Model Drop

GPT-4.1

April 14, 2025 2 min read

GPT-4.1, which OpenAI shipped on April 14, 2025, is an API-only model built for coding, following instructions and reading long documents, with a context window of about 1 million tokens.

Model: GPT-4.1 (gpt-4.1)
Maker: OpenAI
Ship date: April 14, 2025
Pricing: $2 / $8 per M tokens
Context: 1M tokens (1,047,576)
Where you can use it:
- The OpenAI API only, not ChatGPT
- Also comes in mini and nano sizes
When to reach for it:
- Coding agents and diff-style file edits
- Long documents or whole codebases
- Skip it for hard math; o3-mini does better
How it compares:
- Replaces GPT-4.5 Preview in the API
- 5x the context of Claude 3.7 Sonnet
What makes it different:
- Follows instructions more literally
- Long-context requests cost nothing extra
Early reactions: Simon Willison called the 1 million token context a really big deal and was most excited by the cheap nano model.
Read more: Introducing GPT-4.1 in the API

The benchmarks

OpenAI ran GPT-4.1 without reasoning and compared it with GPT-4o (2024-11-20), GPT-4.5 and its reasoning models o1 (high) and o3-mini (high); its SWE-bench score skips 23 of 500 tasks (52.1% if those count as failures).

OpenAI's numbers

SWE-bench Verified (real GitHub fixes): 54.6%. +21.4 over GPT-4o (33.2%), +16.6 over GPT-4.5 (38.0%), +5.3 over o3-mini (high) (49.3%).
Aider Polyglot (diff-format code edits): 52.9%. +34.7 over GPT-4o (18.2%), +8.0 over GPT-4.5 (44.9%), −8.8 under o1 (high) (61.7%).
MultiChallenge (multi-turn instructions): 38.3%. +10.5 over GPT-4o (27.8%), −5.5 under GPT-4.5 (43.8%).
IFEval (verifiable instructions): 87.4%. +6.4 over GPT-4o (81.0%), −0.8 under GPT-4.5 (88.2%).
Video-MME (long videos, no subtitles): 72.0%. +6.7 over GPT-4o (65.3%).
OpenAI-MRCR: 2 needle 128k (long-context retrieval): 57.2%. +25.3 over GPT-4o (31.9%), +18.7 over GPT-4.5 (38.5%).
GPQA Diamond (PhD-level science): 66.3%. +20.3 over GPT-4o (46.0%), −3.2 under GPT-4.5 (69.5%).

Independent numbers

Artificial Analysis Intelligence Index (current v4.3.2, AA estimate): 12.7. −2.6 under Claude 3.7 Sonnet (no thinking) (15.3), +3.0 over DeepSeek V3 0324 (9.7), +4.3 over GPT-4o (8.4).
ARC-AGI-1 (abstract visual puzzles): 5.5%. +1.0 over GPT-4o (4.5%), −4.8 under GPT-4.5 (10.3%), −8.1 under Claude 3.7 Sonnet (no thinking) (13.6%).