TLDR
- OpenAI launches GPT-4.5 with unprecedented API pricing: $75/M input tokens, $150/M output tokens
- New model emphasizes conversational abilities over technical benchmarks
- Release follows launches from competitors Anthropic and xAI
- Mixed benchmark results show improvements over GPT-4o but lag behind o3-mini
- Access starts with $200/month Pro users, expanding to Plus users ($20/month) next week
OpenAI has launched its latest language model, GPT-4.5, setting new records in API pricing while shifting focus toward conversational abilities. The release marks a departure from the industry’s traditional emphasis on technical benchmarks.
The company announced pricing at $75 per million input tokens and $150 per million output tokens for API access. These rates represent a 2900% increase for input and 1300% increase for output compared to GPT-4o.
The launch adds to a busy month in AI development. Anthropic released Claude 3.7 Sonnet yesterday, while xAI’s Grok-3 arrived last week.
OpenAI CEO Sam Altman introduced the model with candid remarks about its resource demands. “Bad news: It is a giant, expensive model,” he stated during the announcement.
The model introduces what OpenAI calls “vibes,” focusing on emotional intelligence and natural conversation flow. This represents a new direction in AI development.
GPT-4.5’s development required building new inference systems. The team used low-precision training across multiple data centers to handle the model’s demands.
The presentation included everyday examples of the model’s capabilities. One demonstration showed GPT-4.5 responding to a user frustrated about canceled plans.
OpenAI created a specialized “Vibes test set” to measure the model’s performance in creative intelligence and conversation quality.
Under the Hood: Technical Performance and Benchmarks
The model’s benchmark results tell an interesting story. GPT-4.5 achieved 71.4% on the GPQA science evaluation, improving from GPT-4o’s 53.6%.
These scores remain below OpenAI’s o3-mini model, which reached 79.7% through its reasoning capabilities.
Mathematics testing showed similar patterns. GPT-4.5 scored 36.7% on AIME ’24, beating GPT-4o’s 9.3% but falling short of o3-mini’s 87.3%.
The model performed better in coding evaluations. It outpaced both its predecessor and o3-mini on the SWE-Lancer Diamond benchmark.
However, GPT-4.5 lagged behind on SWE-Bench Verified compared to reasoning-focused models.
Altman described GPT-4.5 as “the first model that feels like talking to a thoughtful person.” He reported having moments of amazement at receiving “actually good advice from an AI.”
The pricing structure puts GPT-4.5 at ten times the cost of competitor Claude 3.7 Sonnet. This may limit access for smaller developers and startups.
OpenAI explained their dual approach to advancement: unsupervised learning and reasoning. GPT-4.5 emphasizes the former, aiming to increase “word model accuracy and intuition.”
Access to GPT-4.5 starts with Pro users paying $200 monthly. Plus users at $20 monthly will gain access next week.
Stay Ahead of the Market with Benzinga Pro!
Want to trade like a pro? Benzinga Pro gives you the edge you need in today's fast-paced markets. Get real-time news, exclusive insights, and powerful tools trusted by professional traders:
- Breaking market-moving stories before they hit mainstream media
- Live audio squawk for hands-free market updates
- Advanced stock scanner to spot promising trades
- Expert trade ideas and on-demand support
