What is GPT-5?

GPT-5 launched at a lower price point than GPT-4, a result that surprised beta testers and runs counter to assumptions that frontier model releases command premium pricing.

“For five to be cheaper than four, I don't think anyone expected.”

Microsoft's MAI model outperformed GPT-5 on quality for McKinsey tasks while costing 10x less.

“MAI delivered the highest win rate, outperforming GBT55 on quality while being 10x lower on cost.”

Microsoft MAI is worth tracking as an enterprise model that beats GPT-5 on quality and delivers 10x cost savings in real consulting workloads at McKinsey.

“MAI delivered the highest win rate, outperforming GBT55 on quality while being 10x lower on cost.”

OpenAI's decision to lead with efficiency rather than raw capability at the GPT-5 launch is a bearish signal for the broader AI industry, not a triumphant milestone.

“I think it's a bearish sign that the step had to come when they launched GPT5.”

Cursor is teasing a new frontier-scale model trained from scratch, the same size as Claude Opus and GPT-5, using 10 to 20x more compute than its previous Composer model, designed to be generally intelligent rather than coding-only.

“New cursor model being teased at compile. Same size as Claude Opus and GPT55. Trained from scratch. No more Kimmy Base. 10 to 20x more compute versus composer generally intelligent not just coding releases in the next couple of weeks.”

No AI model is expected to deliver a 2x performance gain over the next 6 to 18 months, with the dominant trend being incremental price and performance improvement rather than step-change capability jumps.

“Things are going to keep getting better things will get cheaper the hard agent things will get better and like GBT 5 is not going to be this like crazy super intelligent so I just kind of fleshing out my worldview there so it's like the long slow march of AI is price and performance continuing to be pushed out but not like a 2x performance gain in any new model.”

Cursor's forthcoming model uses 10 to 20x more compute than its prior Composer model, matching the scale of Claude Opus and GPT-5.

“New cursor model being teased at compile. Same size as Claude Opus and GPT55. Trained from scratch. No more Kimmy Base. 10 to 20x more compute versus composer generally intelligent not just coding releases in the next couple of weeks.”

Israetel contends small parameter models are not nearly as capable as large ones for deep cross-domain reasoning, directly challenging benchmark claims that small models are 98 percent as capable.

“People are really impressed by small parameter models doing really cool f*** and they're like oh they're 98% as smart as large parameter models you're based on what kinds of questions eval questions absolutely within distribution logical operations No problem. One of the things I was blown away with by GPT4.5 research preview, which is a larger model than GPT5, way higher parameter count, is the unbelievable nuance and depth and detail and cross-linkage across domains.”

Israetel explains GPT-5 as a GPT-4 architecture that discovered and exploited the reasoning paradigm, optimized for token efficiency and real-world workflows at the cost of deep understanding, distinguishing it from GPT-4.5.

“GPT5 is basically built on our architecture. They basically discovered the reasoning paradigm which GPT4.5 didn't have and they were like f*** we got to exploit the f*** out of this and they managed to do it incredibly efficiently which is why GPT5 is like super token efficient which is like really impressive for real world tasks but absolutely to your point dings its deep understanding because it was not optimized for deep understanding it was optimized for realistic human workflows.”

GPT-4.5 (high parameter count research preview) cited as demonstrating cross-domain nuance and depth that smaller, token-efficient models like GPT-5 trade away, making it a reference point for capability ceiling tracking.

“People are really impressed by small parameter models doing really cool f*** and they're like oh they're 98% as smart as large parameter models you're based on what kinds of questions eval questions absolutely within distribution logical operations No problem. One of the things I was blown away with by GPT4.5 research preview, which is a larger model than GPT5, way higher parameter count, is the unbelievable nuance and depth and detail and cross-linkage across domains.”