cd /news/artificial-intelligence/google-deepmind-releases-gemini-4-ne… · home › topics › artificial-intelligence › article
[ARTICLE · art-172228] src=epyhia.com ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

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

Google DeepMind launched Gemini 4 Neon on November 18, the fourth model in its Gemini 4 series, positioned between the high-end Carbon and the cheaper, faster Hydrogen for complex software engineering, knowledge work and multi-step scientific workflows. Neon offers a 1 million-token context window, text, image, audio and video input with text output, and four thinking levels (low, medium, high as default, and xhigh), and Google claims first-place finishes on Zapier's AutomationBench (56.3%), Andon Labs' Vending-Bench 2 ($16,843), Terminal-Bench Science (70.5%), RiemannBench (84.0%) and HLE-Diamond (71.3%), plus second place on the Agent Arena Index (1336, behind Claude Fable 5.2 Max). Independent evaluator Artificial Analysis put Neon (xhigh) at 61 on its Intelligence Index, one point ahead of Carbon (high) and third behind GPT-6.2 Astra and Claude Fable 5.2, and at 59 on its Coding Agent Index with Antigravity CLI, three points ahead of Carbon at a lower cost per task.

key takeaway

Gemini 4 Neon offers a 1M-token context, text/image/audio/video input and four thinking levels, scoring 61 on Artificial Analysis's Intelligence Index (one point above Carbon) and 59 on its Coding Agent Index at a lower cost per task than Carbon.

read5 min views11 publishedNov 19, 2026

Google DeepMind launched Gemini 4 Neon on November 18, the fourth model in its Gemini 4 series in a span variously described as seven weeks [4] or roughly two and a half months [4]. The company positions Neon between the high-end Carbon and the faster, cheaper Hydrogen, targeting complex software engineering, knowledge work, and scientific workflows that require sustained, multi-step reasoning [1].

A periodic-table lineup #

The Gemini 4 family now includes Argon, Carbon, Hydrogen, and Neon — all named after chemical elements. Carbon handles the hardest, long-horizon projects; Google describes Hydrogen as built for rapid execution across diverse agentic workflows, positioning it as a daily driver for fast, everyday agentic tasks [2]; Neon is pitched as the balance: frontier-level reasoning for high-frequency production workloads, faster and more cost-efficient than Carbon [1].

Google says Neon is adapted from Carbon [6]. The model offers a 1 million-token context window, supports text, image, audio, and video input with text output, and provides four thinking levels: low, medium, high (default), and xhigh [1][3].

Benchmark claims — and the context they need #

Google's blog cites a string of first- or second-place finishes on benchmarks including the Vals Index (70.9% [6], #2 [1]), Zapier's AutomationBench (56.3%, #1), Andon Labs' Vending-Bench 2 ($16,843, #1), Terminal-Bench Science (70.5%, #1), RiemannBench (84.0%, #1), HLE-Diamond (71.3%, #1), and the Agent Arena Index (1336, #2 behind Claude Fable 5.2 Max) [1][4]. It also claims the lowest fabrication rate in internal "Antigravity" testing and strong steerability (#2 on Agent Arena) [1].

Independent evaluator Artificial Analysis confirms Neon (xhigh) scores 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) [3]. On the Coding Agent Index with Antigravity CLI, Neon scores 59, three points ahead of Carbon and in line with Claude Opus 5.5 and Claude Fable 5.2, at a lower cost per task [3]. Arena.ai's new Agent Arena Index, launched the same day, places Neon second overall with 1336 from over 4,200 real-world agent sessions [4][5].

Several caveats apply. Many of Google's cited wins come from its own Antigravity environment or recently introduced benchmarks such as the Agent Arena Index, Vals Index, and AutomationBench [1]. The "lowest fabrication rate" claim rests on internal testing [1]. And the benchmark landscape is crowded: labs have a history of tuning models to excel on specific leaderboards [4].

Pricing and the Carbon comparison #

Neon is priced at $3 per million input tokens ($0.15 cached, a 95% discount) and $15 per million output tokens [1][3][5]. That is 25% below Carbon's standard $4/$20 pricing, though Carbon has been discounted to $2/$10 until November 30 [3]. At its default high thinking level, Google says Neon matches or exceeds Carbon on overall quality at 3x the speed and 60% of the cost [1]. Artificial Analysis estimates Neon (xhigh) costs ~17% less per task than Carbon at non-discounted pricing, driven by lower token prices and higher output speed (~110 vs ~57 tokens/second) [3]. Third-party analysis reports approximately 32% lower cost per task at the default high thinking level [5].

Availability and the "quiet rollout" pattern #

Neon is rolling out immediately in the Gemini API (Google AI Studio), Gemini for Business, Google Antigravity (for paid Pro and Ultra subscribers), and the Gemini app (Pro and Ultra) [1]. A cybersecurity variant, "Gemini 4 Neon for cyber," is reaching trusted defenders via the Fairwind Program in the coming days [1].

Cryptobriefing noted that Neon appeared in web-app code on November 16, in Google Cloud SKU listings on November 18, and was reportedly live in AI Studio before Google's official blog post — a pattern that Cryptobriefing notes Neon followed closely [5]. The company's blog post is dated November 18 [1].

Safety and frontier-risk assessment #

The model card states Neon underwent a Material Capability Change assessment relative to Argon. General capability increased but fell below the threshold requiring a full critical-capability re-evaluation, so Argon's Threshold/Critical Capability Level determinations are carried forward [6][7]. Those include: CBRN Biological CCL and Cybersecurity CCL "cannot be ruled out"; Loss of Control Stealth and Situational Awareness TCL "cannot be ruled out." Mitigations continue to be deployed [6][7].

Google claims Neon maintains Carbon-level protections against severe misuse and prompt injection while refusing far less legitimate work (1.98% refusal on Terminal-Bench 4.0, 0% on Terminal-Bench-Science). It reports 0.0% cheating on Terminal-Bench 4.0, 2.45% shortcut rate on a coding cheating eval (lowest of any Gemini 4 model), 0.09% misaligned action rate on its agentic safety suite, and 99.2% pass rate on agent honesty [1]. On Gray Swan's indirect prompt injection benchmark, Neon with real-time safeguards at xhigh resisted every attack (0.0% success rate) [1]. The full Safety & Security Report is said to be coming soon [6].

What remains unknown #

  • Independent replication of the Antigravity-specific results (SWE-sweep 39.0% vs 12.8% on mini-SWE) and the claim that Neon matches Carbon at three times the speed and 60% of the cost.
  • How Neon performs on sustained, production workloads outside Google's evaluation harnesses.
  • Whether the carried-forward CCL determinations for biological and cybersecurity uplift will be revisited as Neon sees wider deployment.
  • The timeline for the full Safety & Security Report and for broader Fairwind Program access.

Google DeepMind has now shipped four Gemini 4 models in rapid succession. Neon's benchmark scores are competitive, its pricing undercuts Carbon, and independent indexes place it near the top of the current agent leaderboard. Whether the quality-per-dollar advantage holds across diverse production workloads — and how the frontier-risk profile evolves with deployment — will be the test in the weeks ahead.

Sources #

1. [Introducing Gemini 4 Neon](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon) (blog.google)
2. [Introducing Gemini 4 Hydrogen for Agentic Workflows](https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-hydrogen) (blog.google)
3. [Google continues to accelerate its release cadence with Gemini 4 Neon](https://artificialanalysis.ai/articles/gemini-4-neon-intelligence-gain) (artificialanalysis.ai)
4. [Gemini 4 Neon takes second place on Agent Arena Index, trailing only ...](https://cryptobriefing.com/gemini-4-neon-agent-arena-index-second) (cryptobriefing.com)
5. [Google DeepMind releases Gemini 4 Neon, a cheaper alternative to Carbon](https://cryptobriefing.com/google-deepmind-gemini-4-neon-launch) (cryptobriefing.com)
6. [Gemini 4 Neon - Model Card — Google DeepMind](https://deepmind.google/models/model-cards/gemini-4-neon) (deepmind.google)
7. [Gemini 4 Neon Model Card](https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Neon-Model-Card.pdf) (storage.googleapis.com)
── more in #artificial-intelligence 4 stories · sorted by recency
── more on @google deepmind 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/google-deepmind-rele…] indexed:0 read:5min 2026-11-19 · —