News

Google Releases Gemini 4 Neon: 1M Context at $3 per Million Input Tokens (Nov 19, 2026)

Google released Gemini 4 Neon on November 19, 2026, positioned between the Carbon flagship and the Hydrogen efficiency model. Neon is priced at $3 per million input tokens, $15 per million output tokens and $0.15 for cached input, with a 1 million-token context window, roughly 3x the speed and 60% of the cost of Carbon. It went live the same day across AI Studio, the Gemini app, the enterprise API and Antigravity, alongside a cyber-security variant.

One-line summary

$3, $15, one million tokens. Neon delivers roughly three times the speed of Carbon at 25% below Carbon's standard price — and unlike the previous releases, it shipped as available stock rather than a waitlisted promise.

Background ​

Google released Gemini 4 Neon on November 19, 2026, three weeks after the flagship Gemini 4 Carbon. Demis Hassabis described Neon as "Alchemy at its finest."

Unlike the earlier Gemini 4 releases, Neon was not gated behind a waitlist or limited access. It went live the same day in Google AI Studio, the Gemini app, the enterprise API, and Google's own coding environment, Antigravity.

Where Neon sits in the family ​

Gemini 4 now spans four tiers:

ModelRolePrimary use
CarbonFlagshipLong-horizon, hardest tasks
NeonMid-tier workhorseComplex coding, analytical reasoning, production enterprise workflows
ArgonIntermediateGeneral reasoning and multimodal work
HydrogenEfficiencyHigh-frequency everyday tasks, latency-sensitive work

Neon is explicitly positioned between Carbon and Hydrogen: frontier-class reasoning on complex coding and analysis, but faster and cheaper. Google's stated figures are roughly 3x the speed of Carbon and 60% of its cost.

Pricing and comparison ​

ItemGemini 4 NeonGemini 4 Carbon (standard)Gemini 4 Carbon (promo)
Input per 1M tokens$3$4$2
Cached input per 1M tokens$0.15——
Output per 1M tokens$15$20$10
Context window1M tokens1M tokens1M tokens
StatusEffective at launchStandard ratePromo through Nov 30, 2026

At $3/$15, Neon undercuts Carbon's standard $4/$20 by 25%, but sits above Carbon's current promotional $2/$10. Until November 30, Carbon remains cheaper on a per-token basis; from December 1, Neon becomes the more economical tier in the family.

Capability and safety ​

On the third-party Artificial Analysis composite index, Neon scored 61, placing third globally. For context, the GPT-6.2 Astra released the same week scored 62 and GPT-6.2 Sol scored 57. Some outlets also report that Neon cuts cost by roughly 17% and time by about 40% on representative tasks versus the prior generation.

Neon ships with a 1 million-token context window and four adjustable reasoning-depth levels, letting developers trade speed against depth per task. Google emphasizes that a single 1M-token window can hold spreadsheets, PDFs, charts and audio/video material at once, then consolidate them into structured deliverables.

On safety, Google says Neon carries forward Carbon's protections against severe misuse and prompt injection, holding up at its highest reasoning level in the Grayswan indirect prompt-injection test. A dedicated cyber-security variant was announced alongside the general model: AutoBaxBench hard tier 76%, WebExploitBench v1.2 51.26%, SEC-Bench Pro vulnerability detection 90.5%, and CWE-Bench v2 root-cause remediation 70.7%. Google says the variant outperforms Carbon on security tasks, costs roughly half of large frontier models on some tests, and will open over the coming days through the Fairwind program to vetted cyber defenders.

Two other things on the same day ​

Google also opened Genie 4 to eligible AI Ultra subscribers, letting users generate interactive worlds, change environments mid-exploration with text prompts, and hear responsive audio. Project Genie added an agent for world creation and a community gallery.

Separately, several large US school districts — including New York City, Los Angeles and Chicago — said they are re-evaluating Google Search because its AI features cannot easily be separated from traditional search results for students. That has nothing to do with the model release, but it shows Google's AI products are being scrutinized by both enterprise and education customers at once.

Why the mid-tier is the hardest slot ​

Google has now shipped four Gemini 4 models in roughly three months: Carbon for peak intelligence, Hydrogen for speed, Argon for general use, and Neon for volume. The difference between them is not whether capability exists, but where each sits on the cost curve.

The mid-tier is the hardest slot to fill. It has to be cheap enough to count as mid-priced, strong enough to justify "frontier reasoning," and fast enough to beat the flagship — otherwise developers simply use the cheapest model for bulk work and send the hard cases to Carbon, leaving the middle without a reason to exist. Neon's answer is to decouple context length from price: it keeps the one-million-token window that Carbon gates behind its top tier, then exposes four adjustable reasoning levels so a single model can handle both long-document consolidation and mid-complexity code work that must be right the first time.

What enterprise buyers should check ​

First, the ceiling on unit cost. At $3/$15, a billion output tokens a month lands around $15,000; the same volume at Astra's $40 output rate would be roughly $40,000. That gap is large enough to change a team's model choice.

Second, cache hit rates. Cached input at $0.15 is 5% of the list rate, so for applications that repeatedly query one knowledge base — support bots, document assistants — billing is driven by hit rate. For tasks where every input is new, the cache rate is irrelevant and the list price applies.

Third, regional and compliance availability. Shipping on day one does not mean every region, data-residency option or contract term is available at once. Enterprise buyers still need to confirm residency, log retention and contract terms separately.

Neon does not replace Carbon. A typical configuration runs Neon on high-volume extraction, rewriting, code completion and document work, and reserves Carbon for a small number of long-horizon, high-stakes calls — often at a ratio above nine to one.

Boundary ​

  • Release date, pricing, context window, reasoning levels and availability channels come from Google's own statements and reporting by Cailianshe and 36Kr, and are consistent across sources;
  • Artificial Analysis scores are third-party measurements; the index version may differ from the version used in this site's October Carbon dataset, so cross-version comparisons are not valid;
  • "3x faster" and "60% of the cost" are vendor claims without disclosed test methodology;
  • Cyber-security scores are Google's own evaluation;
  • The "17% lower cost / 40% faster" figures are media reporting, not an official conclusion;
  • This article does not rank the models or recommend a purchase.

Sources ​