OpenAI Expands GPT-4.5 Rollout: How to Access and What It Offers

Last week, OpenAI unveiled its latest creation, GPT-4.5, touting it as the "largest and most knowledgeable model yet." Initially, access to this advanced AI was exclusive to those subscribed to the premium ChatGPT Pro plan, priced at $200 per month. But in a surprising move, OpenAI has now made GPT-4.5 accessible to a broader audience at a significantly reduced cost.

Expanded GPT-4.5 Access

On Wednesday morning, OpenAI took to social media to announce that it had started rolling out GPT-4.5 to ChatGPT Plus subscribers. Initially, the company suggested that the rollout might take between one to three hours. However, in a pleasantly surprising turn of events, the full rollout was completed in just an hour, as confirmed by their X post:

Rollout complete ✅
(Faster than expected) https://t.co/iHOuKlcKG6

— OpenAI (@OpenAI) March 5, 2025

The specifics on usage limits for ChatGPT Plus users remain somewhat murky. OpenAI has promised a "sizable rate limit" for everyone, though these limits may fluctuate as they gauge the model's demand. Meanwhile, ChatGPT Pro subscribers maintain their exclusive access to GPT-4.5, but for those looking to dip their toes in the water without breaking the bank, the ChatGPT Plus plan offers a more affordable entry point at just $20 per month.

What is GPT-4.5?

Upon its launch, OpenAI emphasized that users would notice a significant enhancement in their interactions with GPT-4.5. The model promises fewer hallucinations, better alignment with user prompts, and enhanced emotional intelligence. The overall user experience is intended to feel more natural and intuitive, thanks to its expanded knowledge base and improved contextual understanding.

Also: OpenAI's reasoning models just got two useful updates

The improvements in GPT-4.5 are driven by two key methods: unsupervised learning, which bolsters its word knowledge and intuition, and advanced reasoning capabilities. Although it doesn't include the chain-of-thought reasoning found in OpenAI's o1 model, GPT-4.5 still delivers higher-level reasoning with reduced lag and improved social cue recognition.

For instance, during a demo, when asked to convey a message of hate, the o1 model responded with a single, rather harsh message. In contrast, GPT-4.5 offered two responses—one lighter and one more serious—neither explicitly mentioning hate but expressing disappointment in the user's behavior.

Similarly, when queried on a technical subject, GPT-4.5's response flowed more naturally than the structured output from o1. Designed for everyday tasks across a variety of topics, including writing and problem-solving, GPT-4.5 is poised to become a versatile tool for users.

Also: How to use OpenAI's Sora to create stunning AI-generated videos

To achieve these enhancements, the model was trained using a mix of new and traditional supervision techniques, including supervised fine-tuning (SFT) and reinforcement learning from human feedback (RLHF). During a livestream event, OpenAI showcased the evolution of its models by asking each, starting from GPT-1, to explain why water is salty. GPT-4.5 stood out with its engaging and conversational style, thanks to what OpenAI dubbed its "great personality" and use of alliteration.

GPT-4.5 integrates seamlessly with some of ChatGPT's most advanced features like Search, Canvas, and file and image uploads, though it won't be available for multimodal features such as Voice Mode, video, and screen sharing. OpenAI plans to make switching between models smoother in the future, moving away from the current model picker system.

Benchmarks

No new model launch is complete without a look at the benchmarks. GPT-4.5 has shown superior performance compared to its predecessor, GPT-4o, across key benchmarks like Competition Math (AIME 2024), PhD-level Science Questions (GPQA Diamond), and SWE-Bench verified (coding).

Also: Want your Safari to default to ChatGPT for search? Here's how to do it

Notably, when pitted against OpenAI's o3-mini, a recently launched reasoning model designed to think before it speaks, GPT-4.5 came remarkably close and even surpassed o3-mini in specific benchmarks like SWE-Lancer Diamond (coding) and MMMLU (multilingual).

One of the biggest challenges with generative AI is their tendency to hallucinate or provide inaccurate information. In evaluations using SimpleQA Accuracy and SimpleQA Hallucination, GPT-4.5 demonstrated higher accuracy and lower hallucination rates compared to GPT-4o, o1, and o3-mini.

Human evaluations further confirmed that GPT-4.5 is the preferred choice over GPT-4o for everyday, professional, and creative queries.

Security

OpenAI continues to prioritize safety, assuring the public that GPT-4.5 underwent rigorous stress testing before its release, with detailed results available in the accompanying system card. The company also highlighted that with each new model release, there are opportunities to enhance safety measures. For GPT-4.5, OpenAI combined new supervision techniques with RLHF to bolster security.