Gemini 4 Neon launches, tops coding agent index

Google DeepMind released Gemini 4 Neon on November 18, the middle tier of the Gemini 4 family. By Google's division of labor, Carbon handles the hardest, longest-running projects and Hydrogen covers fast everyday execution, while Neon sits between them, aimed at complex coding, analytical reasoning and enterprise workflows.

Pricing is $3 per million input tokens and $15 per million output tokens, with cached input at $0.15, one-twentieth of the standard input price. The context window is 1 million tokens. Reasoning effort comes in four levels; Google says the default high level costs 32% less per task than Carbon, and the top xhigh level is more than 36% cheaper per task than a competitor's highest setting.

Where it sits on the scoreboard

Google cites quite a few third-party leaderboards. On Artificial Analysis, Neon scores 58.6 on the Coding Agent Index, ranking first, and 61 on the overall Intelligence Index at the xhigh level, ranking third. It reaches 70.9% on the Vals Index, ranking second, and 1336 on Arena's agent arena, based on more than 4,200 real agent sessions, also second. Zapier's AutomationBench (56.3%) and Andon Labs' Vending-Bench 2 ($16,843) both put it first.

On the engineering-oriented tests, Terminal-Bench 4.1 comes in at 65.2%, against 57% for Hydrogen, released last week, on the same test; FrontierCode 2.0 gives 46.2% at the default level. SWE-sweep scores depend heavily on the harness: just 12.8% with the mini-SWE harness, but 39.0% with Google's own Antigravity harness. Most of these figures come from third-party or Google-internal evaluations, and the launch page does not specify uniform test conditions, so be careful when comparing them side by side.

On safety, Google offers several behavioral figures: 2.45% of cases take shortcuts in its coding-cheating evaluation, the misbehavior rate in agent safety testing is 0.09%, and the attack success rate in Gray Swan's indirect prompt injection test is 0. A cybersecurity-focused version will be open only to trusted defenders and rolls out in the coming days through the Fairwind program; applicants need to contact their Google Cloud customer team.

Google writes in its announcement:

“Neon also excels at multi-step instruction following and builds on our ongoing progress to curb fabrication”

Illustration: a control hub directs robot stations in parallel, showing how Gemini 4 Neon coordinates sub-agents

Running the numbers across the three tiers

Put the public prices of the family side by side: Hydrogen costs $1.50 per million input tokens and $7.50 for output; Carbon is at a promotional $2 and $10, returning to $4 and $20 from December; Neon is $3 and $15.

On unit price alone, Neon costs half again as much as Carbon at its promotional rate. Google's “32% lower” is measured per task and relies on Neon using fewer tokens to finish the same job. A rough example: an agent task that reads 200,000 tokens and writes 30,000 would cost about $0.60 + $0.45 = $1.05 at Neon's list price; the same usage on Carbon after December comes to $0.80 + $0.60 = $1.40. How big the real gap is depends on how much output Neon saves on a given task, and Google has not published per-task token usage.

When a task carries a lot of repeated context, such as repeatedly reading the same code repository, the cached price matters more. If all 200,000 input tokens in the example above hit the cache, the input side costs just $0.03, and the cost of a run is driven mainly by output.

Where you can use it

Developers can call it on the paid tier of the Gemini API, and it is already integrated into Google AI Studio and Android Studio; enterprise customers get it through Gemini for business and Antigravity. In Antigravity, Neon can act as an orchestrator, directing multiple parallel Hydrogen sub-agents to split up the work.

In the Gemini app, Google AI Pro and Ultra subscribers can pick it from the model drop-down, and Gemini's tasks feature can also be switched to run on Neon. Google has not said whether free users can access it or what the usage caps in the app are.

Sources: Google's official blog post announcing Gemini 4 Neon, CocoLoop, Artificial Analysis, Vals; per-million-token input, output and cache prices were checked, while third-party leaderboard ranks and scores follow the figures as cited by Google.