OpenAI Releases GPT-4.5 with 1300% Price Increase

TLDR

• OpenAI released GPT-4.5 with a massive price increase – API costs are $75 per million input tokens and $150 per million output tokens, representing a 1300-2900% increase from GPT-4o
• The model focuses on “vibes” and emotional intelligence rather than raw reasoning power, with Sam Altman describing it as “feeling like talking to a thoughtful person”
• GPT-4.5 shows mixed benchmark results – better than GPT-4o but trails behind OpenAI’s o3-mini model in several tests
• The release comes amid intense competition, following launches from Anthropic (Claude 3.7) and xAI (Grok-3)
• Available now for $200/month Pro users, with $20/month Plus users getting access next week

OpenAI has released GPT-4.5, its newest language model, marking a shift in focus from raw intelligence to conversational quality. The launch comes with unprecedented pricing that sets new records in the AI industry.

The model’s API costs $75 per million input tokens and $150 per million output tokens. This represents a dramatic increase from its predecessor, GPT-4o, which charged $2.50 and $10.00 respectively.

The timing of the release is worth noting. It comes just one day after Anthropic’s Claude 3.7 Sonnet launch and a week after xAI’s Grok-3 debut.

OpenAI CEO Sam Altman addressed the pricing head-on during the announcement. He acknowledged the model’s resource-intensive nature, describing it as “a giant, expensive model.”

The company’s focus with GPT-4.5 appears to be on what they term “vibes.” This refers to the model’s emotional intelligence, warmth, and ability to engage in natural conversation.

OpenAI created a special “Vibes test set” to measure these qualities. The test evaluates creative intelligence and conversational quality, areas where GPT-4.5 reportedly excels.

The demonstrations during the presentation focused on everyday scenarios. One example showed the model responding to a user upset about a friend canceling plans.

Another demonstration compared GPT-4.5’s response style to that of OpenAI’s o1 model. The new model showed improved nuance in its responses, though the primary difference was in tone rather than content.

Putting GPT-4.5 to the Test: Benchmark Results

The model’s technical performance presents a mixed picture. On the GPQA science evaluation, GPT-4.5 scored 71.4%, improving upon GPT-4o’s 53.6%.

However, these scores fall short of OpenAI’s o3-mini model, which achieved 79.7% through its reasoning capabilities. The pattern repeats across other benchmarks.

On the AIME ’24 math evaluation, GPT-4.5’s score of 36.7% beat GPT-4o’s 9.3%. Yet it remained far behind o3-mini’s 87.3% performance.

The development of GPT-4.5 required extensive technical innovation. The team built new inference systems and used low-precision training to maximize GPU usage.

Training occurred across multiple data centers simultaneously. This approach helped manage the model’s massive resource requirements.

GPT-4.5 is now available to Pro users who pay $200 monthly. Plus users, who pay $20 monthly, will gain access next week.

The price point puts GPT-4.5 at ten times the cost of Claude 3.7 Sonnet. This makes it potentially out of reach for many developers and startups.

Advertise Here