Google filled the gap in its Gemini 4 lineup on November 18 with a new model called Gemini 4 Neon. I read Google's official announcement and the model card the same day, and one thing stood out to me: this is not a watered-down flagship. It is a workhorse aimed at the jobs developers actually run every day — long coding sessions, messy spreadsheets, and research that spans dozens of documents. Here is what I found, what the numbers say, and where I would still be careful.
Where Neon sits in the Gemini 4 family
Google now has three tiers, and Neon is the middle one. At the top sits Gemini 4 Carbon, the deepest reasoner for the hardest, longest-horizon work. At the fast end sits Gemini 4 Hydrogen, built for everyday execution at speed. Neon is positioned between them: frontier-level intelligence, but tuned for fast end-to-end completion of complex work rather than maximum depth on a single problem. That middle slot is deliberate. Most real workloads are not "solve the hardest problem ever" — they are repeated, complex jobs: refactor a repo, reconcile a pile of invoices, test a hypothesis across ten papers. Neon is Google's bet that this is where the volume — and the money — lives.
The price tag
The developer price is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million — a 95% discount. The context window is 1 million tokens, and the rate stays flat across the full window. One timing detail matters. Through November 30, Carbon is still on an introductory discount at $2 input / $10 output, so Carbon is actually cheaper per token right now. From December 1, Carbon's standard rate doubles to $4 / $20, and Neon becomes 25% cheaper per token than Carbon. Google also claims a 32% lower cost per task than Carbon at the default high thinking level, because Neon finishes jobs with fewer tokens. That is Google's claim, not an independent measurement — I will come back to that. Neon also adds a fourth thinking level. It defaults to high, and adds an xhigh level above it that Carbon and Argon do not offer, so you can trade more reasoning depth against token usage when a task needs it.
The benchmarks I checked
Google's announcement leans heavily on third-party leaderboards, and the numbers are strong — with the caveat that Google chose which ones to highlight: Coding: #1 overall on the Artificial Analysis Coding Agent Index with a score of 58.6, plus 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at its default high setting. General intelligence: #3 overall on the Artificial Analysis Intelligence Index with a score of 61 at its xhigh thinking level, just behind larger, slower models, and #2 on the Vals Index. Science and maths: best in class, according to Google, on Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%). Enterprise work: #1 on most Surge AI benchmarks across finance, accounting, spreadsheet modelling and complex constraint-following. Steerability: #2 on the new Agent Arena Index, which measures how cleanly a model adapts when you correct it mid-task — a genuinely important trait for agentic work. Early users add colour. Glean's AI product lead said in Google's announcement that in its early evaluations, Neon matched or exceeded Carbon on overall quality at 3x the speed and 60% of the cost. Google published no absolute completion times for that claim, so treat the "3x" as an early-user impression rather than a timed benchmark. On safety, Google reports a 99.2% pass rate on its internal agent honesty test and its lowest fabrication rate yet in internal Antigravity testing, plus stronger resistance to indirect prompt injection. A cybersecurity variant is also coming for approved defenders in Google's Fairwind Program.
Built for long, messy work
The part of the announcement that convinced me most is not a benchmark — it is the workflow design. Enterprise work lives in spreadsheets, PDFs, charts, audio and video, and Neon can pull all of those into a single 1-million-token context and turn them into structured deliverables. For coding, it is engineered for long-running agent workflows: it cuts down on redundant file reads, shares progress as it works, and handles context compaction reliably in long sessions. In Google Antigravity it can act as a fast, high-judgment orchestrator directing parallel Hydrogen subagents. That puts Neon in direct competition with the other mid-tier models launched this month — OpenAI's GPT-6.2 Astra and Sol and Anthropic's Claude Sonnet 4.6. Google's pitch is that Neon beats both on the Artificial Analysis Intelligence Index at its highest reasoning setting, despite costing more per token than either at list price. The real contest will be cost per finished task, not cost per token — and that is exactly the metric Google is pushing.
Why this matters — my take
The mid-tier is where AI stops being a demo and starts being infrastructure. Flagships win headlines; mid-tier models win the daily workload. If Neon's benchmark scores hold up in independent testing, it becomes the default engine for agentic coding pipelines — the model you point at a repo on Monday morning and trust to still be coherent by Friday. The steerability ranking is the detail I would watch most. Agents that take mid-task corrections cleanly are the difference between a tool you supervise and a tool you babysit. A #2 ranking there, if it survives independent testing, matters more than a point or two on a coding index.
The honest limits
Every benchmark in this article comes from Google's announcement. Independent verification is still pending, and vendors always publish their best numbers. Carbon is cheaper per token until November 30 under its introductory discount. Neon's price advantage starts on December 1. The "3x speed" claim comes from one early user's evaluation, with no published completion times. The model card lists a knowledge cutoff of March 2026, so very recent events are outside its training. The cybersecurity variant is limited to Fairwind Program defenders — it is not a general release.
The bottom line
Gemini 4 Neon looks like the Gemini 4 model most teams will actually run day to day: flagship-grade coding scores, a 1-million-token context, and a price that undercuts Carbon from December 1. The open question is whether the benchmarks survive contact with independent testers — and whether the "cost per finished task" story holds up in real repos, not just Google's test suite. I will be watching the independent results as they land. Would you switch your daily coding workflow to Neon, or stick with Carbon, Sonnet or GPT-6.2? Tell me in the comments.
Sources
Google, "Introducing Gemini 4 Neon," The Keyword, November 18, 2026: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/
VentureBeat, "Google launches Gemini 4 Neon with Opus 5.5-level coding performance at 25% lower API token prices," November 18, 2026: https://venturebeat.com/technology/google-launches-gemini-4-neon-with-opus-5-5-level-coding-performance-at-25-lower-api-token-prices


Comments
Post a Comment