TLDR
- The release of GPT-4.5 by OpenAI comes with a steep price hike; API access now costs $75 per million tokens for input and $150 for output, a drastic rise compared to GPT-4o.
- GPT-4.5 prioritizes emotional resonance and conversational skill over sheer computational power, with Altman likening the interaction experience to conversing with a considerate individual.
- Performance evaluations of GPT-4.5 are variable, showing improvements over GPT-4o but lagging behind OpenAI’s o3-mini in several areas.
- The launch occurs amidst fierce rivalry, coming on the heels of Anthropic's Claude 3.7 and xAI's Grok-3 introductions.
- Currently, it's accessible to $200/month Pro users, with $20/month Plus users set to join the experience in a week.
Introducing GPT-4.5 marks OpenAI's transition from raw computing strength to enhanced conversational capability, with unique pricing breaking new grounds in the AI sector.
For the GPT-4.5 API, OpenAI has set the cost at $75 for every million input tokens and $150 for each million output tokens, a significant rise from the rates of GPT-4o, which were $2.50 and $10.00.
The strategic timing aligns with competitive movements, launching just a day post-Anthropic's Claude 3.7 Sonnet and a week after xAI's Grok-3.
During the announcement, OpenAI’s Altman candidly addressed the costly nature of the new model, describing it as a 'giant, costly creation' requiring substantial resources.
GPT-4.5 is ready!
On a positive note, this is the first model that offers an experience akin to chatting with a thoughtful human, leaving me amazed at the quality and thoughtfulness of AI-given advice.
On the downside, the model's extensive architecture makes it quite an expensive endeavor, requiring…
— Sam Altman (@sama) February 27, 2025
OpenAI’s emphasis for GPT-4.5 lies on the 'vibes' spectrum, emphasizing emotional acuity and the model’s adeptness at naturally flowing dialogue.
A specialized 'Vibes test set' was crafted by OpenAI to assess these attributes, spotlighting creativity and conversational proficiency where GPT-4.5 seems to excel.
Presentations highlighted the model's everyday utility, such as effectively handling a user's frustration over a friend's last-minute plan cancellation.
In comparisons, GPT-4.5's response nuances stood out more than its predecessors, reflecting a significant tonal evolution alongside content similarities.
Evaluating GPT-4.5: Insights from Benchmark Trials
In technical assessments, GPT-4.5 delivers mixed results, achieving 71.4% on the GPQA science test, a leap from GPT-4o's 53.6%.
Nonetheless, performance trails behind the high-reaching o3-mini model that scored 79.7% largely due to its reasoning strength, a trend seen across various comparisons.
Within the scope of the AIME ‘24 math test, GPT-4.5 posted a 36.7% result, a notable step up from GPT-4o’s 9.3%, yet it still falls significantly short of o3-mini’s dominant 87.3%.
The development phase for GPT-4.5 demanded substantial innovation, involving new inference systems and low-precision training methods for optimal GPU application.
Training was concurrently spread over multiple data centers, a strategy to effectively manage the model's sizable resource consumption.
Presently available to Pro users for a monthly fee of $200, with a rollout to $20 Plus users planned for the following week.
Compared to Claude 3.7 Sonnet, GPT-4.5's pricing is tenfold, potentially limiting access for smaller developers and nascent enterprises.
