January 2026 Summaries

Arenas are a form of online evaluation for models that aim to provide a more accurate reflection of real-life use cases compared to traditional benchmarks, offering a single, clear metric that updates quickly with new model releases. However, they face challenges such as potential bias, lack of context, and issues with style and task mismatch. To address these, the introduction of Arena Mode in Windsurf allows users to evaluate models directly within their workflow, promoting personalized and contextually relevant assessments. This mode supports multiple models to generate responses, enhancing the evaluation process through self-consistency and allowing users to act as code reviewers, selecting the best model output. Arena Mode also includes personal and global leaderboards and offers free access to all battle groups for a week, with plans to expand features in the future. Additionally, Plan Mode has been introduced, providing further functionalities within the Windsurf platform.

GPT-5.2-Codex has been released in Windsurf, providing users with varying levels of reasoning effort—low, medium, high, and extra high—along with temporary discounts on credit usage, which include 0.5x, 1x, and 2x credits. This version of Codex is particularly adept at handling extensive codebases in prolonged sessions. For optimal results in most tasks, the medium reasoning effort is recommended. Users are encouraged to download the latest version of Windsurf to explore these new capabilities, and further details can be found in OpenAI's blog post.