GPT-4.5 'Orion': OpenAI's Biggest Model Can't Reason, Doesn't Convince

OpenAI announced on Thursday the launch of GPT-4.5, the model code-named Orion that much of the industry had been anticipating for months. It's the largest model the company has ever trained, built with more computing power and more data than any of its predecessors. And yet, in its own white paper, OpenAI admits it doesn't consider GPT-4.5 to be a frontier model — meaning it isn't at the absolute cutting edge of the field. The lukewarm way the company itself is presenting it says more about the state of AI right now than any benchmark table could.

What GPT-4.5 Is and Who Can Use It

GPT-4.5 first rolled out to ChatGPT Pro subscribers, OpenAI's $200-a-month plan, as part of a research preview — an early-access version the company uses to study how the model performs in practice. Developers on paid API tiers got access the same day. For everyone else — ChatGPT Plus and ChatGPT Team users — OpenAI told TechCrunch the model would arrive the following week.

The company is emphatic on one point: GPT-4.5 is not meant to be a drop-in replacement for GPT-4o, the general-purpose workhorse that powers most of its API and ChatGPT. It supports file and image uploads and works with ChatGPT's canvas tool, but for now it lacks capabilities that GPT-4o already has, like realistic two-way voice mode.

A Telling Detail: The Line That Disappeared

Hours after the launch, OpenAI quietly removed a line from the white paper that read, verbatim, "GPT-4.5 is not a frontier AI model." The updated version of the document no longer includes it. The detail matters for understanding the messaging around this release: the company admitted in writing the limitations of its biggest, most expensive model — and then walked that admission back.

Beyond that edit, the tone of the announcement itself gives away the caution. "We're sharing GPT-4.5 as a research preview to better understand its strengths and limitations," OpenAI wrote in a blog post. "We're still exploring what it's capable of and are eager to see how people use it in ways we might not have expected." That's not the language of a company unveiling a generational leap.

Where It Excels and Where It Falls Short

GPT-4.5 was built using the same core technique that produced GPT-4, GPT-3, GPT-2, and GPT-1: dramatically scaling up computing power and data during a "pre-training" phase known as unsupervised learning. In every previous generation, scaling this way delivered massive performance jumps in math, writing, and coding.

This time, the pattern breaks. OpenAI says the larger size has given GPT-4.5 "a deeper world knowledge" and "higher emotional intelligence." The model responds in a warmer, more natural tone, holds its own on creative tasks like writing, and, according to the company, understands human intent better.

Where it truly shines is factual reliability. On the SimpleQA benchmark, which measures accuracy on straightforward factual questions, GPT-4.5 beats GPT-4o and OpenAI's own reasoning models, o1 and o3-mini. The company says it hallucinates — makes things up — less often than most models. With an awkward caveat: Perplexity's Deep Research model, which performs similarly to OpenAI's on other tests, actually outperforms GPT-4.5 on this factual-accuracy benchmark.

Coding results are mixed. On a subset of the SWE-Bench Verified benchmark, GPT-4.5 roughly matches GPT-4o and o3-mini but falls short of OpenAI's own deep research model and Anthropic's Claude 3.7 Sonnet. On another test, SWE-Lancer — which measures the ability to build full software features — it beats GPT-4o and o3-mini, but again doesn't catch up to deep research.

And on tough academic benchmarks like AIME and GPQA, GPT-4.5 doesn't reach the level of leading reasoning models: o3-mini, DeepSeek's R1, and Claude 3.7 Sonnet. It matches or beats the best non-reasoning models on those same tests, which shows it handles math and science problems well — but the hierarchy is clear: OpenAI's biggest model trails smaller models that "think" before they answer.

The company points to things benchmarks don't capture. In one informal test, it asked GPT-4.5, GPT-4o, and o3-mini to draw a unicorn in SVG, a graphics format based on mathematical formulas and code. GPT-4.5 was the only one that produced anything resembling a unicorn. In another test, given the prompt "I'm going through a tough time after failing a test," all three models gave useful responses, but GPT-4.5's was the most socially appropriate.

The Price: 15 to 30 Times More Expensive

The factor that will most shape GPT-4.5's future is cost. OpenAI charges developers $75 per million input tokens (roughly 750,000 words) and $150 per million output tokens. The comparison with GPT-4o is brutal: that model costs $2.50 per million input tokens and $10 per million output tokens. In other words, GPT-4.5 is 30 times more expensive on input and 15 times more expensive on output.

It's so costly to run that OpenAI admits it's evaluating whether to keep GPT-4.5 in its API long-term. A company questioning whether to keep a product around on the same day it launches tells you everything about its commercial fit.

The Ceiling of Pre-training Comes Into View

OpenAI says GPT-4.5 sits "at the frontier of what's possible with unsupervised learning." That may well be true — and that's precisely what makes it concerning for the industry's dominant strategy. For years, the formula was simple: more data plus more compute equaled more capability. GPT-4.5 suggests that curve is flattening out.

It's not an isolated concern. OpenAI co-founder and former chief scientist Ilya Sutskever said in December that "we've achieved peak data, and there will be no more," adding that "pre-training as we know it will unquestionably end." His comments echoed worries that investors, founders, and researchers had already been voicing.

The industry's response — OpenAI included — has been to bet on reasoning models, which take longer to solve tasks but tend to be more consistent. The idea is that by giving models more time and compute to "think" through a problem, you can meaningfully boost their capabilities without having to train ever-more-colossal systems.

A Stepping Stone, Not a Summit

OpenAI plans to merge its GPT series with its "o" reasoning series, starting with GPT-5 later this year. GPT-4.5 — reportedly extraordinarily expensive to train, repeatedly delayed, and a disappointment relative to internal expectations — likely won't win any benchmark crowns on its own.

Its real value to the company seems to lie elsewhere: it's a large-scale experiment testing the limits of a technique, and a stepping stone toward the fusion of knowledge and reasoning that GPT-5 promises. For users and businesses, the practical takeaway is more sobering. OpenAI's biggest model to date is also one of the hardest to justify economically — and it arrives just as the industry itself is admitting that making models bigger is no longer enough.