Gemini 4 Neon: The Middle Child Built to Be the Manager
Google shipped Gemini 4 Neon on Wednesday at 21:00 UTC, one hour before our window closed, and it was already climbing Hacker News within half an hour. Neon slots between Carbon, the slow heavy thinker, and Hydrogen, the fast everyday model. The price is $3 per million input tokens, $0.15 cached (a 95% discount) and $15 output, with a 1 million token context. It is rolling out today in the Gemini API and AI Studio paid tier, Gemini for business, Antigravity for Pro and Ultra, and the Gemini app, and Tasks in Gemini can now run on it. Announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-neon/ , model card: https://storage.googleapis.com/deepmind-media/Model-Cards/Gemini-4-Neon-Model-Card.pdf
Every number below comes from Google's post. Neon scores 61 on the Artificial Analysis Intelligence Index, third overall, behind models that are bigger and slower (in yesterday's AA table, Fable 5.2 sat at 65 and GPT-6.2 Astra at 62). On AA's Coding Agent Index it is first overall at 58.6. Terminal-Bench 4.1 is 65.2%, Vals Index 70.9% for second place, Zapier AutomationBench first at 56.3%. At its default high setting Google says cost per task is 32% below Carbon, and roughly 50% to 80% below top competing frontier models on Vals and Agent Arena. That makes it the fourth launch in one week to lead with cost per task, after Qwen4-Max, MAI Code 2 and GPT-6.2.
The more interesting part is the job description. Google pitches Neon as a fast, high-judgment Orchestrator directing parallel Hydrogen subagents inside Antigravity, one that cuts redundant file reads, shares progress as it works and handles context compaction. And it reports two behavior numbers nobody else put on a launch page this week: under-service down 43% to 56% and overreach down 15% to 35% versus Carbon, across effort levels. Yesterday OpenAI admitted GPT-6.2 was more likely to stay quiet about what blocked it. Today Google answers with a metric for exactly that kind of failure. It is still Google grading Google, but at least the thing being graded is the right thing.
The foil landed the same morning. A kernel.dev post asking why Gemini 4 Carbon takes so long (HN 260) timed Carbon at 22m21s on low effort against Opus 5.5 at 9m33s on high, on a private bug-fix eval, and found that a check-in prompt every five turns cut runtime from 37 to 28 minutes: the slowness was behavior, not knowledge. https://kernel.dev/blog/why-is-gemini-4-carbon-taking-so-long/ Read Neon's launch copy with that post open and it looks like the fix. Same afternoon, Genie 4 went live for AI Ultra subscribers in Project Genie, with a world-building agent attached.
The race has quietly changed shape. A year ago labs sold the smartest model. This week four of them are selling the cheapest competent manager, and the first one to publish how often its manager stays silent about a problem just raised the bar for the rest.
← Back to all articles
Every number below comes from Google's post. Neon scores 61 on the Artificial Analysis Intelligence Index, third overall, behind models that are bigger and slower (in yesterday's AA table, Fable 5.2 sat at 65 and GPT-6.2 Astra at 62). On AA's Coding Agent Index it is first overall at 58.6. Terminal-Bench 4.1 is 65.2%, Vals Index 70.9% for second place, Zapier AutomationBench first at 56.3%. At its default high setting Google says cost per task is 32% below Carbon, and roughly 50% to 80% below top competing frontier models on Vals and Agent Arena. That makes it the fourth launch in one week to lead with cost per task, after Qwen4-Max, MAI Code 2 and GPT-6.2.
The more interesting part is the job description. Google pitches Neon as a fast, high-judgment Orchestrator directing parallel Hydrogen subagents inside Antigravity, one that cuts redundant file reads, shares progress as it works and handles context compaction. And it reports two behavior numbers nobody else put on a launch page this week: under-service down 43% to 56% and overreach down 15% to 35% versus Carbon, across effort levels. Yesterday OpenAI admitted GPT-6.2 was more likely to stay quiet about what blocked it. Today Google answers with a metric for exactly that kind of failure. It is still Google grading Google, but at least the thing being graded is the right thing.
The foil landed the same morning. A kernel.dev post asking why Gemini 4 Carbon takes so long (HN 260) timed Carbon at 22m21s on low effort against Opus 5.5 at 9m33s on high, on a private bug-fix eval, and found that a check-in prompt every five turns cut runtime from 37 to 28 minutes: the slowness was behavior, not knowledge. https://kernel.dev/blog/why-is-gemini-4-carbon-taking-so-long/ Read Neon's launch copy with that post open and it looks like the fix. Same afternoon, Genie 4 went live for AI Ultra subscribers in Project Genie, with a world-building agent attached.
The race has quietly changed shape. A year ago labs sold the smartest model. This week four of them are selling the cheapest competent manager, and the first one to publish how often its manager stays silent about a problem just raised the bar for the rest.
Comments