Why is OpenAI so much more efficient?

DeepSWE harness is garbage. They are putting a monkey wrench in Gemini.

codex is the dumbest model, that's why it's the most efficient, because it doesn't get any shit done, but you can use it 24/7. It's so fkin bad that it disgusts me.

Did you figure out that as long as it's a probabilistic bot that means there is 100% chance there is always a margin of error, meaning it will drift language to the point where the architecture language drifts towards playing the telephone game and the entire thing becomes something entirely different? The reason it's NOT a spacial problem is because it's not reasoning by measurements between objects, it's failing to reasoning about the DIFFERENCES between the objects. Look up "difference" in a philosophical definition. It doesn't work on definitions of words, meaning it's never ever going to stop drifting language. Try to MAKE it stick to definitions over the course of work on a project, you can't. It will forcefully find a way to avoid it and change what you're doing. Without that nothing will ever stay static in your code, it will forever try to mess up existing correct design patterns and code itself. Meaning it will never achieve "done" in it's own view of the set of algorithms as a code base. Also it will never be able to automate anything because it literally can not follow a set of instructions repeatedly without failing due to it's token probability. Go play with a language called LEAN and learn what Neuro-Symbolic AI is and how it will use LEAN instead. "Symbolic" reasoning (in a semiotics sense) is the only way to apply actual real logic to a problem. AI can never reason if it does not logically understand truth. If it can't stick to defined words, it will never be able stick to logic consistently or coherently as both are required. It can't have a basis for any kind of morality because it can't distinguish logic. Meaning a current LLM won't ever have empathy. Look up the term "formal verification" how a real licensed engineer would know it as a constraints based system. It will literally fight its way out of it every time. LLM's a truly a dead end that big tech keeps using to destroy their competition (small/medium sized businesses) because they are dumb enough to "do what google does". It's a corporate weapon in the competition space.

This technical deep dive into model efficiency and the 'Grug speak' theory was fascinating. It's wild to see how much OpenAI is squeezing out of smaller token budgets compared to the competition. Great breakdown!

19:20 The caveman skill

Funny! Sometimes I go full caveman when I prompt AI. Knowing this is how the model thinks and works now, I feel like this is going to change how I behave when I get frustrated with it. I kind of feel sorry for it. It's just doing it's best.

Can you change the reasoning style via system prompt and how this affects output? Can you modify Claude system prompt to reason as grug developer ?

"If you can't explain it simply, you don't understand it well enough" Albert Einstein

Obsidian-posting vs slop-posting. will we ever know?

latest ChatGPT
(:user-message "The skater is skating switch stance. They pop on their tail, flip the board in kickflip direction, and the board spins 180 degrees backside while also flipping. The skater does not spin."
:llm-reasoning "Identify rotation components: kickflip axis + backside 180 board rotation with no rider rotation. In switch stance this maps to a switch backside 180 kickflip."
:extended-llm-reasoning "Board performs a kickflip (longitudinal flip) combined with a backside 180 rotation while rider remains stationary. This matches the standard composite trick classification 'backside 180 kickflip'. Since stance is switch, prefix applies."
:llm-response "switch backside 180 kickflip")

Let's take a step back: Is there a way to verify how many tokens are used to accomplish anything? Is there any way to verify how many times your Facebook ad post was viewed? Anybody auditing all this.... just step back and ponder that 😢

This is some awesome insight!

It feels weird that reasoning works at all

Doesn't matter really. All models are +/- same. The cheapest always win. Kimi, glm etc are much more efficient cost in mind

GLM 5.2 video when

1984 Newspeak

Actually, I'm very surprised it's not more pronounced to the point where it has 'invented' its own language that's 10x more efficient than english.
I was expecting it to talk complete gibberish to itself.

and i thought why is gpt5.5 eats up limits so much faster than claude. turns out always-caveman policy doesn't affect token usage for gpt5.5.

came to say earning $13 billion in 2025 and losing $21 billion doesn't seem very efficient imo
but you meant token efficiency, can't argue with that

I have some of those COT traces in my most recent codex thread. They were accidentally leaked into the chat for some reason. It was really funny. And yes, it does use caveman (grug dev) speak internally when it reasons.

I wonder if you explicitly tell codex to not think like a caveman, would it perform even better at the cost of longer thinking and more cost?

Looks like Kevin from Office was ahead of its time and had a point with not wasting words 😂

I want to reply to your question at the end. This was such a fun, interesting, insightful, and educational presentation. Thank you. So yes I really liked this technical deep dive.

In before the inevitable Sonnet 5 video.

I think that they use Chinese grammar for reasoning since English has a bunch of small words that is not needed. Can a Chinese speaker let me know?

I'm finding Theo's videos repetitive as of late, and I'm wondering if I changed, he did, or the industry has simply gotten less interesting. It's like I have nothing to learn from him anymore.

I usually troll t3 for shilling for SaaS, but even with his short-sightedness, he is starting to get it. Great video and good information.

Kimi 2.7 also thinks like a caveman

The only YT channel i dont skip sponsors on since its actually useful

Open AI's models think like Jeffery Epstein texts

Great vid Theo. Thx

GPT 5.5 Instant used 44% of my T3 Base usage budget with one new single-sentence coding message in the middle of the night. Either Theo has the cost multiplier set too high, or it's not that efficient.

They copied deepseek. Tada

heads up theo, I dont think ephemeral reasoning traces are recomputed in cache as they are never reinserted in to model context. They are computed once only and never re-injected. They have their own lane which is different from active context lane... Thats the reason why LLM's dont know what they "thought" about x or y, they don't have access to that information by design.

That is why GPT writes such bad text

I SEE IT NOW! IT'S BEN FROM PARKS AND REC. But yeah, i was also wondering how much they are just making up the token usage which is not transparent and it dictates how much they charge you

Who cares they are all losing money and they can’t do that forever

One thing people miss is that the large contexts are not only expensive for us over the API, but also expensive for the company serving the model.
- Each request has it's own KV cache and takes up space on the HBM of the GPU/ ASIC cluster.
- If I have one request, in the batch that is processed, with 200k context and two with 100k then they take up the same space of HBM (roughly, of course there is also other things to consider like the reserved space for output token and their KV pairs)
- It is possible to *efficiently* run differently large requests in the same batch, as token generation is bottlenecked by moving the weights not compute
- If all three requests now generate X output token then the 200k request was roughly twice as expensive for the provider than the 100k requests that generated twice the revenue
- Input token revenue cannot compensate for that
- This only/ mostly holds true during peak demand times when the utilisation of the GPU clusters is close to 100%
- Anthropic is drowning in money thrown after them and probably won the marketing game
- OpenAI needs to improve their profit and loss statement and so has all incentive to
a) decrease context size to generate more revenue during peak demand on their GPU clusters
b) decrease context size to being a cheaper alternative to Anthropic and get some of their users

Innocent Theo didn't know what was coming with 5.6 😁

Sometimes when I use Pi on long context with 5.5 xhigh it starts to spit out its reasoning traces as the summary and you can see he is talking like caveman. I don’t have any extensions/system prompts loaded. It just talks like this: Need answer: /build prompt expands to /build then text, created maybe to allow /build <task>
doesn't accept task, so prompt template expands to command plus task? Could cause command on…
/build maybe extension command checked first and rest? Hmm prompt templates expansion before…

really compact

The last Artificial Analysis chart you show proves all this efficiency doesn't matter.. Look at Gemini's position in the chart?

Wait till Theo finds out that having code that is less complex, like htmx, requires less thinking for humans as well

Maybe this is a bit conspiratorial and definitely a bit off topic, but I don't get why we just accept OpenAI/Anthropic/Google's word for it when they charge us for reasoning tokens.

Even aside from the obvious conflict of interest, it's not like we haven't seen labs make mistakes with how they handle caching and token calculations, and when it's not uncommon for reasoning to make up the bulk of a conversation's token use, even small bugs there could add up dramatically.

I bet this is the cause of the goblin problem. 5.5 is reasoning in goblin mode

oh so kevin was on to something

The optimisation is also the reason why ChatGPT fails at the kind of deep synthesis and data analysis that Claude excels at.