Model Drop

GPT-4.5

February 27, 2025 2 min read

GPT-4.5, which OpenAI shipped on February 27, 2025, is a research preview of its largest chat model, built by scaling up pre-training rather than step-by-step reasoning.

Model: GPT-4.5 (gpt-4.5-preview)
Maker: OpenAI
Ship date: February 27, 2025
Pricing: $75 / $150 per M tokens
Context: 128K tokens
Where you can use it:
- ChatGPT Pro first, then Plus and Team
- The API on all paid usage tiers
When to reach for it:
- Writing, brainstorming and coaching
- Questions that lean on world knowledge
- Skip it for hard math; o3-mini does better
How it compares:
- Far pricier per token than GPT-4o
- No thinking mode, unlike Claude 3.7 Sonnet
What makes it different:
- OpenAI's largest pre-trained model yet
- Tuned for warmer, higher-EQ conversation
Early reactions: Simon Willison called it very expensive and weirdly slow in the API, with coding scores similar to o3-mini.
Read more: Introducing GPT-4.5, OpenAI GPT-4.5 System Card

The benchmarks

GPT-4.5 answers without a reasoning step; OpenAI compared it with GPT-4o and o3-mini (high), plus o1 on SimpleQA, and its SWE-bench and SWE-Lancer figures are its best internal runs. OpenAI labeled the GPQA row only "GPQA (science)"; its later GPT-4.1 post put GPT-4.5 at 69.5% on GPQA Diamond.

OpenAI's numbers

SimpleQA (short factual questions): 62.5%. +24.3 over GPT-4o (38.2%), +15.5 over o1 (47%), +47.5 over o3-mini (15%).
SimpleQA hallucination rate (lower is better): 37.1%. −24.7 under GPT-4o (61.8%), −6.9 under o1 (44%), −43.2 under o3-mini (80.3%).
GPQA (PhD-level science): 71.4%. +17.8 over GPT-4o (53.6%), −8.3 under o3-mini (high) (79.7%).
AIME 2024 (competition math): 36.7%. +27.4 over GPT-4o (9.3%), −50.6 under o3-mini (high) (87.3%).
MMMLU (multilingual knowledge): 85.1%. +3.6 over GPT-4o (81.5%), +4.0 over o3-mini (high) (81.1%).
MMMU (multimodal reasoning): 74.4%. +5.3 over GPT-4o (69.1%).
SWE-Lancer Diamond (paid freelance coding): 32.6%. +9.3 over GPT-4o (23.3%), +21.8 over o3-mini (high) (10.8%).
SWE-bench Verified (real GitHub fixes): 38.0%. +7.3 over GPT-4o (30.7%), −23.0 under o3-mini (high) (61.0%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 9.6. −5.7 under Claude 3.7 Sonnet (no thinking) (15.3), −2.5 under Grok 3 (non-reasoning) (12.1), +1.2 over GPT-4o (8.4).
ARC-AGI-1 (abstract visual puzzles): 10.3%. −3.3 under Claude 3.7 Sonnet (no thinking) (13.6%), +5.8 over GPT-4o (4.5%).
Aider Polyglot (multi-language code edits): 44.9%. −15.5 under Claude 3.7 Sonnet (no thinking) (60.4%), −3.5 under DeepSeek V3 (48.4%), +26.7 over GPT-4o (18.2%).