Google DeepMind announced Gemini 4 Neon on Nov 18, 2026. Google positions it as a model that finishes multi-step tasks quickly across software engineering, knowledge work and science.
Where it sits and where to get it
Neon is the middle model in the Gemini 4 family. Google says Carbon handles the hardest long-horizon projects and Hydrogen handles fast everyday execution, while Neon balances depth and efficiency for complex workflows that run often.
The rollout covers the Gemini API in Google AI Studio and Gemini for business, through the API or the model dropdown. Paid Google AI Pro and Ultra subscribers get it in Google Antigravity, and Pro and Ultra subscribers get it in the Gemini app. Tasks inside the Gemini app can now run on Neon.
Price and thinking levels
Input costs $3 per million tokens and output costs $15 per million. Cached input costs $0.15 per million, which Google describes as a 95% discount. The context window holds 1 million tokens.
Neon has four thinking levels, so a team can trade intelligence against token spend per workload. Google says the default high level costs 32% less per task than Carbon on the Artificial Analysis index. It also states that the cost per task is about 50% to 80% below that of leading rival frontier models on the Vals Index and Agent Arena. At xhigh, Google says cost per task is still at least 36% below competitors at their highest setting.
What Google reports on performance
All figures below come from Google's announcement and the third-party benchmarks it cites.
- Artificial Analysis Intelligence Index: 61 at xhigh, third overall.
- Vals Index: second, at 70.9%.
- Zapier's AutomationBench: first, at 56.3%.
- Legal Research Bench and Tax Agent Bench: first on both.
- Agent Arena Index: second, at 1336 across more than 4,200 real agent sessions.
- Steerability, meaning how well the model adapts to corrections mid-task: second on the same index.
Google also says Neon recorded its lowest fabrication rate yet in internal testing in Google Antigravity.
Security
On Gray Swan's indirect prompt injection benchmark, which plants malicious instructions in web pages, emails, documents or code, Google says Neon with real-time safeguards at its highest thinking level resisted every attack. A separate Neon variant for cyber defense goes to trusted defenders through the Fairwind Program in the coming days. Eligible members are asked to contact their Cloud account team.
Top comments (1)
The $0.15 cached-input line is the most consequential number in this announcement, and it is easy to skim past. At $3 input / $15 output with no cache reuse, a task's bill is dominated by output. Add a stable prefix — system prompt, tool schema, house style, retrieved documents that don't change between calls — and cached reads at 95% off flip which side owns the bill. Two tools both listed at "$3 input / $15 output" can differ three to five times per task purely on prefix_tokens / total_input_tokens, so that ratio deserves to sit next to any price comparison.
Second thing to pin down: four thinking levels means the price sheet needs a per-level row. "32% less cost per task than Carbon at the default high level" is a cost-per-task claim, not a per-token one, and at xhigh the same model can easily cost more per task than a cheaper tier at its default. If you're forecasting a budget rather than a demo, the units you need are cost per completed task at the level you will actually run, plus the retry rate at that level — the second number is usually what eats the first.
Small caveat on the 95%: it applies to input only. Long-generation tasks see a much smaller effective saving than the headline suggests, because output tokens are unaffected by caching.