📑 Table of contents

Gemini 4 Neon: Google accelerates again, 24 hours after Hydrogen

LLM & Modèles 🟢 Beginner ⏱️ 14 min read 📅 2026-11-19

Gemini 4 Neon: Google Accelerates Further, 24 Hours After Hydrogen

🔎 A Pace That Is No Longer Normal

On November 18, 2026, Google DeepMind unveiled Gemini 4 Neon. Its fourth model in the Gemini 4 family in barely two months. To put things in perspective: between Gemini 3.1 Pro (February 2026) and the Gemini 4 wave, Google had taken its time. Today, the machine has gone into overdrive.

The pace is dizzying. According to data from Arena, 28 days passed between Argon and Carbon, 15 days between Carbon and Hydrogen, and only 6 days between Hydrogen and Neon. The site LLM Releases, which tracks every model release, has simply never recorded such a density of major launches from a single laboratory.

The title of this article mentions "24 hours"—this is a simplification of Google's communication, which announced Neon immediately following Hydrogen's release on November 12. In reality, the exact gap is 6 days. But the perceived effect is the same: you barely finish testing one model before the next one arrives. This is no longer a release strategy; it is a saturation strategy.

What this frenzy says about the market, how Neon positions itself against Carbon and Argon, and most importantly, what you need to do about it: we break it down, backed by sourced figures.


The essentials

  • Gemini 4 Neon was released on November 18, 2026, the fourth Gemini 4 model in roughly 7 weeks, following Argon, Carbon and Hydrogen (latent.space).
  • Pricing: $3/M input tokens, $15/M output tokens, positioned between Hydrogen ($1.50/$7.50) and Carbon ($2 intro then $4/$10 intro then $20) (shattered.io).
  • 1M token context and 4 adjustable "thinking" levels, a first for Google, letting users choose the quality/latency/cost trade-off.
  • Neon replaces Argon in the Gemini app according to 9to5Google, while Argon's status on the API side remains unclear.
  • Positioning: a direct competitor to Anthropic's "Opus/Sol" models, i.e., the frontier reasoning niche for high-frequency workflows.
  • Google's release cadence is now nearly monthly, as Sundar Pichai had announced back in July (emergent.sh).

Model Main use case API input/output pricing (November 2026, check blog.google) Best for
Gemini 4 Hydrogen Everyday use, fast $1.50 / $7.50 per M tokens Common tasks, Flash replacement
Gemini 4 Argon Code and analysis $2 intro then $4 / $10 intro then $20 Development, data analysis
Gemini 4 Carbon Most complex work $2 intro then $4 / $10 intro then $20 Long-horizon coding, research
Gemini 4 Neon Frontier reasoning $3 / $15 per M tokens High-frequency enterprise workflows

To choose between these models based on your use case, our monthly comparison of the best LLMs is updated with every major release — and it's had plenty of work to do these past few months.


What does Gemini 4 Neon actually deliver?

Neon is the "frontier" model of the Gemini 4 lineup, designed for heavy reasoning at high frequency, with a 1M token context and four adjustable thinking levels. That is essentially the reading offered by shattered.io, which published the most detailed analysis of the model to date.

A deliberate pricing position: between the fast and the massive

The pricing grid of the Gemini 4 family tells Google's strategy all on its own:

Model Input ($/M tokens) Output ($/M tokens) Role
Hydrogen $1.50 $7.50 Volume, everyday
Argon $2 intro then $4 $10 intro then $20 Code and analysis
Carbon $2 intro then $4 $10 intro then $20 Maximum capacity
Neon $3 $15 High-frequency frontier reasoning

Neon sits at $3/$15: twice Hydrogen's input price, but half Carbon's output price once the introductory pricing ends. The logic is crystal clear: Neon is built for companies that need very high-level reasoning, but at volumes that make Carbon economically hard to justify.

This is exactly the niche Anthropic occupies with its Opus models — and latent.space doesn't miss it, headlining Neon directly as a "contender" to Opus/Sol at Google DeepMind.

Four thinking levels: the trade-off becomes a parameter

The other novelty of Neon is its 4 "thinking" levels. Concretely, you choose how much the model thinks before answering — from fast mode for simple tasks to deep reasoning for complex problems.

This is no small thing. Until now, that choice was made implicitly by picking a model (the fast one to move quickly, the heavy one to think). With Neon, a single model covers the entire spectrum, and it's an API call parameter. For teams building products on AI APIs, it's a welcome simplification: one model to integrate, one slider to adjust depending on the use case.

1M tokens of context: the norm settles in

Neon's 1M token context is no longer an innovation in itself — Hydrogen already offers a flat price across its entire million tokens (emergent.sh). But combining it with frontier reasoning changes the game: analyzing a complete codebase, a legal file several hundred pages long, or an entire research corpus in a single call, with maximum thinking level, becomes a standard use case rather than a feat.


Why is Google releasing four models in two months?

Because the frontier model market has become a war of release cadence, and Google has decided to win it through saturation. Sundar Pichai announced it back in July: a near-monthly cadence for Gemini models (emergent.sh). He kept his word — and then accelerated even further.

Fragmentation by use case, not increment

What strikes you about the Gemini 4 family is that these four models are not successive iterations of the same product. They are four distinct products:

  • Argon: the versatile code and analysis model, first in the series.
  • Carbon: the heavyweight model, launched two weeks after Argon, initially reserved for the Fairwind Program for cyber defenders and trusted partners (blog.google).
  • Hydrogen: the fast, low-cost model, successor to Flash, released on November 12 — we covered it in detail in our article on Gemini 4 Hydrogen and its 51 points on the Artificial Analysis index.
  • Neon: the high-frequency frontier reasoning model, released on November 18.

Google is no longer releasing "a better model": it is releasing a lineup. This is the transposition of hardware logic (an iPhone lineup at multiple price points) to the world of foundation models. And it works because use cases have genuinely specialized: nobody needs a frontier model to summarize an email, and nobody uses a Flash model for a sensitive code audit.

The competitive pressure is real

This cadence is not vanity. When TheDecoder notes that Neon is the "4th Gemini 4 model in 4 weeks," it is also a message sent to competitors: occupy the space. Each launch fills a pricing and capability niche before OpenAI or Anthropic can do so.

The internal hierarchy is explicit, moreover: Carbon remains the maximum-capability model, and Neon slots in below it with a better capability-to-price ratio for large volumes. Google cannibalizes its own models before the competition does.

For a cross-cutting look at the Google vs OpenAI duel, our ChatGPT vs Gemini comparison remains the most effective entry point.


Neon replaces Argon in the Gemini app: what about the API?

In the Gemini app, Neon replaces Argon as the default model. On the API side, Argon's status is not specified — and that's the major point of friction for developers.

This is the most important piece of information for technical teams, and the least documented. If you've built on Argon since its launch, you're in a zone of uncertainty typical of accelerated release cycles: your production model may become a "legacy" model six weeks after its release.

Three scenarios are emerging:

  1. Argon remains available via the API with its introductory pricing extended ($4/$20), and becomes the stable mid-range option.
  2. Argon is gradually deprecated, and Neon + Hydrogen cover its niche from above and below.
  3. Argon merges into Neon as a thinking level, which Neon's 4 levels make technically plausible.

Google has not publicly ruled on any of these scenarios. My advice: if you're running Argon in production, build an abstraction layer over your API calls now. At this pace, the question isn't whether your model will be replaced, but when. And for projects that can't afford this instability, running an LLM locally remains a stability option that cloud APIs will never offer.


What this release frenzy says about the market

The frontier LLM market has entered its commercial maturity phase: differentiation no longer comes from some magical capability, but from pricing packaging and specialization by use case.

The end of "one model to rule them all"

Look at the Gemini 4 lineup as a whole: one model at $1.50, two at $4, one at $3. None of them crushes the others across the board. Carbon is the most capable but the most expensive on output. Hydrogen is the cheapest but plateaus on complex reasoning. Neon occupies the middle ground with a precise value proposition.

This is exactly the structure of a mature market. Compare that with 2023-2024, when every launch aimed for the absolute top spot in the rankings. Today, Google is releasing four models for four segments — a logic that OpenAI is also following with GPT-6.1 Terra, its half-price mid-range model — and which forces comparisons to rank by use case, not by overall score.

Benchmarks become secondary, cost per task becomes central

A sign of the times: Neon coverage focuses less on raw scores than on the capability-to-price ratio. vals.ai has already opened its benchmarks and cost page for the model, but the dominant discussion in the analyses (shattered.io, latent.space) revolves around pricing positioning and thinking levels, not MMLU points.

For businesses, the question is no longer "which is the best model?" but "which model offers the best cost for MY task, at this volume?". It's a paradigm shift that structurally favors Google, whose infrastructure enables aggressive pricing across the entire lineup.

The risk for users: decision fatigue

There's a flip side. Four models in two months also means four spec sheets to read, four potential migrations, four prompt recalibrations. Teams that optimized their applications on Argon must already be looking at Neon. Those who had calibrated on Hydrogen are discovering that a "better" model exists for barely more money.

My position is simple: at this pace, don't migrate on reflex. Migrate when a model solves a concrete problem that yours doesn't, or when the cost per task justifies the migration work. Otherwise, you'll spend your time migrating instead of building.


How to choose within the Gemini 4 range?

Choose by use case and volume, not by prestige. Here's the decision grid I recommend, in light of the data from shattered.io:

Your situation Recommended model Why
Chatbot, summarization, daily tasks Hydrogen ($1.50/$7.50) The best price/performance ratio on the market
Development, code analysis Argon or Neon Argon if budget is tight, Neon if complex reasoning is recurring
Research, audit, long-horizon coding Carbon Maximum capability, despite the $20 output price after the intro period
Agents in production, high volumes Neon Frontier reasoning at $15/M output, sustainable

Two important caveats:

  • The introductory prices for Argon and Carbon ($2/$10) are temporary and will rise to $4/$20. If you're planning an annual budget, calculate based on the final prices, not the promotional ones.
  • The input cache at a 95% discount (introduced with Carbon) applies to the whole family — for agentic applications that repeat long system contexts, it's a major cost-saving lever, often more effective than switching models.

And if your budget is zero, the good news is that the best free LLMs now include models from this range with limited access — more than enough for personal use.


❌ Common Mistakes

Mistake 1: Migrating to Neon "because it's the latest"

The newest release isn't automatically the best for your use case. Neon is optimized for high-frequency frontier reasoning — if your use case is document summarization or customer support, Hydrogen will do the job at half the price. Solution: define your metric (cost per successful task, p95 latency) and test the new model on your real use cases before any migration.

Mistake 2: Building on introductory pricing

The $2/$10 rates of Argon and Carbon are teaser prices. Building a business model on them is building on sand. Solution: model your costs on the target prices ($4/$20) and treat the intro prices as a temporary margin.

Mistake 3: Ignoring Argon's replacement in the Gemini app

If your team uses the Gemini app with Argon, it switched to Neon without necessarily knowing it. Results may differ — sometimes for the better, sometimes at a latency cost. Solution: audit which models your team actually uses, and document the behavioral changes observed since November 18.

Mistake 4: Confusing release cadence with real progress

Six major models in two months doesn't mean six times the capability. Part of this cadence is marketing segmentation of capabilities that already existed in the research models. Solution: follow independent benchmarks like vals.ai rather than announcements, and check out our article on Carbon reclaiming first place in the Intelligence Index to distinguish real leaps from adjustments.


❓ Frequently Asked Questions

Does Gemini 4 Neon replace Carbon?

No. Carbon remains the maximum-capability model in the lineup, oriented toward long-horizon coding and research. Neon sits below it, offering frontier reasoning at a controlled cost ($3/$15) for high-frequency workflows. Neon replaces Argon in the Gemini app, not Carbon in the hierarchy.

What is the exact price of Gemini 4 Neon?

$3 per million input tokens and $15 per million output tokens, according to the pricing published at its launch on November 18, 2026. Unlike Argon and Carbon, Neon does not appear to have a temporary introductory price — its pricing is presented as structural. Check the up-to-date rates on Google's official page.

What are Neon's 4 thinking levels?

Four adjustable reasoning-depth tiers that can be set at call time. The low level prioritizes latency and cost, the high level deep reasoning. This makes it possible to use a single model for varied use cases, adjusting the quality/cost trade-off per request rather than switching models.

Is Neon available via API?

Yes, with its $3/$15 pricing and its 1M token context. However, Argon's API status is not specified by Google at the time of Neon's launch, which creates uncertainty for developers who built on Argon. Watch for official deprecation announcements.

Should you abandon Claude or ChatGPT for Neon?

Not automatically. Neon is an excellent frontier reasoning model at an aggressive price, but the choice depends on your use case: our Gemini vs ChatGPT vs Claude comparison details the strengths of each ecosystem. Test Neon on your real use cases before migrating, especially if your prompts are calibrated for another model.

Why is Google releasing so many models so fast?

A strategy openly embraced by Sundar Pichai since July 2026: a near-monthly cadence. The goal is threefold: occupy every pricing tier before the competition, segment use cases (fast, code, frontier, max capability), and saturate media attention. It's a war of cadence as much as of capability.


✅ Conclusion

Gemini 4 Neon confirms that the real revolution of late 2026 isn't any single model, but the speed at which Google now knows how to ship them: four models in seven weeks, each with a precise price point, and a lineup that finally covers the entire spectrum from $1.50 to $20 per million tokens. The advice worth its weight in gold: try Neon if you have high-volume reasoning workflows, otherwise let the dust settle — and follow our monthly comparison of the best LLMs to know when a switch is truly worth the migration.