News

Google Ships Gemini 4 Neon, a Cheaper Model Built to Finish the Task Rather Than Win Benchmarks

Google released Gemini 4 Neon on November 18 at $3 per million input tokens and $15 per million output, third overall on the Artificial Analysis Intelligence Index and second for steerability on the new Agent Arena Index.

Google's newest model is not its largest, and that is the point. Gemini 4 Neon, released on November 18, is built to finish tasks quickly at a price that lets agents run all day, and it lands third overall on the Artificial Analysis Intelligence Index with a score of 61 at its highest thinking level.

The launch post puts the claim narrowly: Neon ranks behind larger, slower models on raw intelligence and ahead of them on the cost of finishing a task. Google prices it at $3 per million input tokens, $0.15 per million for cached input, and $15 per million output, with a 1 million token context window. At its default high thinking level it costs 32 percent less per task than Carbon, the model Google opened to all paying developers on November 10.

What steerable means in practice

The result that says most about daily use is not the intelligence score. Neon placed second on the new Agent Arena Index for steerability, which measures how cleanly a model absorbs a user correcting it mid-task, across 4,200 recorded agent sessions. Anyone who has watched an agent keep building the wrong thing after being told to stop knows why that matters more than another point of reasoning.

Google also reports its lowest fabrication rate so far in internal Antigravity testing, a 0.09 percent misaligned action rate on its agentic safety suite and a 2.45 percent shortcut rate on its coding cheating evaluation. Those are vendor figures on vendor evaluations, and they describe behavior under test conditions rather than in someone's repository.

The cyber edition

Neon for cyber scores 76 percent on the hard tier of AutoBaxBench, which tests whether a model writes secure code, 51.26 percent on WebExploitBench v1.2 for finding exploitable flaws, 90.5 percent on SEC-bench Pro for locating vulnerabilities and 70.7 percent on CWE-bench v2 for patching them. Google says it reaches those numbers at roughly half the cost of larger frontier models. It goes to trusted defenders through the Fairwind program first, not to the general API.

The model ships with four thinking levels rather than one, so a team can pick how much reasoning a workload deserves. In Antigravity it can also act as an orchestrator, handing work to parallel Hydrogen subagents and compacting its own context when a long session fills up.

Where it is available

The launch covers the Gemini API, Google AI Studio, Gemini for business through the API and the model dropdown, paid Google AI Pro and Ultra subscribers in Antigravity, and Pro and Ultra subscribers in the Gemini app.

The interesting question for buyers is no longer whether the most expensive model is good enough. It is whether the one that costs a third less per finished task is close enough that the expensive model stops being the default.

Some offers on this page may be paid placements or contain affiliate links.

Subscribe to Techpresso

Free daily newsletter, read in 5 minutes.

Subscribe free