OpenAI starts rolling out GPT-4.5 to Plus users

For the past year, every new flagship AI model arrived with promises of shattering every benchmark in sight. This time, OpenAI took the exact opposite approach.

OpenAI CEO Sam Altman announced that rollout has begun for GPT-4.5 to ChatGPT Plus subscribers, with access gradually arriving over the next one to three days. It represents OpenAI's largest and most computationally intensive model to date, though Altman noted that rollout speed is limited by server capacity.

Great conversationalist, terrible math whiz

Under the hood, GPT-4.5 was trained with significantly more data and compute than anything OpenAI has released before. That massive footprint shows up in its worldly knowledge, conversational nuance, and noticeably lower hallucination rates.

OpenAI highlighted that the model shines at practical problem-solving, everyday chats, and creative writing. Intriguingly, internal benchmarks showed it excels at persuading other AI models in dialogue.

Here is the catch: it falls short when things get logically demanding.

Altman bluntly admitted on X that GPT-4.5 is not a reasoning model and will not crush benchmarks. In complex, multi-step problem-solving, it trails rivals like Anthropic's Claude and DeepSeek's recent models.

Then there is the sticker shock. At $150 per million output tokens, GPT-4.5 is roughly 15 times more expensive than GPT-4o, making large-scale production use a tough sell.

Why it matters

For everyday ChatGPT Plus users, GPT-4.5 is worth taking for a spin if you do a lot of long-form writing, persona-heavy roleplay, or complex brainstorming. Its depth of tone and reduced inaccuracies are genuine quality-of-life upgrades. But for developers building automated reasoning pipelines or cost-sensitive tools, this is not the daily driver.

Industry-wide, this release cements the bifurcation of modern AI. One track focuses on reinforcement learning and deliberate chain-of-thought reasoning, while the brute-force pre-training track pushes parameter size for broader real-world knowledge and EQ. GPT-4.5 sits squarely in the latter camp.

We think Altman was smart to proactively deflate the hype bubble before users started testing math prompts. Still, a 15x price tag over GPT-4o exposes the brutal diminishing returns of traditional scaling. Brute-forcing world knowledge through sheer size gets insanely expensive, and without specialized architectural breakthroughs, giant conversationalists may struggle to justify their massive computational overhead.