January 2025 Summaries

OpenRouter has introduced a new feature called Reasoning Tokens, which allows users to observe the reasoning steps of models in both chatroom interactions and API requests. By including the parameter `include_reasoning: true` in API requests, users can access the reasoning field in the model's response. Initially available for DeepSeek R1 models, this feature will soon support Gemini Thinking models as well. The functionality is primarily intended for transparency in model reasoning but can also be used to enhance less advanced models by integrating reasoning insights. The feature demonstrates potential for innovation in complex workflows, although its practical applications are still being explored.

Reasoning Tokens is a new feature introduced by OpenRouter that provides insight into the reasoning process of models like DeepSeek R1, with plans to extend support to Gemini Thinking models. By adding include_reasoning: true to API requests, users can access the reasoning steps in the reasoning field of each message, offering transparency in model decision-making. This feature not only allows users to inspect the reasoning but can also enhance complex workflows by injecting reasoning into less advanced models to improve their performance. An example demonstrates using Reasoning Tokens to make a basic model smarter by integrating the reasoning output from a more advanced model, showcasing the potential for evolving use cases in AI model interactions.

OpenRouter has implemented standardized finish reasons for model responses to align with OpenAI-style conventions, aiming to simplify the interpretation of responses and enhance debugging and understanding. This includes the introduction of normalized finish_reason values, while also preserving model-specific details through native_finish_reason. The initiative is designed to improve insight into a model's decision-making process and ensure cross-model consistency, thereby streamlining development across reasoning-capable models.

The OpenRouter API has introduced a web search feature that allows developers to integrate real-time web search results into their workflows by appending :online to model slugs or activating a web plugin. Priced at $4 per 1,000 results and defaulting to five results per request, it utilizes Exa.ai to fetch and summarize web data, which is then incorporated into user prompts with markdown-cited sources. This feature is customizable, enabling developers to adjust the number of results and tailor prompts for result integration. It can be activated by appending :online to model slugs or using the web plugin in API requests, with further details available in the documentation or via OpenRouter Chat by selecting the 'web' icon.

OpenRouter has introduced standardized finish reasons to align with OpenAI-style conventions, enhancing the consistency and interpretability of model responses. This update includes both a normalized finish_reason for uniformity across different models and a native_finish_reason to preserve model-specific details. These changes aim to improve debugging and understanding of model behavior, while also facilitating cross-model consistency, thereby streamlining the development process by providing unified finish reasons for reasoning-capable models.

OpenRouter has introduced a web search feature via its API, enabling developers to integrate real-time web search results into their applications by appending :online to the model slug or using a web plugin. This feature, powered by Exa.ai, allows users to fetch up to five web results per request and costs $4 per 1,000 results. With options for customization, developers can adjust the number of results and modify the search prompt to tailor how results are incorporated into responses. The integration can be specified through API requests by using the :online shortcut or directly activating the web plugin, ensuring that up-to-date information is included in workflows.

Starting off 2025 with advancements, OpenRouter has introduced an upgraded Auto Router featuring enhanced capabilities, including routing to 19 different models for optimized quality. The update also introduces a more transparent Chatroom UI that indicates the model being used and is accessible via API. Additionally, OpenRouter is experimenting with routing powered by Not Diamond's meta model, with the API currently available for free, though changes and increased customizability are anticipated in the future.

As of January 1, 2025, the Auto Router has been upgraded to route to 19 different models, enhancing quality by optimizing model selection, and the chatroom interface has been updated to clearly display the model utilized, with these features accessible via API. The introduction of Not Diamond's meta model is being tested to power this routing, and while the API is currently free, future changes may include more customizability options and potential costs.