
Google has added a third model to its Gemini 4 family, and it sits in a deliberately practical spot. Gemini 4 Neon, announced on November 18, 2026, is pitched as the model for complex, high-frequency work: long coding sessions, spreadsheet-heavy analysis and technical research. It is not the biggest model in the lineup, and that is the point.
What Google announced
In its launch post, Google describes Neon as bringing frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work and science workflows. The company says Neon ranks third overall, with a score of 61, on the Artificial Analysis Intelligence Index, and first on that firm’s Coding Agent Index.
The model is rolling out in the Gemini API in Google AI Studio, with a 1 million-token context window. Google lists pricing at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million, a 95 percent discount. It is also appearing in the Gemini app and in Gemini for business.
Built for long-running agent work
Much of the announcement focuses on agentic coding. Google says Neon is engineered to hold performance through complex, tool-heavy tasks, whether it runs on its own or acts as an orchestrator that directs parallel Gemini 4 Hydrogen subagents inside Google Antigravity. According to the post, it cuts redundant file reads, reports progress as it works and handles context compaction reliably in long sessions.
Google backs this with demonstration projects it says were built in Antigravity. One is a first-person WebGL2 ski simulator on real Mont Blanc terrain, with more than 10,000 lines of custom shaders. Another turned a single image into a 100,000-cell Google Sheet using over 500,000 live formulas and no Apps Script. These are vendor showcases rather than independent tests, but they show the kind of long, multi-step builds the model is meant to sustain.
Enterprise inputs and thinking levels
For business users, the pitch is multimodal synthesis. Enterprise work often lives in spreadsheets, PDFs, charts, audio and video, and Google says Neon can combine these inputs in a single context window and turn them into structured deliverables. It cites independent results such as the Vals Index and Surge AI comparisons, and says Neon is best in class on Terminal-Bench Science for scientific and mathematical reasoning.
Neon also offers four thinking levels so teams can trade intelligence against token spend. At the default high level, Google says it posts a high Intelligence Index score at a cost per task 32 percent lower than Gemini 4 Carbon. It also cites roughly 50 to 80 percent lower cost per task than top competing frontier models on the Vals Index and Agent Arena, and says the top xhigh setting still costs at least 36 percent less per task than rivals at their highest settings. Those figures come from third-party benchmarks as reported by Google, so results on your own workloads will vary.
Where Neon fits in the family
Google positions the three models as tiers. Carbon takes the hardest, long-horizon projects. Hydrogen handles fast everyday execution. Neon balances depth and efficiency for complex workflows that run often. In practice that gives teams a way to route work: a heavy planner on one tier, many cheaper workers on another, and a middle model coordinating.
What it means
The interesting shift is economic rather than raw capability. As agents run longer and call more tools, cost per completed task matters more than cost per token. A model tuned to finish tasks with fewer wasted steps, plus a steep cached-input discount, speaks directly to that. Developers building agent pipelines should compare Neon on their own tasks, watching total cost per finished job rather than headline benchmark rank.
It also reinforces a pattern across the industry: model families are becoming routing menus. The practical skill is choosing the right tier for each step, not picking one model for everything.
Bottom line
Gemini 4 Neon is Google’s bid for the sweet spot between capability and cost in agentic coding and enterprise analysis. The claims are strong and mostly rest on third-party benchmarks reported by Google, so it is worth testing before committing. Availability starts today in AI Studio and the Gemini API.