). At ultra-high tier, per-task cost is at least 36% below the highest competitor tier.
Cost vs. Intelligence Scatter Plot
A chart plots per-task cost (x-axis) against intelligence index (y-axis). Neon's curve spans low to ultra-high tiers. Carbon and Opus 5.5 reach comparable intelligence only at higher cost (further right). Sol's points cluster on the lower-cost, lower-intelligence side.
Cybersecurity Benchmarks
Hard-tier code generation safety vs. Carbon: 76%
SEC-bench Pro vulnerability detection: Neon 90.5% , Sol 88.3% , Opus 5.5 not reported
Web exploitation: Neon 51.3% , Opus 5.5 37.1%
The blog notes some tasks cost roughly half as much (still per-task cost). Fairwind for defenders is not yet released alongside the API.
Knowledge Cutoff and Safety
Knowledge cutoff: March 2026 . Hallucination rates are within known limits but no proportion is given. A full safety report is forthcoming.
Sources
Google Blog, Nov 18:
https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/Model Card, Nov 18:
https://deepmind.google/models/model-cards/gemini-4-neon/Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
