April 15 News.OpenAI The company published a blog post today (April 15) announcing the release of the API as a GPT-4.1 series of models.Covers GPT-4.1, GPT-4.1 mini and GPT-4.1 nano.
These models comprehensively outperform their predecessors, GPT-4o and GPT-4o mini, in programming, instruction adherence, and long text understanding, with a context window of up to 1 million tokens and knowledge updated through June 2024.
It is important to note that this series of models is now exclusively for developers, and is only available through the developer API, so ordinary users cannot experience the models through the ChatGPT page for the time being.
OpenAI says that on the programming side, the GPT-4.1 model's code generation speed soared by 401 TP3T compared to the GPT-4o model, and the cost of user-entered queries was reduced by 801 TP3T.
New Model Performance
In an official blog post, OpenAI said that the GPT-4.1 family of models outperforms GPT-4o and GPT-4o mini across the board in terms of programming, instruction adherence, and long text processing.
GPT-4.1 scored 54.6% in the Programming test SWE-bench Verified, an improvement of 21.4 percentage points over GPT-4o, 10.5 percentage points in the Command Compliance test MultiChallenge, and a new record of 72.0% in the Multimodal Long Text test Video-MME.
The GPT-4.1 mini and nano demonstrate the great potential of smaller models. gpt-4.1 mini matches or exceeds gpt-4o in several benchmarks, with latency reduced by nearly half and cost reduced by 83%.
The GPT-4.1 nano, the fastest and most cost-effective option with a context window of 1 million tokens and a score of 80.1% in the MMLU test, is suitable for categorization and auto-completion tasks.
These models significantly reduce the first response time by optimizing the inference stack and hint caching techniques, providing developers with an efficient and cost-effective solution.
The GPT-4.1 series of models have outstanding performance in real-world applications and are particularly suitable for building intelligent agents to handle complex tasks. For example, Windsurf tests show that GPT-4.1 improves programming efficiency by 30% and reduces unnecessary edits by 50%; CoCounsel, a legal AI assistant from Thomson Reuters, improves the accuracy rate of multi-document review by 17% after using GPT-4.1.
Naming confusion raises concerns
The release of GPT-4.1 has exacerbated the complexity of naming OpenAI products.
ChatGPT now includes a variety of model options such as GPT-4o, GPT-4o mini, o1-pro, etc. OpenAI CEO Sam Altman acknowledged the naming issue back in 2024 in February.
He said at Platform X that the product line is too unwieldy and that he plans to consolidate the branding with a future GPT-5, and that OpenAI plans to phase out the GPT-4.5 Preview model in the API by July 2025, thereby alleviating naming confusion.
This interim model, launched in February 2024, has been criticized as a "failure" and developers will need to migrate to a different model by July 2025, however, GPT-4.5 remains in ChatGPT for now and is unaffected.
cost
In terms of API price, the OpenAI GPT-4.1 model costs $2 per 1 million tokens input (note: the current exchange rate is about RMB 14.6) and $8 per 1 million tokens output (the current exchange rate is about RMB 58.3). Compared with GPT-4o in medium queries, GPT-4.1 not only provides stronger performance, but is also 26% cheaper.
In addition, the OpenAI GPT-4.1 nano is OpenAI's cheapest and fastest model.