Innovation & Tech

Google DeepMind Ships Gemini 4 Neon at $15 Per Million Output Tokens

Google DeepMind launched Gemini 4 Neon on November 18 at $3 per million input tokens and $15 per million output tokens, 25% under Carbon, and near the top of independent agent indexes.

By Daniel Okafor

5 min read

Google DeepMind Releases Gemini 4 Neon, Fourth Model in Rapid Succession
Google DeepMind Releases Gemini 4 Neon, Fourth Model in Rapid SuccessionPexels / pixabay

What's News

  • Gemini 4 Neon launched November 18 at $3 per million input tokens and $15 per million output tokens, 25% below Carbon's standard $4/$20.
  • Google says Neon matches or exceeds Carbon on quality at 3x speed and 60% of cost at its default high thinking level.
  • Artificial Analysis scores Neon (xhigh) 61 on its Intelligence Index, one point ahead of Carbon and third overall.
  • Arena.ai's Agent Arena Index, launched the same day, ranks Neon second with 1336 from more than 4,200 real-world agent sessions.
  • Neon's model card carries forward Argon's findings that CBRN Biological and Cybersecurity capability levels "cannot be ruled out."

By Daniel Okafor

Google DeepMind launched Gemini 4 Neon on November 18, priced at $3 per million input tokens and $15 per million output tokens — 25% below the standard price of the Gemini 4 Carbon model. The release makes Neon the fourth Gemini 4 model in a span that coverage describes as seven weeks or roughly two and a half months.

Those numbers carry the commercial pitch. Google says Neon matches or exceeds Carbon on overall quality at three times the speed and 60% of the cost, at its default high thinking level. Artificial Analysis, an independent evaluator, puts the saving at roughly 17% per task at non-discounted pricing. Third-party analysis reports about 32% lower cost per task at the default high thinking level.

Where does Neon fit in the Gemini 4 lineup?

Google DeepMind named the Gemini 4 family after chemical elements: Argon, Carbon, Hydrogen and Neon. Carbon handles the hardest, longest-horizon projects. Google describes Hydrogen as built for rapid execution across diverse agentic workflows, positioned as a daily driver for everyday agentic tasks. Neon sits between them, and Google says it adapted Neon from Carbon.

Neon carries a 1 million-token context window. It accepts text, image, audio and video input and returns text output. Four thinking levels — low, medium, high (default) and xhigh — let buyers trade latency for reasoning depth without changing models. Google aimed Neon at complex software engineering, knowledge work and scientific workflows that require sustained, multi-step reasoning.

What do Google's own benchmark claims show?

Google's blog posts a long list of first- and second-place finishes:

  • Vals Index: 70.9%, second place
  • Zapier's AutomationBench: 56.3%, first
  • Andon Labs' Vending-Bench 2: $16,843, first
  • Terminal-Bench Science: 70.5%, first
  • RiemannBench: 84.0%, first
  • HLE-Diamond: 71.3%, first
  • Agent Arena Index: 1336, second behind Claude Fable 5.2 Max

Google also claims the lowest fabrication rate in internal "Antigravity" testing, plus strong steerability, with a second place on Agent Arena.

What do independent evaluators say?

Artificial Analysis scores Neon (xhigh) at 61 on its Intelligence Index, one point ahead of Carbon (high) and third overall, behind GPT-6.2 Astra (max) and Claude Fable 5.2 (max). On the Coding Agent Index run with Antigravity CLI, Neon scores 59 — three points ahead of Carbon, level with Claude Opus 5.5 and Claude Fable 5.2 — at a lower cost per task.

Arena.ai launched its Agent Arena Index the same day. It places Neon second overall with 1336, drawn from more than 4,200 real-world agent sessions.

Those scores come with caveats. Many of Google's cited wins come from its own Antigravity environment or from benchmarks introduced recently, such as the Agent Arena Index, Vals Index and AutomationBench. The "lowest fabrication rate" claim rests on internal testing. And the field is crowded: labs have a history of tuning models to excel on specific leaderboards.

How much does Neon undercut Carbon?

Neon costs $3 per million input tokens, with cached input at $0.15 — a 95% discount — and $15 per million output tokens. Carbon normally costs $4 and $20, though Google has discounted Carbon to $2 and $10 until November 30. Artificial Analysis attributes Neon's edge to lower token prices and higher output speed, roughly 110 tokens per second against about 57 for Carbon.

Where can customers get Neon?

Customers can reach Neon now through the Gemini API in Google AI Studio, Gemini for Business, Google Antigravity for paid Pro and Ultra subscribers, and the Gemini app for Pro and Ultra plans. A cybersecurity variant, "Gemini 4 Neon for cyber," reaches trusted defenders through the Fairwind Program in the coming days.

Cryptobriefing tracked the model before the announcement. It found Neon in web-app code on November 16, in Google Cloud SKU listings on November 18, and reportedly live in AI Studio ahead of Google's official blog post, which carries the same November 18 date.

What does the model card say about risk?

Neon's model card records a Material Capability Change assessment against Argon. General capability rose, but stayed below the threshold that would force a full critical-capability re-evaluation, so Argon's determinations carry forward. That includes CBRN Biological and Cybersecurity critical capability levels, which "cannot be ruled out," and Loss of Control Stealth and Situational Awareness threshold levels, which also "cannot be ruled out." Google says mitigations continue to deploy.

On refusals and misuse, Google claims Neon maintains Carbon-level protection against severe misuse and prompt injection while refusing less legitimate work: 1.98% refusal on Terminal-Bench 4.0 and 0% on Terminal-Bench-Science. It reports 0.0% cheating on Terminal-Bench 4.0, a 2.45% shortcut rate on a coding cheating eval — the lowest of any Gemini 4 model — a 0.09% misaligned action rate on its agentic safety suite and a 99.2% pass rate on agent honesty. On Gray Swan's indirect prompt injection benchmark, Neon with real-time safeguards at xhigh resisted every attack, at 0.0% success rate. Google says the full Safety & Security Report will come soon.

What remains unproven

  • Independent replication of the Antigravity-specific results, including SWE-sweep at 39.0% against 12.8% on mini-SWE
  • Whether Neon matches Carbon at three times the speed and 60% of the cost outside Google's own evaluation harnesses
  • How the model performs on sustained, production workloads
  • Whether the carried-forward biological and cybersecurity capability determinations get revisited as deployment widens
  • Timing for the full Safety & Security Report and for broader Fairwind Program access

Independent replication will decide whether Neon's quality-per-dollar edge holds across production workloads, and the pending safety report will show whether carried-forward risk determinations survive a wider rollout.

Source: epyhia.com

Share this article:

More from Daniel Okafor

Daniel Okafor

Show full bio

Correspondent covering money, funding and market shifts around entrepreneurship, startups and venture funding at Business Bearings.

2,360 articles

Related articles

« Previous articleNext article »