Just as foreign media reported, OpenAI really did release the GPT-4.1 series models this week, including GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, with performance even stronger than GPT-4o. OpenAI has provided supporting data to back this up, but the special thing this time is that this series is API-only, meaning it’s expected not to be integrated into ChatGPT.
The GPT-4.1 series models are officially unveiled, with performance differences compared to GPT-4o.
To help users quickly understand the performance gap between the GPT-4.1 series and GPT-4o, OpenAI shared this chart:
From the figure, it can be observed that GPT-4.1’s MMLU slightly leads GPT-4o, while maintaining the same latency and being cheaper.
Compared to GPT-4o mini, GPT-4.1 mini shows the biggest improvement, and more importantly, its MMLU performance is still close to GPT-4o—that’s really impressive.
OpenAI also noted that GPT-4.1 mini has made a significant leap in small model performance, even surpassing GPT-4o on many benchmark tests, with latency reduced by nearly half and costs 83% lower.
Although GPT-4 nano’s MMLU performance is still worse than GPT-4o mini, its latency is much lower, making it suitable for low-latency tasks. Combined with a 1 million token context window, it delivers quite impressive performance.
Note: All GPT-4.1 series models can reliably reference the full 1 million token context information.
Looking closely at the various tests, GPT-4.1 leads GPT-4o by a considerable margin, such as SWE-bench Verified with a 55% completion rate, higher than GPT-4o’s 33%:
GPT-4.1 also outperforms GPT-4o in both whole and diff on Aider’s polyglot benchmark, and even GPT-4.1 mini is slightly ahead of GPT-4o:
GPT-4.1 also demonstrates more reliable instruction-following performance, almost on par with o1-high and o3-mini-high:
The official OpenAI introduction shares more. GPT-4.1’s test dataThose who are interested can go over and take a look.
What is the API pricing for the GPT-4.1 series models?
The following are the API costs for the GPT-4.1 series models:
The chart below compares with GPT-4o: GPT-4.1’s input is $2 USD, while GPT-4o’s is $2.5 USD. Output is $8 vs $10 USD. As for GPT-4.1 mini, it goes without saying—it’s way cheaper:
Why is the GPT-4.1 series limited to API use only?
To this, OpenAI only stated: “GPT-4.1 will be available exclusively through the API. In ChatGPT, improvements in instruction following, coding, and intelligence have already been gradually incorporated into the latest version of GPT-4o, and we will continue to integrate more improvements in future versions.”
Although the API is adding the new GPT-4.1 series models, it will also phase out GPT-4.5 Preview, since GPT-4.1 offers better or comparable performance on many key capabilities at significantly lower cost and latency. It is expected to be retired on July 14, 2025.
Meanwhile, ChatGPT will continue to keep GPT‑4.5 Preview (if no new version is released by then), so I guess that’s mainly the reason—just have users use this one.
GPT-4.1 live launch video:
Source: KOCPC Chinese