Google DeepMind releases Gemini 4 Neon, its fourth model in seven weeks
Original title: Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks ago
Neon has a one-million-token window, trails GPT-6.2 Astra and Claude Fable 5.2, and undercuts Carbon on list price.
Google DeepMind released Gemini 4 Neon and positions it between Carbon and Hydrogen, with a one-million-token context window for text, images, audio and video. Artificial Analysis ranks Neon third on its index: at the xhigh reasoning level it scores 61, just above Carbon's 60, behind OpenAI's GPT-6.2 Astra and Anthropic's Claude Fable 5.2. Google says a Fairwind Program cyber variant leads SEC-bench Pro and WebExploitBench v1.2, and the brief model card claims Neon blocks every indirect prompt-injection case on the Gray-Swan benchmark.
Unconfirmed: Benchmark rankings and the claim of blocking 100 percent of Gray-Swan indirect prompt injections come from Google or third-party evals, not independent confirmation.
- One million input tokens cost $3, or $0.15 cached; one million output tokens cost $15.
- That list price is 25 percent below Carbon's regular rate but above Carbon's November promo price.
- At high, Neon matches Carbon's intelligence score while cutting per-task cost in half.
- Neon generates about 110 tokens per second versus Carbon's 57, cutting a benchmark from about 36 minutes to 21.
- Inside Google Antigravity, Neon at xhigh scores 59 on the Coding Agent Index.
- Vendors
- Organizations
- Google DeepMindArtificial AnalysisOpenAIAnthropic
From the original article
3 min read at the-decoder.com
Google DeepMind has released Gemini 4 Neon, its fourth model in seven weeks, which is nearly 2x faster than its sibling Carbon, outperforms it on most benchmarks and comes in at a significantly lower price. The article Google's latest AI model is faster, cheaper and smarter than its flagship predecessor from just weeks ago appeared first on The Decoder.
By Maximilian Schreiner · the-decoder.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.