Breaking
Loading news...

Google's Gemini 4 Neon Does Carbon's Job for 32% Less - and It Is the Fourth Model in a Month

Google's Gemini 4 Neon AI model does Carbon's job for less

Google introduced a new AI model on November 18, and the pitch is simple: nearly all of the smarts, for a lot less money. Gemini 4 Neon sits between Hydrogen, the fast everyday model, and Carbon, the heavyweight Google recommends for its hardest jobs. It is the fourth Gemini 4 model in about a month, after Argon in October and Carbon and Hydrogen earlier in November, and Google says it matches or beats Carbon on several benchmarks while costing 32% less per task at its default setting. Most of those numbers come from Google itself, hours after launch, so the victory lap is a claim for now rather than a settled fact.

What Neon actually does

Neon is built for the work that fills most of a developer's day: long coding sessions, multi-file debugging, and turning piles of documents into answers. It takes a 1 million-token context window, which means it can chew through spreadsheets, PDFs, charts, audio, and video in a single pass. On the benchmarks Google published at launch, Neon ranks third overall on the Artificial Analysis Intelligence Index with a score of 61 at its highest thinking level, first on the Artificial Analysis Coding Agent Index at 58.6, and second on the Vals Index at 70.9%. It also posts 65.2% on Terminal-Bench 4.1 and 84% on RiemannBench. Those are strong numbers for a mid-tier model, landing just behind the larger, slower models on the overall index.

Google is also selling Neon as an orchestrator, a model that directs fleets of faster Hydrogen subagents in its Antigravity coding platform while keeping the big picture in its own head. Early users quoted in the announcement sound pleased: Thai Tran, the AI product lead at Glean, said Neon matched or exceeded Carbon on overall quality at three times the speed and 60% of the cost. That is a customer talking, not a benchmark chart. Neon also records Google's lowest fabrication rate yet in internal Antigravity testing, which is a polite way of saying it makes things up less often than its siblings.

The price is the story

Neon costs $3 per million input tokens and $15 per million output tokens, with cached input at $0.15, a 95% discount. That puts its headline rates 25% below Anthropic's flagship Opus 5.5, which charges $4 and $20, and about 50% above the mid-range Claude Sonnet 5.5 and OpenAI's GPT-6.2 Sol, though Google says Neon beats both of those on the Artificial Analysis Intelligence Index at its highest setting. The wrinkle is the calendar. Carbon is still on introductory pricing of $2 per million input tokens through November 30, so for the next eleven days Neon actually costs more per token than its bigger sibling. From December 1, when Carbon jumps to $4 and $20, Neon becomes the cheaper option by 25%. Anyone budgeting a migration should check the date before doing the math.

Neon also comes with four thinking levels, so teams can dial the reasoning up or down and pay for only as much brain as the job needs. Google says the model pairs a high intelligence score with a 32% lower cost per task than Carbon at its default high setting, and roughly 50% to 80% lower cost per task than competing frontier models on the same benchmarks. Those per-task figures are Google's own calculations, and independent testing will either confirm them or not over the coming weeks.

A cyber version is coming too

Google is also shipping a security-focused variant, Gemini 4 Neon for cyber, to trusted defenders through its Fairwind Program over the next few days. The variant is built for the full code-hardening loop: finding vulnerabilities, checking which ones are actually exploitable, and patching them. Google reports 90.5% on SEC-bench Pro for finding flaws, 51.26% on WebExploitBench for validating them, and 70.7% on CWE-bench v2 for patching, each at roughly half the cost of competing large models. On the safety side, Neon keeps Carbon-level protections against severe misuse and prompt injection, and Google says it resisted every attack in Gray Swan's indirect prompt injection benchmark at its highest thinking level. That last claim comes from Google's own testing, but the direction is clear: Google wants Neon running in production, not just in demos.

Neon is rolling out now in the Gemini API in Google AI Studio, in Gemini for business, and in the Gemini app and Antigravity for paid Pro and Ultra subscribers. Four models in a month is a lot to keep straight, and the names will not help anyone remember which one does what. But if the early numbers hold up outside Google's own tests, Neon is the model most teams building with Gemini will actually reach for: close enough to the flagship to do the hard jobs, cheap enough to run all day, and fast enough that nobody has to wait around for it.

---------------
Author: Blake Taylor
New York News Desk