Gemini 4 Neon Takes Top Coding Rank Amid Google's Rapid-Fire Model Sprint and Pricing Paradox
Google introduced Gemini 4 Neon on November 18, 2026, positioning it between Hydrogen and Carbon for complex coding, enterprise analysis, and science workflows. While the model claims #1 on the Artificial Analysis Coding Agent Index and strong cybersecurity scores, its launch exposes temporary pricing inversions, a dropped sibling model, and heavy reliance on company-reported evaluations.
Story timeline · Put Gemini to work for you

On November 18, 2026, Google introduced Gemini 4 Neon, the fourth model in its Gemini 4 family in just over five weeks 123. Aimed at software engineering, enterprise knowledge work, and scientific research, Neon immediately claimed the top spot on the Artificial Analysis Coding Agent Index and drew attention for a specialized cybersecurity variant entering limited rollout 42. Yet behind the headline benchmarks lies a more complicated story about how Google is structuring its flagship AI lineup, how it prices reasoning depth, and how much of Neon's early acclaim rests on the company's own test harnesses 563.
What Happened and Why It Matters
Announced by Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect at Google, Gemini 4 Neon was built to sit between the heavyweight Gemini 4 Carbon and the fast, everyday Gemini 4 Hydrogen 278. Google says Neon delivers frontier-level reasoning with faster end-to-end execution across multi-step agent workflows, processing text, images, video, audio, and PDFs inside a unified context window of roughly one million tokens—specifically 1,048,576 tokens 1239.

For developers and enterprise teams, Neon is available immediately through the Gemini API under the identifier `gemini-4-neon`, as well as on Google AI Studio, Android Studio, Vertex-connected tools, Vercel's AI Gateway, and Google AI Pro and Ultra subscriptions 1043. It supports function calling, structured outputs, code execution, computer use, file search, and grounding via Google Search and Maps, while omitting image generation, audio generation, and the Live API 123.
The release matters because it signals a shift in how Google competes against OpenAI and Anthropic: rather than relying on a single monolithic flagship, Google is splitting agentic workloads across specialized models and configurable thinking tiers 423. Alongside the main release, Google unveiled practical demonstrations—such as a 3D freeride ski simulator built on Mont Blanc terrain, an interactive 3D reconstruction of Alan Turing's office from archival photos, and a 100,000-cell spreadsheet using 500,000 formulas to render depth and lighting from a single image 4.
The Evidence and How Strong It Is
The core claim driving headlines is Neon's first-place ranking on the Artificial Analysis Coding Agent Index with a score of 58.6, paired with a third-place overall ranking (score of 61) on the Artificial Analysis Intelligence Index v5.0.1 at its highest reasoning setting, trailing only Anthropic's Claude Fable 5.2 and OpenAI's GPT-6.2 Astra 429.
According to Artificial Analysis's published methodology, the Coding Agent Index averages scores across three benchmarks: FrontierCode-AA 2.0 (132 software engineering tasks), Terminal-Bench 4.1 (66 agentic terminal tasks), and SWE-Atlas-QnA-AA (124 technical Q&A tasks) 6.
However, Artificial Analysis explicitly cautions that its composite score does not mean a top-ranked model is best for every workflow and notes that outcomes vary significantly by model settings and execution harnesses 6.
That harness caveat is visible in Google's own numbers. On SWE-sweep, Neon scores 12.8% on the standard mini-SWE harness, a figure that more than triples to 39.0% when run inside Google's managed Antigravity harness 279. In other words, nearly two-thirds of that benchmark gain comes from the surrounding agentic orchestration environment rather than raw model weights alone 2.
Most of the quantitative claims circulating about Neon originate from Google DeepMind's November 18 launch post and model card rather than independent audits 298. Moreover, independent and secondary sources diverge on two notable technical points.
First, Ukrainian outlet Dzerkalo Tyzhdnya reported that Neon debuted in third place on the new Agent Arena Index after more than 4,200 real-world agent sessions 4, whereas Google DeepMind's primary announcement—echoed by FoneArena and AF.net—states Neon debuted at #2 globally with a score of 1,336 278.
Second, The Decoder wrote that Neon can produce up to one million output tokens 1, while developer documentation reviewed by Unite.AI and Kie.ai clarifies that input, thinking, and output tokens share a single 1,048,576-token window, with synchronous responses capped at 65,536 tokens by default 39.
Early independent hands-on testing also tempers the marketing narrative. Software engineer Simon Willison ran Neon across low, medium, and high reasoning levels on his comparative benchmark prompt and observed that Neon's outputs looked better than Hydrogen's but worse than Carbon's 5.
A Four-Week Sprint in Context
Neon's arrival caps an unusually compressed release schedule, even by the standards of Google's post-2022 AI acceleration. Google's AI trajectory spans foundational work from machine learning in Search spelling (2001), Google Translate (2006), TensorFlow (2015), AlphaGo and custom TPUs (2016), and the Transformer architecture (2017) to BERT (2019), AlphaFold (2020), and PaLM 2 (2023) 11.
Yet its modern consumer race began after OpenAI launched ChatGPT in November 2022, prompting emergency meetings at Google and the February 6, 2023 announcement of Bard—an initial rollout marred by an inaccurate James Webb Space Telescope demo that wiped 8%, or $100 billion, off Google's market value 12.
Google rebranded its effort as Gemini on December 6, 2023, retired the Bard name in February 2024, and iterated through the Gemini 1.5 and Gemini 3 series across 2025 12.
In autumn 2026, that pace compressed into days. According to API release documentation cited by Unite.AI, Google launched Gemini 4 Argon on October 13, 2026, Gemini 4 Carbon on November 10, 2026, Gemini 4 Hydrogen on November 12, 2026, and Gemini 4 Neon on November 18, 2026 3. DeepMind's model documentation confirms that Neon was directly adapted from Carbon 9.
Within the immediate 48-hour window around Neon's debut, Google also pushed adjacent Gemini updates: on November 17, it rolled out task-handoff features under its 'Put Gemini to work for you' campaign, and later on November 18 it introduced temporary 'incognito' chats in Search AI Mode alongside Genie 4, an interactive 3D world model under Project Genie. While those consumer and research announcements share corporate timing rather than technical causality with Neon, they underscore Google's broad November product offensive.
Actors, Positioning, and the Pricing Paradox
Google positions Neon as the workhorse for high-frequency complex tasks, reserving Carbon for the hardest long-horizon architecture problems and Hydrogen for fast everyday execution 23. In Google Antigravity, Neon can either work solo or act as an orchestrator directing parallel Hydrogen subagents 427. Tech analysts at The Decoder liken Carbon and Argon to OpenAI's Astra or Anthropic's Fable, casting Neon as Google's counterpart to Claude Opus or Sol 1. Vercel moved on launch day to integrate `google/gemini-4-neon` into its AI Gateway for coding agents like Codex and Cursor 10.
Yet Google's pricing structure has drawn sharp criticism from developers. Neon is priced at $3.00 per million input tokens, $0.15 per million cached input tokens (a 95% discount applied automatically via implicit caching on prompts of at least 4,096 tokens), and $15.00 per million output tokens, with thinking tokens billed at the output rate 123. That sits above Hydrogen ($1.50 input / $7.50 output, which also offers a free API tier) and, crucially, above Carbon and Argon's introductory rate of $2.00 input / $10.00 output 513.
Because Carbon and Argon's 50% discount runs through November 30, 2026, before doubling to $4.00 input and $20.00 output on December 1, Neon enters the market temporarily more expensive per token than the larger flagship Carbon it sits beneath 513.
Following ChatGPT's November 2022 debut, Google launched Bard in February 2023 and rebranded its generative AI platform as Gemini in December 2023, rolling out Gemini 1.5 and 3 across 2025 before shipping four Gemini 4 models—Argon, Carbon, Hydrogen, and Neon—between October 13 and November 18, 2026.
On December 1, 2026, introductory half-price discounts on Gemini 4 Carbon and Argon expire, resolving the temporary pricing inversion where Neon is costlier per token than Carbon. Developer adoption via Vercel, Cursor, and Antigravity, alongside the limited Fairwind Program rollout of Neon for cyber, will test Google's cost-per-task claims.
Written with AI from the cited sources and checked against them. Open the sources to verify.