Google DeepMind has released Gemini 4 Neon, a reasoning model positioned between Gemini 4 Carbon and Gemini 4 Hydrogen. The tool targets coding agents, knowledge work, and science workflows. Pricing stands at $3 per 1M input tokens and $15 per 1M output tokens.
In this article
TL;DR
- Size: Unified context window of 1M tokens, 64K default output, up to 1M via background long decode.
- Runs on: Google-hosted infrastructure only. Weights are closed; no local GPU, quantized, or edge builds are available.
- Performance: Scores 61 on the AA Intelligence Index, placing it ahead of Claude Opus 5.5 (59) and GPT-6.2 Sol (57).
- Best: Achieves 95.2% on GraphWalks (256K to 1M context), compared with 80.1% for GPT-6.2 Sol.
- Worst: Scores 46.2% on FrontierCode 2.0, trailing Claude Opus 5.5 at 47.3%.
- Bottom line: Leads in long-context and agentic coding at $3/$15. It is closed, uses self-reported metrics, and costs more per token than GPT-6.2 Sol.
What is Gemini 4 Neon?
Neon occupies the middle tier of the Gemini 4 family. Carbon handles the hardest long-horizon projects. Hydrogen handles fast everyday execution. Neon balances depth and efficiency between the two.
According to the model card, Neon is adapted from Gemini 4 Carbon. Google does not disclose parameter count or architecture. Inputs cover text, images, audio and video. The knowledge cutoff is March 2026.
How does Neon work inside agent workflows?
Neon exposes multiple thinking levels. High is the default. xhigh is reserved for the hardest engineering, analytical and science workloads.
For long agent runs, Neon supports context compaction and makes fewer redundant file reads. In Google Antigravity, Neon can act as an orchestrator. It directs parallel Gemini 4 Hydrogen subagents on smaller subtasks.
How does Gemini 4 Neon perform on benchmarks?
Google compares Neon against Claude Opus 5.5 and GPT-6.2 Sol. All scores below are reported by Google, not independently reproduced.
- Coding: #1 on the AA Coding Agent Index at 58.6. SWE-sweep with provider harness hits 39.0%, vs 10.1% for Opus 5.5.
- Knowledge work: 56.3% on Zapier AutomationBench, vs 42.5% for Opus 5.5. 75.8% on Agents’ Last Exam v1.1.
- Science and reasoning: 84.0% on RiemannBench, 70.5% on Terminal-Bench-Science and 71.3% on HLE-Diamond.
- Long context: 95.2% on GraphWalks, vs 66.8% for Opus 5.5.
- Cyber: 90.5% on SEC-bench Pro, 70.7% on CWE-Bench v2 and 51.3% on WebExploitBench v1.2.
Neon does not win every row. Opus 5.5 leads Terminal-bench 4.1 (65.8% vs 65.2%) and FrontierCode 2.0 (47.3% vs 46.2%). GPT-6.2 Sol leads Chartography (75.1% vs 73.5%).
On the Vals Index, Neon scores 70.9% and ranks #2. On the Agent Arena index it also ranks #2, with a score of 1336.
What does Gemini 4 Neon cost?
Neon lists at $3 input, $0.15 cached input and $15 output per 1M tokens. Google says Neon at high thinking costs 32% less per task than Carbon. It also claims 50% to 80% lower cost per task vs rivals on Vals Index and Agent Arena. At xhigh, Google cites 36%+ lower cost than rivals at their highest settings.
Those are per-task claims, not per-token prices. GPT-6.2 Sol lists at $2/$10. Claude Opus 5.5 lists at $4/$20. Gemini 4 Carbon is $2/$10 until November 30, then $4/$20.
Where can developers use Gemini 4 Neon?
- Gemini API paid tier, via Google AI Studio and Android Studio
- Gemini for business, via API and the model dropdown
- Google Antigravity, for paid Google AI Pro and Ultra subscribers
- The Gemini app, for Pro and Ultra subscribers
What do the safety numbers show?
Neon runs on the Gemini 4 safety and security stack. Google reports a 0.09% misaligned action rate on its agentic safety suite. It passes 99.2% of the agent honesty evaluation. It shows a 0.0% attack success rate on Gray Swan’s prompt injection benchmark at the highest thinking level. Its 2.45% shortcut rate on a coding cheating test is the lowest of any Gemini 4 model.
Gemini 4 Neon vs closest competitors
| Feature | Gemini 4 Neon | Gemini 4 Carbon | Claude Opus 5.5 | GPT-6.2 Sol |
|---|---|---|---|---|
| Developer | Google DeepMind | Google DeepMind | Anthropic | OpenAI |
| Parameters | Not disclosed | Not disclosed | Not disclosed | Not disclosed |
| Weights | Closed | Closed | Closed | Closed |
| Context window | 1M | Not disclosed in launch post | 1M | 1M |
| Max output | 64K default, 1M background | Not disclosed | 128K | Not disclosed |
| Input / output per 1M | $3 / $15 | $2 / $10 to Nov 30, then $4 / $20 | $4 / $20 | $2 / $10 |
| AA Intelligence Index | 61 | Not disclosed | 59 | 57 |
| Vals Index v2.2* | 70.9% | Not disclosed | 70.1% | 66.1% |
| Terminal-bench 4.1* | 65.2% | Not disclosed | 65.8% | 61.5% |
| SWE-sweep (provider harness)* | 39.0% | Not disclosed | 10.1% | 8.3% |
| GraphWalks 256K to 1M* | 95.2% | Not disclosed | 66.8% | 80.1% |
*Scores from Google’s launch comparison table, self-reported. AA scores for Opus 5.5 and GPT-6.2 Sol match Artificial Analysis.
What it means
Developers using coding agents will find Neon offers a specific trade-off. It provides superior performance on long-context recall tasks like GraphWalks and leads on the AA Coding Agent Index. However, the model remains closed and relies on self-reported benchmarks. Users seeking the highest scores on FrontierCode 2.0 or Terminal-bench 4.1 may still prefer Opus 5.5. The pricing is higher per token than GPT-6.2 Sol, though Google claims lower costs per completed task at high thinking levels.


