Google has finished filling in the Gemini 4 lineup. After Carbon (deep, long-horizon work) and Hydrogen (fast, cheap execution), Gemini 4 Neon arrived on November 18, 2026 as the middle tier. Google pitches it as frontier-level reasoning for complex work that runs often, where Carbon is too slow or expensive to use on every call.
It’s rolling out in the Gemini API (paid tier, generateContent and the Interactions API), Google AI Studio, Android Studio, Gemini for business, Google Antigravity for paid AI Pro and Ultra subscribers, and the Gemini app for Pro and Ultra.
Price and limits
| Gemini 4 Neon | |
|---|---|
| Model ID | gemini-4-neon |
| Input / 1M tokens | $3.00 |
| Cached input / 1M tokens | $0.15 (implicit caching, prompts ≥ 4,096 tokens) |
| Output / 1M tokens (incl. thinking) | $15.00 |
| Context | 1,048,576 tokens, shared by input, thinking and output |
| Output per request | 65,536 synchronous; up to ~1M with background: true |
| Thinking levels | low, medium, high (default), xhigh |
| Free tier | No |
Pricing is flat across the whole window, with no long-context surcharge. For reference, our Hydrogen coverage lists Hydrogen at $1.50 / $7.50. Carbon is $2 / $10 at its introductory price, rising to $4 / $20 on December 1.
What Google claims
All numbers below are vendor-reported:
- Coding: #1 on the Artificial Analysis Coding Agent Index (58.6), 65.2% on Terminal-Bench 4.1, 46.2% on FrontierCode 2.0 at the default high setting. On SWE-sweep it scores 12.8% in the mini-SWE harness and 39.0% in Antigravity’s own harness. That gap is as much about the harness as the model.
- General: 61 on the Artificial Analysis Intelligence Index at xhigh (#3 overall), #2 on the Vals Index (70.9%), and #2 on Arena’s new Agent Arena Index (1336).
- Cost per task: at high, 32% lower than Carbon, and roughly 50–80% lower than top competing frontier models on Vals and Agent Arena, per Google.
- Agent behavior: fewer redundant file reads, reliable context compaction in long sessions, 0.0% cheating on Terminal-Bench 4.0 and a 2.45% shortcut rate on Google’s coding-cheating eval.
Google frames Neon as the orchestrator model: it can work alone or direct parallel Hydrogen subagents in Antigravity. The Antigravity agent in Gemini Managed Agents supports Neon on its own or paired with Hydrogen. A security-tuned Neon for cyber is going to vetted defenders through the Fairwind Program over the next few days.
Migration gotchas
If you’re moving from Gemini 3.x, or even from Carbon, read the migration notes before you swap the model string:
- Thinking defaults change. Neon defaults to
highand supportsxhigh. Carbon and Argon default tomediumand cap athigh.minimalreturns an error. - Output cap. One synchronous request returns at most 65,536 tokens, and
max_output_tokenscovers thinking plus answer. Remove old caps like 8,192. For longer output, setbackground: truewith a highermax_output_tokens. - No explicit caching.
cachedContents/caches.createaren’t supported, so removecached_content. Implicit caching applies automatically. - Sampling knobs are ignored. In-range
temperature,top_pandtop_kare ignored, and out-of-range values return a 400. - Declare tools natively. Prompting for raw-text JSON tool calls with no declared tools can return
MALFORMED_FUNCTION_CALL. - Raise client timeouts to at least 15 minutes, or use background execution. Hard prompts can take minutes.
Where it fits
For a planner/executor setup, Neon is the obvious planner or reviewer over Hydrogen workers when Carbon is overkill. At $3 / $15 it costs twice as much as Hydrogen per token, so it’s worth testing whether fewer steps and fewer redundant reads actually make it cheaper per task on your own workload, as Google claims. Run your evals at each thinking level and pick the lowest one that clears your quality bar. Google’s docs recommend exactly that.