OpenAI has announced a new model called Tolan. In short, it is a system that can handle very large volumes of text — up to 2 million tokens in a single request — while performing better at tasks requiring action planning, reasoning, and coding.
What the Model Is and Why It's Needed
Tolan is not just another GPT with an increased context window. The model is designed to efficiently process long documents, codebases, and research materials — everything that previously had to be split into parts or compressed.
A context window of 2 million tokens is roughly equivalent to one and a half million words. For comparison, this is the equivalent of several medium-sized novels or extensive technical documentation. Such a volume allows the model to see the «big picture» without losing details during processing.
Besides having a long context, Tolan is designed for planning tasks and structured thinking. OpenAI claims that the model shows noticeable progress in its ability to build multi-step strategies, break down complex tasks into stages, and maintain a logical chain throughout a long dialogue or workflow.
Improvements in Code and Reasoning
One of the key directions of Tolan's development is working with code. The model was trained with an emphasis on understanding program structure, debugging, refactoring, and code generation in conditions where the large context of a project needs to be considered.
Simply put, if you upload an entire repository into the model, it can understand the connections between modules, see dependencies, and suggest changes that take into account the architecture of the whole project, not just a single file.
The model's reasoning ability has also been improved. OpenAI notes that Tolan handles tasks better where sequential logical thinking, finding contradictions, comparing alternatives, and justifying conclusions are required. This is especially useful in scenarios involving data analysis, legal documents, or scientific research.
How It Works Technically
OpenAI does not disclose all the architectural details, but it is known that Tolan is built on an improved version of the transformer architecture with an optimized attention mechanism. This allows the model to process long sequences efficiently without a critical increase in computational costs.
Model training included both traditional pre-training methods on large text corpora and specialized fine-tuning stages with an emphasis on planning, coding, and long-document tasks. Reinforcement learning techniques were used to improve the model's ability to select optimal task-solving strategies.
Who Will Find This Useful
The model is geared towards professional users and developers. Primarily, these are:
- Software developers working with large codebases
- Researchers analyzing large volumes of scientific publications
- Lawyers and analysts processing complex contracts and regulatory documents
- Teams engaged in strategic planning and long-term forecasting
In each of these scenarios, the key role is played by the model's ability not just to read a large volume of information but to extract meaning, connections, and structure from it.
Availability and Limitations
For now, Tolan is available via the OpenAI API. The company reports no plans to open-source the model or provide it for local use. Given the size of the context window and computational requirements, this is not surprising — running such a model requires significant resources.
It is worth noting that working with 2 million tokens in a single request is expensive in terms of API costs. OpenAI has not yet disclosed pricing details, but it is likely that utilizing the model to its full capacity will be accessible mainly to corporate clients.
Questions also remain regarding the model's accuracy when working with maximum context. It is known that even modern systems with long windows have problems retaining information from the middle of the context — the so-called «lost in the middle» effect. OpenAI claims that Tolan handles this better, but real-world tests will show how true this is in practice.
What's Next
The release of Tolan is part of OpenAI's strategy to develop models capable of solving complex, multi-stage tasks. The company continues to move towards systems that don't just generate text but can act as full-fledged assistants in professional work.
Interestingly, work on improving the models' ability for planning and structured thinking is proceeding in parallel with increasing the context window. This suggests that the industry is gradually shifting its focus from «more context» to «better reasoning».
It is too early to say whether Tolan will become a breakthrough or just another step in the evolution of large language models. But the fact that OpenAI continues to actively develop the capabilities of its systems toward practical application is undeniable.