Realtime AI News
Google DeepMind ships Gemini 4 Neon, pitching frontier performance at a lower cost per task
Google DeepMind released Gemini 4 Neon on Nov. 18, positioning it as a frontier model for fast end-to-end work in software engineering, knowledge work and science. At $3 per million input tokens and $15 per million output tokens with a 1 million-token context window, Google says it costs 32% less per task than Carbon.

Google DeepMind released Gemini 4 Neon on Nov. 18, positioning it as a frontier model built for fast end-to-end task completion across complex software engineering, knowledge work and scientific workflows. Crypto Briefing framed the launch simply: a cheaper alternative to Carbon.
The pricing is the headline. Neon costs $3 per million input tokens, $0.15 per million cached input tokens and $15 per million output tokens, with a 1 million-token context window — a combination clearly aimed at teams that run agents continuously rather than in demos.
Google's benchmark table is broad. Neon scores 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, third overall and behind only larger, slower models; it ranks second on the Vals Index and lands on the Pareto frontier; and it debuts second on Arena's new Agent Arena Index with 1,336 points across more than 4,200 real-world agent sessions.
On task-specific work, Google says Neon ranks first on Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843), a long-horizon business simulation, as well as on the Legal Research Bench and Tax Agent Bench. It scores 73.5% on Chartography for reading dense professional charts.
Coding gets the most detail. Neon tops the Artificial Analysis Coding Agent Index at 58.6 and reaches 65.2% on Terminal-Bench 4.1; Google notes it manages 12.8% on the mini-SWE harness but jumps to a state-of-the-art 39.0% inside Google Antigravity, an argument that harness design now matters as much as the model.
The efficiency claim is the competitive one. At its default high thinking level, Neon delivers a 32% lower cost per task than Carbon, and Google says it sits on the Pareto frontier of the Vals and Agent Arena indices at roughly 50% to 80% below top frontier competitors on cost per task. The xhigh setting adds deeper multi-step reasoning while, Google claims, still costing at least 36% less per task than rivals at their highest setting.
Google is also shipping a cyber-tuned variant. Gemini 4 Neon for cyber is said to beat Carbon on secure-code generation across easy, medium and hard AutoBaxBench challenges, reach a state-of-the-art 76% on the hard tier, and lead WebExploitBench v1.2 at roughly half the cost of larger models.
Neon is available from today through the Gemini API and Google AI Studio, in Gemini for business, in Google Antigravity for paid AI Pro and Ultra subscribers, and in the Gemini app. The number that matters next is not the leaderboard rank but the independently measured cost per completed task in production.
Why it matters
The pricing pushes frontier-class capability into the range where teams can run agents continuously rather than in demos, squeezing rivals on cost in coding and office automation. The next test is measured cost per completed task on real production workloads, not benchmark position.
Nearby Updates
All11/19, 05:57
Nvidia's delayed 10-Q turns a routine filing into a Wall Street mystery
Nvidia told investors on Tuesday's earnings call that it would not immediately file its 10-Q with the SEC, saying only that it would arrive in the coming weeks and offering no reason. A UBS analyst later learned the delay relates to the company's investigation into Huang and whether new risk factors must be added.
11/19, 05:34
Palmer Luckey warns AI-powered cyberattacks are outpacing defenses
Speaking at The New York Times' DealBook Summit on Nov. 18, Anduril co-founder Palmer Luckey warned that AI makes crafting exploits easy and that outdated systems cannot keep pace with the new attack speed. He said adversaries are exploiting the mismatch, singled out North Korea, and urged defenders to move faster.
11/19, 06:04
Anthropic prices IPO at $83 a share, topping its marketed range
Anthropic has set the price of its initial public offering at $83 per share, above the top of its previously marketed range. The Tech Buzz reports the pricing landed amid heavy investor appetite for AI assets, making it one of the most closely watched AI listings of the week.
11/19, 04:44
Hugging Face Transformers 5.23.0 ships seven model integrations and a rewritten exporter API
Hugging Face has released Transformers 5.23.0, adding seven new model integrations and rewriting the library's exporter API. The release also carries twelve breaking changes, so teams that depend on it for training or inference should plan a migration and regression pass before upgrading.