Gemini 4 Neon is Google’s new model for complex software engineering, knowledge work and scientific workflows. The announcement was published on November 18, 2026. Neon is meant to sit between the existing siblings in the Gemini 4 family: Carbon for the hardest, long-horizon projects and Hydrogen for fast everyday work. Unlike the larger models, Google is emphasizing speed and cost for Neon: 3 dollars per million input tokens, 15 dollars per million output tokens, with a context window of one million tokens.
Gemini 4 Neon: where it sits in the Gemini 4 family
According to Google’s announcement, Neon is “adapted from Gemini 4 Carbon.” The Google DeepMind model card confirms that. The model is multimodal and handles text, images, audio and video in a shared context of up to one million tokens. Output is up to 64,000 tokens by default, or up to one million in background long decode. The knowledge cutoff is March 2026, and in some domains the model’s knowledge may be less current.
Neon is rolling out now in the Gemini API in Google AI Studio, in Gemini for business, and for paying Google AI Pro and Ultra subscribers in Google Antigravity and the Gemini app. Google also mentions Android Studio for developers. Exact availability depends on the channel and subscription tier.
Compared with the other new models of recent days, Neon is a typical mid-tier offering: faster and cheaper than the flagships, but intended to keep up on typical enterprise and coding work. Our reports on Gemini 4 Hydrogen and on GPT-6.2 Astra and Sol describe the direct competitors.
The benchmarks according to Google
Google cites a long list of test results. Many come from third parties, others from internal Google tests. Independent verification was not available at the time of the announcement. Some of the main figures:
- Artificial Analysis Intelligence Index: a score of 61 at the highest thinking level, “xhigh,” ranking third overall, behind larger and slower models.
- Vals Index: second place at 70.9 percent, and according to Google first place on the Legal Research Bench and Tax Agent Bench.
- Agent Arena Index: second place globally with a score of 1,336 across more than 4,200 real-world agent sessions, and also second on steerability (how cleanly the model adapts to mid-task corrections).
- Coding: first place on the Artificial Analysis Coding Agent Index (58.6), 65.2 percent on Terminal-Bench 4.1 and 46.2 percent on FrontierCode 2.0 at the default “high” setting.
- Science and math: 70.5 percent on Terminal-Bench Science, 84.0 percent on RiemannBench and 71.3 percent on HLE-Diamond.
- Other scores: 75.8 percent on Agents’ Last Exam v1.1, 73.5 percent on Chartography, first place on Zapier’s AutomationBench at 56.3 percent, and 16,843 dollars on Andon Labs’ Vending-Bench 2.
On SWE-sweep, Google reports 12.8 percent on the mini-SWE harness and 39.0 percent in Google Antigravity. The large jump therefore depends on the harness, not on the model alone. At the default “high” thinking level, Neon is said to cost 32 percent less per task than Carbon on the Artificial Analysis Intelligence Index. At the highest “xhigh” level, Google speaks of more than 36 percent lower cost per task than competing models at their highest setting. On the Vals Index and Agent Arena, costs at the “high” level are, according to Google, roughly 50 to 80 percent lower than those of leading competing frontier models.
Four thinking levels and a dedicated cyber mode
Neon has four thinking levels that let teams trade off effort and token use. The highest level, “xhigh,” is intended for the hardest engineering, analytical and scientific work. Google also stresses that Neon can act as a fast orchestrator directing parallel Hydrogen subagents in Google Antigravity.
Google is also announcing “Gemini 4 Neon for cyber.” This variant is meant to produce more secure code and perform better on typical defensive tasks. Figures cited include 76 percent on the hard tier of AutoBaxBench, 51.26 percent on WebExploitBench v1.2, 90.5 percent on SEC-bench Pro and 70.7 percent on Collinear’s CWE-bench v2. Access is scheduled to roll out in the coming days through the Fairwind Program for trusted defenders.
A full safety and security report for Neon is promised only in the coming days. The model card carries forward the critical capability assessments from Gemini 4 Argon: a critical capability threshold cannot be ruled out for biological risk and cybersecurity, and is not reached for chemical, radiological and nuclear scenarios or harmful manipulation. For loss of control, the card states that an early-warning threshold for stealth and situational awareness cannot be ruled out either.
Safety, alignment and known limits
According to Google, Neon maintains Carbon-level protections against severe misuse and prompt injection. On Gray Swan’s IPI benchmark, the attack success rate with real-time safeguards at the highest thinking level was 0.0 percent, matching other Gemini models. Internal figures Google itself cites include 0.0 percent cheating on Terminal-Bench 4.0, shortcuts on 2.45 percent of tasks in an internal coding-cheating evaluation, a 0.09 percent misaligned action rate and a 99.2 percent pass rate on an agent honesty evaluation. Despite the safeguards against cyber and biosecurity misuse, Neon is said to decline far less legitimate work in these areas; Google cites refusals on 1.98 percent of tasks on Terminal-Bench 4.0 and 0 percent on Terminal-Bench Science.
The model card also compares Neon with Gemini 4 Hydrogen: comparable text-to-text and multilingual safety, plus 1.90 percentage points on image-to-text safety, plus 1.25 percentage points on refusal tone, and minus 3.00 percentage points on unjustified refusals. As a known limitation, Google names hallucinations. The company also says it continues to improve jailbreak resistance.
In the Gemini app, Neon is intended mainly for longer tasks, for reading spreadsheets, PDFs and photos, and for drawing on details from Gmail and Docs. We described Google’s recent changes to how long chats are stored in our article on Gemini chat history.
Assessment
Gemini 4 Neon is not a new flagship, but a model Google is deliberately placing between Carbon and Hydrogen. The pricing of 3 and 15 dollars per million tokens and the benchmark claims target the same market that OpenAI is addressing with GPT-6.2 Sol and Microsoft with MAI Code 2: a lot of everyday work, as cheaply and quickly as possible. Whether the many first-place finishes in tests selected by Google will hold up in practice remains to be seen. The full safety report is also still outstanding.


