OpenAI Unveils GPT-4.5: A Focus on Conversational Nuance

OpenAI has pulled back the curtain on GPT-4.5, the latest iteration of its flagship large language model (LLM). The company is touting it as its most capable model yet for general conversation, representing what OpenAI research scientist Mia Glaese calls “a step forward for us.”

This release clarifies OpenAI’s recent strategy, which seems to involve two distinct product lines. Alongside its “reasoning models” (like the previously mentioned o1 and o3), GPT-4.5 continues what research scientist Nick Ryder refers to as “an installment in the classic GPT series” – models focused primarily on broad conversational ability rather than step-by-step problem-solving.

Users with a premium ChatGPT Pro subscription (currently $200/month) can access GPT-4.5 immediately, with a broader rollout planned for the following week.

Scaling Up: The Familiar OpenAI Playbook

OpenAI has long operated on the principle that bigger models yield better results. Despite recent industry chatter – including from OpenAI’s former chief scientist Ilya Sutskever – suggesting that simply scaling up might be hitting diminishing returns, the claims around GPT-4.5 seem to reaffirm OpenAI’s commitment to this approach.

Ryder explains the core idea: larger models can detect increasingly subtle patterns in the vast datasets they train on. Beyond basic syntax and facts, they begin to grasp nuances like emotional cues in language. “All of these subtle patterns that come through a human conversation—those are the bits that these larger and larger models will pick up on,” he notes.

“It has the ability to engage in warm, intuitive, natural, flowing conversations,” adds Glaese. “And we think that it has a stronger understanding of what users mean, especially when their expectations are more implicit, leading to nuanced and thoughtful responses.”

While OpenAI remains tight-lipped about the exact parameter count, they claim the leap in scale from GPT-4o to GPT-4.5 mirrors the jump from GPT-3.5 to GPT-4o. For context, GPT-4 was estimated by experts to potentially have around 1.8 trillion parameters. The training methodology reportedly builds on GPT-4o’s techniques, including human fine-tuning and reinforcement learning with human feedback (RLHF).

“We kind of know what the engine looks like at this point, and now it’s really about making it hum,” says Ryder, emphasizing scaling compute, data, and training efficiency as the primary drivers.

Performance: A Mixed Bag?

Compared to the step-by-step processing of reasoning models like o1 and o3, “classic” LLMs like GPT-4.5 generate responses more immediately. OpenAI highlights GPT-4.5’s strength as a generalist.

• On SimpleQA (an OpenAI general knowledge benchmark), GPT-4.5 scored 62.5%, significantly outperforming GPT-4o (38.6%) and o3-mini (15%).
• Crucially, OpenAI claims GPT-4.5 exhibits fewer “hallucinations” (made-up answers) on this test, fabricating responses 37.1% of the time versus 59.8% for GPT-4o and 80.3% for o3-mini.

However, the picture is nuanced:

• On more common LLM benchmarks like MMLU, GPT-4.5’s lead over previous OpenAI models is reportedly smaller.
• On standard science and math benchmarks, GPT-4.5 actually scores lower than the reasoning-focused o3-mini.

The Charm Offensive: Conversation is King?

Where GPT-4.5 seems engineered to shine is in its conversational abilities. OpenAI’s internal human testers reportedly preferred GPT-4.5 over GPT-4o for everyday chats, professional queries, and creative tasks like writing poetry. Ryder even notes its proficiency in generating ASCII art.

The difference lies in social nuance. For example, when told a user is having a rough time, GPT-4.5 might offer sympathy and ask if the user wants to talk or prefers a distraction. In contrast, GPT-4o might jump directly to offering solutions, potentially misreading the user’s immediate need.

Industry Skepticism and the Road Ahead

Despite the focus on conversational polish, OpenAI faces scrutiny. Waseem Alshikh, cofounder and CTO of enterprise LLM startup Writer, sees the emotional intelligence focus as valuable for niche uses but questions the overall impact.

“GPT-4.5 feels like a shiny new coat of paint on the same old car,” Alshikh remarks. “Throwing more compute and data at a model can make it sound smoother, but it’s not a game-changer.”

He raises concerns about the energy costs versus perceived benefits for average users, suggesting a pivot towards efficiency or specialized problem-solving might be more valuable than simply “supersizing the same recipe.” Alshikh speculates this might be an interim release: “GPT-4.5 is OpenAI phoning it in while they cook up something bigger behind closed doors.”

Indeed, CEO Sam Altman has previously indicated that GPT-4.5 might be the final release in the “classic” series, with GPT-5 planned as a hybrid model combining general LLM capabilities with advanced reasoning.

OpenAI, however, maintains faith in its scaling strategy. “Personally, I’m very optimistic about finding ways through those bottlenecks and continuing to scale,” Ryder states. “I think there’s something extremely profound and exciting about pattern-matching across all of human knowledge.”

This story was originally featured on MIT Technology Review.