GPT-4.5, OpenAI's big new model, represents a significant step forward for AI's industry leader. It could also be the end of an era.

The big picture: 4.5 is "a giant, expensive model," as OpenAI CEO Sam Altman put it. The company has also described it as "our last non-chain-of-thought model," meaning — unlike the newer "reasoning" models — it doesn't take its time to respond or share its "thinking" process.

Why it matters: The pure bigger-is-better approach to model pre-training now faces enormous costs, dwindling availability of good data and diminishing returns — which is why the industry has begun exploring different roads to continuing to improve each new AI generation.

Between the lines: Building and powering the massive data centers required to build and run the latest models has become an enormous burden, while assembling ever-bigger datasets has become challenging, since today's models already use nearly all the data available on the public internet.

Yes, but: Although pre-training may have hit a wall, most of the industry remains bullish on new gains to be made with reasoning.

Catch up quick: OpenAI yesterday released an early version of GPT-4.5, a major update to the large language model underlying ChatGPT that OpenAI says will be better at recognizing patterns and drawing connections.

• This is OpenAI's largest model yet — though the company declined to offer details about its size or the computing resources it took to train it.
• While OpenAI isn't sharing details, the cost is clearly substantial, as evidenced by the fact that developers are being charged 30 times as much for their use of GPT-4.5 compared with the current cost of GPT-4o.
• OpenAI says GPT-4.5 should hallucinate less, follow instructions better and deliver interactions that feel more natural.

OpenAI turned on the new model yesterday, but only for subscribers of the $200-per-month ChatGPT Pro and developers who use OpenAI's API.

• Next week it will be available for some other paid subscribers, including the $20-per-month ChatGPT Plus service, with paid enterprise and educational customers getting access the following week.

What they're saying: Altman wrote in a post on X, "Good news: it is the first model that feels like talking to a thoughtful person to me. ... It's a different kind of intelligence and there's a magic to it I haven't felt before."

• One skill the new model seems to have mastered is "reading the room," for instance when a user might prefer a conversation rather than be handed a pile of facts.

Zoom in: Box CEO Aaron Levie, whose company has been testing GPT-4.5, says that it shines in certain areas, such as accurately extracting the proper information from very large datasets. In such tasks, Levie told Axios, GPT-4.5 is about 20% better.

• "We're very much in the camp of not diminishing returns yet," Levie said in an interview, adding that updates like GPT-4.5 are "continuing to drive new step function improvements on reasoning capabilities, logic, math, a bunch of things that really matter to our world in the enterprise."
• Levie said it makes sense to use models like GPT-4o for some tasks, such as summarizing documents, especially given how much the cost of that model has come down.
• "But if you go to a bank or a large law firm and they need to run mission critical operations on their data, then they would absolutely pay the five or 10 times increase on these more powerful models — because it's still far cheaper than their alternative of just throwing humans at the problem."

Yes, but: Levie said the next era of gains will likely come from improving the reasoning that sits on top of large language models.

• "If the foundation model is extremely powerful and then you're doing chain-of-thought thinking on top of that model, then you get very, very high-impact results," he said.
• Altman has already said that the next big release, GPT-5, will integrate reasoning capabilities from its inception.

Between the lines: Former OpenAI chief research officer Bob McGrew says the question isn't either/or, but comes down to where AI companies commit their resources.

• "That o1 is better than GPT-4.5 on most problems tells us that pre-training isn't the optimal place to spend compute in 2025," McGrew said in a thread on X.
• "There's a lot of low-hanging fruit in reasoning still. But pre-training isn't dead, it's just waiting for reasoning to catch up to log-linear returns."

A key question about GPT-4.5 is how much better it is than its predecessors — and whether that edge justifies its higher cost. So I asked GPT-4.5 to make the case for itself.

Zoom in: Axios had access to GPT-4.5 via a demo ChatGPT Pro account.

• So I asked the model to make the case that it offers more than incremental improvements and is worth the billions it surely cost to do all that pre-training. (OpenAI isn't sharing details on the cost or computing that went into training GPT-4.5.)

What it's saying: The model enthusiastically took on the challenge, offering five areas that "underline the significant leap from GPT-4o (and other predecessors)."

1. "Advanced Reasoning & Complex Problem Solving:" "I can handle multi-step, nuanced reasoning tasks more accurately and efficiently than GPT-4o," GPT-4.5 said, offering a specific example of a logic problem that has challenged earlier models.

2. "Enhanced Creativity & Flexibility:" As evidence, GPT-4.5 served up a "brief, poetic description of artificial intelligence from the perspective of a vintage typewriter." (My thought bubble: Does it know I have a vintage typewriter on my desk?)

3. "Superior Contextual Memory:" Here GPT touted its ability to manage complex multi-turn conversations "with enhanced recall of subtle contextual nuances." It provided an example conversation in which a user asked if it remembers their favorite Lego theme and it responded, "Absolutely — I remember clearly! Your favorite Lego theme is space." (My thought bubble: OK, it clearly knows me.)

4. "Real-time Integration & Dynamic Information:" While prior models were also able to draw on the current web, GPT 4.5 knocked its older sibling, saying: "GPT-4o typically required manual guidance to achieve this smoothly."

5. "Improved Instruction Interpretation & Safety:" To highlight its improvements in understanding nuance and observing ethical, legal and practical boundaries, GPT-4.5 cited its refusal to lay out step-by-step instructions for hacking into someone else's social media account, avoiding common GPT-4o pitfalls. (Yes, but: GPT-4o rejected the same query with a similar response.)

Between the lines: Not content to let the examples speak for themselves, GPT 4.5 offered several points to explain why OpenAI's massive investment was justified, including greater efficiency, precision and reliability and the ability for entirely new business cases.

It even offered up an Axios-style "Bottom Line."

• "The leap from GPT-4o to GPT-4.5 isn't incremental — it's transformative," it said. "The significant investment translates directly into measurable improvements in capability, creativity, reliability, and real-world value."

• Microsoft released a native Mac app for Copilot. (The Verge)
• Meta is planning to release a standalone AI app and will test a paid version of the service, sources told CNBC.

I love learning about clever ways to do stuff better, like using a fork to help hang pictures on the wall.