I ranked every AI model.

For those who just wanted to see the final table.

We def need more of those comparisons Against harnesses, AI labs, & more :)

Dude, you don't even explain why you think 3.7 Flash is F Tier except for ranting about the price... I use a wide range of models for my day job and I still enjoy using Flash 3.7 for personal projects a lot! It's incredibly fast, gets many jobs done, and price doesn't matter if you are on a subscription.

Flash is slower and more expensive than gpt-5.6-sol on medium. Stop looking at TPS it is rotting your brain.

Anyhow, I feel you are giving Google too hard of a time. Gemini Pro models feel pretty useless atm, but I enjoy using Flash much more than Sonnet for example.

Can't wait for this to be outdated in just a few days.

It's impressive that Google now has its own category. IMO, Opus > Sol and Fable (cost). I know it's an unpopular opinion, but for me, the Opus 5 Medium works great. On High and Max, it becomes an annoying smartass

Curious, if you have had a chance to play with Qwen models.

Great video! My experience with Design + SwiftUI work: Sol eats the cake! Fable 5 is really good at design, but Sol is just right behind it. I agree; once OpenAI releases the new model, I hope that gap gets closed. 🤞🏼

Not every, you forgot Ox Alpha. We need a genuine ranking for this one with all the ongoing hype

Grok4.6 and Muse at same tier level ❌

Thanks for ruining the "codex usage after limits expired" perk. I hadn't written "shut the f..k up" just not to increase your tweet's visibility one inch more, now I can say that you are a 90-IQ imposter who financially harmed thousands of people who'd need it.

Terra high is good for /review on codex, Luna Max/xhigh can get stuck in compaction loop, Sol high will take about 10x long and cost 5x as much and go into unnecessary edge cases and such. At the very least, a loop of Terra first cause it's cheaper, then Sol. That's the only use I have for Terra.

When you run out of Fable. Going back to Opus / Sonnet is a bigger downgrade than what this tier list states.

Thanks for the kindness bro, thanks for ruining that up

Terra is much better than you give it credit for, especially when you up a reasoning level over Sol, it’s still much cheaper and has vision, so really doesn’t belong in the group with text only models.

Crazy informative. Theo out here simplifying all of this for everyone

Great one. Did you know that you can add thumbnails and CTA links to videos on X? The feature is quite hidden, but for bigger videos like these, it would make sense to consider it.

Can’t believe Google got SSS++ tier on every model

opus, my beloved

One of the best videos online right now.

stealing this idea lol ty

Sleeping on Grok man

absolutely noot needed just use sol and luna

wait, how do you use multiple subs without breaking their tos?

Benchmarking AI models is like ranking every car by top speed while ignoring the fact that you can't afford gas. Useful? Sure. Accurate to reality? Please stop torturing yourself with spreadsheets

Make your own AI

Couldn't have the issues you had with Sol just have been the wrapper you had around the model such as T3.

Great comparison. The interesting part is actually putting these models to work. We’ve been building Qubrid as a place where developers can test and compare open models through one API. Would love to see some of these rankings using real-world workloads.

someone pls summarize the list

give us the tier list template pls

can feel this is gonna be fire

看 Theo 盘最新一波模型，他聊 Token 效率那段最戳我。做 Agent 容易掉进两个坑：光看每百万 Token 标价谁便宜，或者迷信跑分里的每秒生成速度。 算总账根本不能看单价。有的模型虽然便宜、TPS 极高，但跑一个任务动不动吐出七八万甚至十万字思考过程，最后又贵又慢。按 Theo 的说法，这就跟数自行车轮子转速一样，轮子大小根本不同。5.6 Sol 开到 High 挡位，两三万 Token 就能把活干利落，算下来反而省钱。 我现在搭架构也彻底把层级拆开了。轻活杂活扔给 5.6 Luna，拿来提取标题、把杂乱上下文转成 JSON，顺便做点可逆的工具调用。这层每天调用量巨大，小模型反应快，成本基本可以忽略。 写长流程代码或者复杂编排，我才留给 Sol 或更聪明的模型。至于开源权重，Kimi K3 在设计感和视觉上挺惊艳，但实际计费和 Token 消耗量摆在那，真跑起来可能没大家想的那么便宜。DeepSeek V4 Flash 拿来做后台处理很舒服，但没带视觉让我踩了坑，流水线直接卡住。 做 Agent 跑长链路，模型能用多少 Token 把事情解决、工具调用会不会抽风，直接决定了系统能不能稳住。那些账面参数好看但输出极度冗长的模型，放到实际工程里大概率得交学费。

Which one best?

Every model ranked, so naturally the leaderboard will be obsolete by lunch.

THEO ! STOP USING THE xAI LOGO !! IT'S SPACEXAI.