Edited: For formatting and from feedback
So if I am getting this right?
+-------------------------------+--------------+--------+--------+----------+-------+
| Lab | XL | L | M+ | M | S/XS |
+-------------------------------+--------------+--------+--------+----------+-------+
| Google (Gemini 4 Family #1) | | Argon | | | |
| Google (Gemini 4 Family #2) | | Carbon | Neon | Hydrogen | |
| Anthropic (Claude 5.2 Family) | Fable/Mythos | | | | |
| Anthropic (Claude 5.5 Family) | | Opus | | Sonnet | Haiku |
| OpenAI (GPT 6.2 Family) | Astra | Sol | | | |
| OpenAI (GPT 6.1 Family) | | | | Terra | |
| OpenAI (GPT 6 Family) | | | | | Luna |
+-------------------------------+--------------+--------+--------+----------+-------+
Notes:
* Gemini 4 Family
- Argon -- ++Train --> Carbon
- Carbon -- Distill --> Neon, Hydrogen
* Gemini 3 Family
- 3 Pro -- ++Train --> 3.1 Pro
- 3 Pro -- Distill --> 3 Flash
- 3 Flash -- ++Train --> 3.5 Flash ... -- ++Train --> ... 3.8 Flash
* "S/XS" classification is due to weirdly low pricing for models by Ant/OAI and not by raw perf
I'm not sure it even makes sense to align models on price given that Neon appears to be considerably better than Sonnet/Terra. Not to mention that there are rumors of a flash-lite type model even faster than Hydrogen.
Yeah; I guess. Updated
Yeah Neon competes with Opus 5.5
Neon is within about 1% of Fable 5.2's performance for about 40% the cost on vals.ai, so surely it beats opus 5.5 and competes with Fable 5.2?https://www.vals.ai/benchmarks/vals_index
> I'm not sure it even makes sense to align models on price given that Neon appears to be considerably better than Sonnet/Terra.
Models are nowadays compared in terms of their position with regards to a Pareto front assessing traits such as capabilities and cost because there are clear tradeoffs between cost and capabilities.
I'm sure we are going to see speed show up in the analysis in the near future.
So total names to remember, hydrogen, argon, carbon, neon, luna, sol, terra, astra, opus, fable, sonnet, haiku
Then each one has a version. Then each one has an effort of thinking, low, mid, high, extra, ultra.
Good lord.
Yeah but all the names are rankable with the exception of fable.
But wait! There’s more!
Yeah, I just use whatever agy sets as the model when I fire up the IDE or CLI. I don't hit the limit on my personal projects (I'm retired), and it's fast and good every time. And free because I get some kind of bonkers usage caps with my Google Fi.
I know I need to change the hard-coded model name in my foreign language acquisition app because I guess with 4 being the "mid-range", the 3.8 I use is going to be deprecated?
Yeah but you don't have to remember let, for, var, const... and a few thousand rules, any class names or what your program does or uses.
This is where the consultants come in.
Only if you subscribe to all of them. Most people have one paid AI subscription with one provider so they have to learn ~four names for the models. Four is not good-lord-tier for the amount of value it brings. The thinking effort names are just regular English words.
Would you say the same about gadgets like oh my god I have to remember iPhone, iPad, iMac, MacBook pro and Air and Samsung Galaxy etc, or with gpus or with CPUs or with car brands or literally anything??
I had a chat with Gemini this morning, Argon seems to be deprecated and replaced by Neon, so neon just occupies argon’s tier now. Argon is no longer selectable on my ultra configured antigravity account.
Carbon costs the same as Argon did. Neon is a cheaper model.
And after short session, I don't think these models are comparable.
Feels like another rug pull.
At what point did Google even offer a decent rug?
the 1.5 through 2.5 generations. you were given quite some generous usage for free in Google AI Studio API. I knew it would break pretty soon, but in the meantime I built a pretty decent reverse-engineering sidekick. LLMs are much better at pattern-matching disassembler output than I am, but someone still needs (yes, still needs) to draw circles and arrows to capture a program's behavior.
Even early in the 3.1 era the combination of 3.1 pro and Antigravity made an excellent agent. Then they reduced the quota for pro models from very generous down to zero.
Currently Neon is more expensive on OpenRouter...
$3/$15 vs $2/$10
Carbon is 50% off, while Neon is not; at full price Carbon is the "premium" model.
Argon is also supposed to be 50% off
But Neon appears to be stronger than both in benchmarks
Confusing line up!
It's a guest house where rugs are constantly changing, often while the guests are still dancing, it's just part of the scenery at this point. Just remember to not stop dancing...
As I understand it, Gemini 4 family should be shifted 1 step to the left. Argon / Carbon are supposed to be XL and in the same class as Fable / Astra.
They might be weight/parameter count wise, but definitely not perf (this is while over indexing on coding though) or price wise.
Well... This industry moves fast. Carbon had a higher AA score than Astra and Fable when it was announced, though the other labs have updated their models since. Fable 5.2 is clearly better than Carbon; with Astra 6.2 it's a little less clear. I don't think there's much question it's meant to compete against the other labs super-big models.
It's cheaper than astra/fable, but that's likely because they know they've been out of the game for awhile and need to earn back mindshare. Also Carbon is pretty token heavy.
And another thing is that AFAIK in the past Google replaced lack of skill (relative to Ant/OAI) with size (as they had plenty of hw). So Geminis were always the largest model sizes, larger than Opus/GPT family. I'd assume this holds even today, with Carbon vs Fable/Astra.
Carbon's relative slowness would support this as well.
To me Carbon feels like it's in the same class as Fable and Astra.
For a recent prototype I used Carbon as a big-brain planner and Astra-6.2 an implementer. It cut some corners while implementing so it is somewhat Sonnet-ish if you know what I mean
People also report that Carbon solved some hard tasks which Fable and Astra could not
This is illegible on mobile.
Mobile sucks that way.
Rotating screen 90 degrees seems to fix it
Not on mine
I took the further step of toggling "Desktop site" and then the whole thing displayed beautifully. shrug
I have my rotation locked but still started turning my phone anyway like an idiot. The coffee clearly hasn't kicked in yet.
better go buy a Duo then I guess ;-)
Try another 90
Give it to an AI to reformat.
It's a comment, not a product you paid for.
From looking at the model selection in Gemini Web and Antigravity:
Flash -> Hydrogen (medium level) -> Argon -> Neon Pro -> Carbon
Because Argon and Carbon existed for a while at the same time but treated differently. Pro users get Argon but only Ultra users get Carbon.
The issue is that this assumed the models are not nerfed. But in real life you can find Astra performing worse than Terra or Fable being more or less the same as Sonnet as these companies engage in dirty tricks (selective scamming - some people apparently get full fat models whilst the others don't, so you have people coming and saying works for me you must be dumb and the others say e.g. Opus is crap)
I like this. I wish the big sites would categorize like this. For most people XL is not needed. M or M+ is good enough for most. If you really are pushing it then go L.
I disagree, but maybe I'm not most people. XL is my daily driver, I can't imagine dealing with the mess an M model creates.
I can only use Fable now as an orchestrator just because it feels so much more intelligent to talk with. Even for silly tasks, it feels so much nicer.
YMMV; IME Opus-5.5 medium is excellent for many purposes.
I don't see it as clear cut; Opus 5.5 M is very good at doing what it is told; XL often suggests other approaches that just confuse the current work.
This is a great visualization. Thanks for putting it together. If you'll placate my OCD brain for a second, it does bother me that Google and Anthropic have the oldest models at the top and OpenAI has the oldest model at the bottom.
As someone who stayed away from Gemini models as they were apparently unable to reliably use tools for a long time, they could have named the models with UUIDs and would cause less confusion.
Argon was really good, sad to see it go.
Oh big Thank You. I was about to post a ranty comment about a month or so ago but decided not too. Am I suppose to remember every single name of every single model and their size classification?
At this point I just wish they all name it as Ultra, Pro Max, Pro, $Version Number and SE .
Whose idiot idea was it to go for a period table naming scheme? Are we supposed to just think that the heavier elements are the smarter models? That doesn't seem to hold up in practice.
I have based this on model pricing (which I guess also corresponds to their labs's idea of performance/how they compete) and for google their model cards which specify which model is derived from what
I don't know how fair this is to openAI. Astra and Sol cost and speed wise could be in two groups, xl+l and l+m respectively.
Put Astra in L ...
Why do you say so? OAI clearly pits it against Fable no? Are you talking from actual perf perspective?
All of lists like mine are still subjective honestly despite me/others trying to make it rooted in as much objective things as possible. For now the current table is what I think reflects that best
Cost and speed. It should be in xl+l
how would you fit in Deepseek