OpenAI drops GPT-4.5 research preview

Just when everyone thought the industry had fully moved on to slow-thinking reasoning models, OpenAI decided to show what happens when you keep pushing brute-force pre-training.

OpenAI just launched a research preview of GPT-4.5. Billed as its largest and best chat model to date, the system was trained on Microsoft Azure AI supercomputers and doubles down on unsupervised learning rather than test-time compute.

Raw intuition over chain-of-thought

Unlike OpenAI o1 or o3-mini, GPT-4.5 does not stop to generate a chain of thought before speaking. Instead, OpenAI is explicitly dividing intelligence into two tracks: reasoning models that tackle heavy math and coding, and unsupervised scaling models that build world models and intuition.

By scaling compute and data, GPT-4.5 shows a noticeable drop in hallucination rates on benchmarks like SimpleQA. OpenAI also used a technique where smaller models generate supervision data to train the larger architecture, aiming to make its tone warmer, more intuitive, and better aligned with human intent.

Early evaluations suggest the model has higher emotional intelligence and stronger aesthetic judgment, excelling at creative writing, nuanced conversation, and high-level agentic planning.

There is a massive catch, though. GPT-4.5 is compute-heavy and significantly more expensive than GPT-4o. OpenAI openly admitted it is not a GPT-4o replacement and is currently evaluating whether it can even justify keeping it in the API long-term.

ChatGPT Pro gets it first

The rollout begins today for ChatGPT Pro subscribers across web, desktop, and mobile. Plus and Team tiers will receive access next week, followed by Enterprise and Edu users.

While GPT-4.5 supports file uploads, web browsing, image inputs, and Canvas workflows, it completely lacks native multimodal features like advanced Voice Mode, real-time video, and screen sharing.

For developers on paid API tiers, it is accessible via the Chat Completions, Assistants, and Batch endpoints, complete with function calling and structured outputs.

According to the accompanying system card, GPT-4.5 rates low in cybersecurity and model autonomy risks, and medium in CBRN and persuasion, showing no significant jump in safety hazards compared to earlier releases.

Why it matters

We think GPT-4.5 sends a crucial signal to the AI ecosystem: traditional pre-training is far from dead. While competitors rush to duplicate o1-style reasoning tricks, OpenAI is proving that raw scale still yields unmatched conversational nuance, creative flair, and factual reliability.

For power users and writers, having a model that intuitively gets subtle prompts without needing pages of instructions is a genuine upgrade. But for developers, the eye-watering inference costs make broad adoption unrealistic right now.

OpenAI deserves credit for brutal honesty by telling developers upfront that the model might not stick around forever if the unit economics do not pan out. Still, the underlying challenge remains clear. Unless OpenAI finds a way to distill this level of world knowledge into cheaper, nimbler weights, GPT-4.5 risks being admired as a technical triumph that is simply too expensive for the real world to run every day.