Google releases Gemini 4 Neon, a mid-tier model priced below Carbon
원문을 읽지 못해 Hacker News 토론으로 요약했습니다
- Commenters report Neon costs $3/$15 per million tokens, below Carbon at $4/$20 after Nov 30.
- A commenter quoting the Vals Index says Neon ranks #2 at 70.9% and is first on the Legal Research Bench and Tax Agent Bench.
- Commenters say the model size class names do not map to actual size, since Hydrogen, Neon, and Carbon do not follow atomic number or alphabetical order.
- Commenters report Neon is available on the $20 plan in agy, while Carbon requires the $100+ tier, and student accounts hit an entitlement bug.
Hacker News 의견들
Neon is the long-awaited successor to 3.1 Pro, so the tiers would be Carbon/Argon as XL, Astra/Fable as L, Neon/Sol as M, Hydrogen as S. Argon doesn't feel XL to me though.
Other people are saying Neon replaced Argon, so I'm not sure the lineup is settled yet.
Will it show up in Google Workspace? I don't want to pay for a separate AI license just to get it.
It works in agy for me on Pro, and it has its own orchestrator that delegates to Hydrogen.
It's in my $20 sub too. Only Carbon needs the $100+ tier.
Student accounts apparently can't use these at all. There's an entitlement sync bug on the dev forum.
The naming is a mess, but OpenAI and Anthropic aren't better, I'll give Google that.
I thought they were going by atomic number, but that doesn't match size. Carbon is 6 and Neon is 10, and Carbon is the big one.
From the link, Neon looks better, faster, and cheaper than Carbon. I'm not sure what Carbon is for then.
On Artificial Analysis, OpenAI looks like it equals or beats everything Google has at half the price.
Hydrogen covers most of what people actually do. Nobody needs Carbon to plan a birthday party.
As far as I can tell the naming is arbitrary. Argon, Boron, Carbon went alphabetical, then Hydrogen and Neon broke the pattern.
Honestly I like it. The periodic table is an ordered list everyone knows, so it's a rough guide to what the team expects.
I'd guess they're tuning something other than size, like sequence length or sparsity, and shipping all four so the public can pick.
So which model am I supposed to use for what? Nobody in the thread has given me a clean answer.