Google Releases Gemini 4 Neon: Twice as Fast as the Flagship, but Current Pricing Is Higher Than the Model It Replaces

Google has released Gemini 4 Neon, positioned as the efficiency-depth balance point in the Gemini 4 family, with roughly twice the output speed of the flag

On November 18, 2026, Koray Kavukcuoglu, Senior Vice President at Google DeepMind and Chief AI Architect, announced the launch of Gemini 4 Neon, positioning it as the "efficiency and depth balance point" of the Gemini 4 family—faster than the flagship Carbon and stronger than the lightweight Hydrogen. It became fully available on the day of launch, with the model ID gemini-4-neon and a knowledge cutoff date of March 2026.

But after reading the launch announcement, one number is worth pausing to consider: this model positioned as the "middle layer" is currently priced at $3 per million input tokens—while the flagship Carbon, which it is meant to balance, now costs $2 per million input tokens. This is not a typo but a pricing structure deliberately designed by Google.

A Story of Pricing Reversal

The Gemini 4 family currently includes four models: Argon, Carbon, Neon, and Hydrogen. Carbon is the current flagship, with a codename pointing to the hardest element and a focus on the hardest long-horizon tasks; Hydrogen is the lightweight contender at $1.5/$7.5 per million tokens and the only Gemini 4 model offering a free tier; Neon sits in the middle at $3/$15 per million tokens.

The issue is that Carbon is enjoying a temporary promotional price: according to the Google AI Studio pricing page, the Standard pricing for Carbon and Argon is $2/$10, valid until November 30, and will rise to $4/$20 starting December 1. Under this schedule, starting in December, Neon ($3/$15) will be 25% cheaper than Carbon ($4/$20). Google is not releasing a "cheaper model"; rather, it is using a promotional window to create a brief price inversion, then letting Neon naturally become the "more affordable choice" after Carbon's price increases.

This structure itself is a message: Google's long-term confidence in Neon is higher than its commitment to Carbon's current pricing.

Speed Is the Real Weapon

From benchmarks, Neon scores 61 on the Artificial Analysis Intelligence Index (xhigh thinking level), ranking third globally and behind larger, slower models. It ranks second on the Vals Index (70.9%) and simultaneously takes first place on both the Legal Research Bench and the Tax Agent Bench.

But these numbers obscure a more substantive difference: speed. As measured by Artificial Analysis, Neon's output speed is about 110 tokens per second, while Carbon is 57 tokens/s, nearly a twofold gap. In more than 4,200 real-world agent sessions in the Agent Arena Index, Neon took an average of 21 minutes to complete an Intelligence Index task, while Carbon took 36 minutes; the cost per task was $6.44 and $7.76, respectively.

For agent workflows that chain dozens of model calls, this speed gap is amplified linearly. A codebase-level debugging task requiring 100 rounds of tool calls might require waiting 60 hours on Carbon and is compressed to just over 30 hours on Neon. This is not minor tuning optimization but a threshold difference that determines whether a workflow is feasible.

The Underrated Steerability Metric

One data point in the announcement is easy to overlook amid extensive benchmark discussions: Neon ranks second globally on the "steerability" dimension of the newly launched Agent Arena Index—its ability to adapt to user corrections mid-task—with a score of 1,336, based on more than 4,200 real-world agent sessions.

Raw intelligence scores reflect a model's theoretical ceiling, while steerability reflects actual behavior in production environments. The pain point for enterprises deploying AI is often not whether the model is smart enough, but whether, when requirements change mid-task, the model will reinterpret instructions as expected or stubbornly complete what it "thinks" is the original task. The second-place steerability ranking deserves more attention than most tech media have given it.

In scientific and mathematical reasoning, Neon scores 70.5% on Terminal-Bench Science, reaches 84.0% on RiemannBench, and scores 71.3% on HLE-Diamond (hard reasoning), all three the highest in the industry.

Precise Targeting of Enterprise Scenarios

Looking at all the first-place results announced for Neon: Zapier AutomationBench (56.3%), Andon Labs Vending-Bench 2 ($16,843, simulating long-horizon business operations), Legal Research Bench, and Tax Agent Bench—these are not general AI capability contests but high-frequency, high-value enterprise vertical scenarios: legal research, tax agency, financial analysis, and automated workflows.

Google's official blog specifically notes that Neon ranks first in finance, accounting, spreadsheet modeling, and complex constraint following categories on the Surge AI platform, and scores 73.5% on reading professionally dense charts (Chartography). This combination of capabilities points to the same type of user: knowledge workers in professional services who deal daily with PDFs, spreadsheets, charts, and audio, and need models to integrate these cross-modal inputs into structured outputs.

Google describes Neon as "frontier-level reasoning for complex, high-frequency workflows," and the 1,000,000-token context window is the infrastructure underpinning this narrative—allowing it to digest an entire codebase, a complete audit report, or hours-long meeting recordings in a single request.

Head-to-Head Comparison with Anthropic

According to Artificial Analysis data, in a direct comparison between Gemini 4 Neon and Claude Opus 5.5, the former scores 61 on the Intelligence Index and the latter 59; on per-million-token pricing, Neon ($3/$15) is about 25% cheaper than Opus 5.5; on output speed, Neon is about 105-127 tokens/s and Opus 5.5 is about 42-118 tokens/s, with some tests showing Neon demonstrating nearly a 2.5x speed advantage in interactive development scenarios. On context window, Neon is 1,050,000 tokens and Opus 5.5 is 1,000,000 tokens.

This comparison itself involves some methodological differences; different evaluation platforms measure speed at anywhere from 105 to 127 tokens/s.

What Else Google Promised

Google's announcement also mentions that Neon recorded its "lowest hallucination rate to date" in internal Google Antigravity tests. This figure was not disclosed in an externally reproducible way and is provided only as a directional signal. The same announcement emphasizes that Neon performs excellently on multi-step instruction following—echoing the steerability ranking and pointing to a model characteristic that, in complex tasks, not only executes but also responds to changes.

In terms of access, Neon has no free tier: on the API side, it requires the paid Standard, Batch, Flex, or Priority tiers; on the consumer side, it requires a Google AI Pro or Ultra subscription. This differentiates it from Hydrogen's strategy—the latter is the only model in the Gemini 4 family that retains a free entry point.

Conclusion

Gemini 4 Neon is a model with a clear goal: it does not aim to be the most powerful, but to be a choice that is sufficient across speed, cost, and availability simultaneously. For developers and enterprises, the current price window before November 30 creates a strange phase—Neon is more expensive than Carbon, which will briefly suppress adoption willingness. But starting in December, its relative price advantage will emerge.