A new set of API models has been introduced that brings major advances in coding, understanding instructions, and handling extended contexts. These models—comprising a primary version along with two smaller variants—demonstrate significant performance gains over previous releases. They offer improvements in solving coding challenges, reliably following prompts, and efficiently processing context windows that now support up to one million tokens, making them well-suited for complex and large-scale applications.
The enhancements are evident across several dimensions:
- Coding Performance: The flagship model now achieves substantially higher scores on industry benchmarks. For instance, its performance on a real-world software engineering test has improved dramatically compared to earlier iterations, making it a top choice for generating and refining code, from frontend development tasks to producing precise code patches.
- Instruction Following: Across multiple evaluations that test adherence to specific formats, negative instructions, and ordered directives, the new model has marked improvements. It also shows better coherence in multi-turn conversations, reliably integrating context from previous interactions.
- Long Context Understanding: With the unprecedented ability to process contexts as large as one million tokens, the models now perform reliably in tasks that require retrieving subtle details from massive documents, training data, or codebases. Early evaluations demonstrate its accuracy in finding relevant information regardless of its position within a long text.
An accompanying diagram illustrates the relationship between performance and response latency across the model family. This visualization underscores how the improvements have not only boosted task accuracy but also reduced response times and overall costs.
The new offerings include three distinct models:
- Main Model: This version provides advanced coding, nuanced instruction following, and improved long-context comprehension. It is ideal for high-stakes applications like software engineering and multi-document analysis.
- Mini Model: Designed to offer the power of the main model in a smaller package, it achieves comparable or superior performance on intelligence evaluations while significantly reducing latency and cost.
- Nano Model: Engineered for tasks where speed is paramount, this smallest variant is exceptionally cost-effective and swift. Despite its size, it supports extensive contexts and performs impressively on benchmarks measuring classification, autocompletion, and coding assistance.
Real-world applications have already begun to benefit from these improvements. Developers report faster iteration cycles, more precise code changes, and enhanced reliability in applications that require nuanced attention to detail. For example, early testers in domains ranging from legal document review to financial data extraction highlight improvements in both speed and accuracy, translating into smoother workflows and more reliable outcomes.
In addition to performance, the pricing structure has been refined by reducing costs along with enhanced efficiency. The improvements in the inference stack and expanded prompt caching ensure that long-context requests incur no extra expense beyond standard per-token costs. This makes the new models exceptionally attractive for applications that require processing vast amounts of text with minimal delay.
The models are available exclusively via the API, with many of the advancements in instruction following, coding, and long-context comprehension being gradually incorporated into other interfaces. Developers interested in building intelligent agents or systems—such as automated code editors, document analysis tools, or chat-based assistants—will find these tools particularly valuable when combined with features like the Responses API.
A second illustration highlights the model’s demonstrated ability to accurately retrieve critical details from extensive text passages. This “needle in a haystack” performance evaluation confirms that the models can consistently extract relevant information from context windows ranging up to one million tokens.
Overall, these innovative API models represent a significant step forward in making artificial intelligence more capable, efficient, and accessible for real-world applications. Developers are encouraged to explore these capabilities in environments like the Playground and leverage the new tools to build sophisticated and responsive intelligent systems.
Image credit: OpenAI News | OpenAI
