Introducing Gemini 4 Neon
Gemini 4 Neon is Google’s middle Gemini 4 model for long, complex work: agents, coding, analytics and science. Google says it pairs frontier-level capability with faster end-to-end task completion and lower cost per task than Carbon at its default high thinking level.
Transcript
Pippa So Google just dropped Gemini 4 Neon, and the pitch is basically Carbon-level brains that actually finish gnarly tasks end to end… without eating your whole budget.
Tyler And aimed right at the stuff that hurts today. Long coding agents, spreadsheet-from-hell analytics, science workflows where code, charts, and papers are all tangled together.
Pippa Which is exactly where people hit the wall with Hydrogen. It's fast until you ask it to babysit a repo for three hours.
Tyler Then everyone quietly flips back to Carbon. Or even out to Sol.
Pippa So Neon slots in between. One-million-token context, three dollars per million input tokens, fifteen per million output.
Tyler And it's the new "deep work" model on Google AI Pro and Ultra. Hydrogen for everyday, Neon for serious work, Carbon for the absolute worst projects.
Pippa Okay but… is this just another IQ bump, or is there actually something different here for builders?
Tyler Google's trying to make it about tasks, not raw scores. Neon's #3 overall on the Artificial Analysis Intelligence Index, a 61 at xhigh. But it's #2 on Agent Arena, across 4,200-plus real-world agent sessions.
Pippa Huh, okay.
Pippa And #1 on Zapier's AutomationBench, and Andon's Vending-Bench 2, that long business simulation. So… more workflows finished, cheaper?
Tyler That's Google's claim. At default high, they say Neon costs 32% less per task than Carbon. And it sits on the Pareto frontier on Vals and Agent Arena.
Pippa Thirty-two, huh.
Pippa Pareto frontier being…
Tyler Capability versus cost. Nothing else on that chart is both cheaper AND better.
Pippa And the thinking levels are your dial for that, low through xhigh. High for most deep work, medium for routine jobs, xhigh for the weird research rabbit holes.
Tyler Especially since output is fifteen dollars per million tokens versus three for input. If an agent writes a pile of code and logs… that's the line you feel.
Pippa So the promise is fewer retries, fewer stuck agents, the whole job costs less.
Tyler On coding it's #1 on the Artificial Analysis Coding Agent Index. 65.2% on Terminal-Bench 4.1, 46.2% on FrontierCode 2.0, both at default high. Multi-step, tool-using tests, not just "write a function."
Pippa Mm.
Pippa The harness thing got me, though. On SWE-sweep, Neon gets 12.8% in the mini-SWE harness… and then 39.0% end to end in Google Antigravity.
Tyler Same model. Which doesn't prove your stack gets the same jump, but it does make the orchestration layer part of the result.
Pippa That's the through-line from Hydrogen, right? With Antigravity and /swarm, long work either holds together or explodes.
Tyler And the big demos are there to sell exactly that. The 3D lab bench, the ski sim, the Turing study, the hundred-thousand-cell Sheet
Pippa Spectacle. All of it.
Tyler Sure. But underneath the spectacle, Google's saying Neon can drive tools for a long time, cut redundant file reads, compact context, and show you progress instead of leaving you at a spinner.
Pippa As someone who hates blind spinners, that part makes me oddly happy. Like, tell me what you're doing while you burn my tokens.
Tyler The safety story's aimed at that same use case. Carbon-level protections for severe misuse and prompt injection, plus a monitor watching for chain-of-thought and action misalignment.
Pippa With numbers?
Tyler Google reports 0.0% cheating on Terminal-Bench 4.0. Also a 0.09% misaligned-action rate on its agentic safety suite, and lower alert rates than Carbon in internal coding use.
Pippa Zero point ZERO?
Pippa So if you're handing it tools in prod, that monitor matters as much as one more point on an index.
Tyler My read… this isn't necessarily Google jumping a whole tier. It's a more coherent middle model for long-running agents, and the cost and harness behavior are the things I'd want to test myself.
Pippa Alright, let's wrap there. Next time you're ranking Carbon, Argon, Hydrogen, and Neon by vibes… and I am ABSOLUTELY keeping score.
Tyler Vibes are not a benchmark, Pippa.