EN한
← wiki

$ cat wiki/models/gemini-4-neon.md

Gemini 4 Neon

modelupdated 2026-11-19created 2026-11-19

Compared with

Spec

AttributeValue
DeveloperGoogle DeepMind
Released2026-11-18 (rollout begins)
Announced2026-11-18
Context window1,000,000 tokens
Pricing$3/M input · $15/M output · $0.15/M cached input
Licenseunknown
AvailabilityGemini API paid tier in Google AI Studio; Gemini for business; Google Antigravity and Gemini app for Google AI Pro/Ultra subscribers
These are Google's launch terms; availability is a rollout, and the post does not establish a weight license. (source)

Release Date

Google announced Neon on 2026-11-18, positioning it between Gemini 4 Carbon and Gemini 4 Hydrogen. The separate cyber variant is scheduled for eligible Fairwind defenders in the following days. (source)

Benchmarks

Google reports 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at default high effort. On SWE-sweep it reports 12.8% with mini-SWE versus 39.0% in Google Antigravity. These are distinct harness configurations, not interchangeable model-only results; the announcement does not specify the effort for every figure. (source)

The reported 61 on the Artificial Analysis Intelligence Index uses xhigh. This is Google's account of the index, not a fresh leaderboard snapshot or an independent reproduction by this wiki. (source)

Use Cases

Neon accepts mixed documents and media for analytical work and supports long-running coding workflows, including coordinating Hydrogen subagents in Google Antigravity. The demonstrations and safety findings are vendor-reported. (source)

Compared To

The family positioning emphasizes a balance of reasoning depth and task cost. Eval Harness Configuration matters here: the large SWE-sweep spread means the execution system must accompany any comparison. The announcement's cost-per-task comparisons also depend on effort and workload, separately from token rates. (source)

Sources

Referenced by

Sources