February 2026 Summaries

Gemini 3.1 Pro has been released in the Windsurf platform, featuring two thinking variants: Low Thinking, which consumes 0.5x credits, and High Thinking, which uses 1x credits. This version is accessible to Pro, Teams, Enterprise, and Trial users, with a limited-time promotional pricing on credit usage for self-serve users. To experience these features, users are encouraged to download the latest version of Windsurf.

Claude Sonnet 4.6 has been launched in Windsurf, offering a special promotional pricing structure for self-serve users, with 2x credits for "no thinking" and 3x credits for "with thinking" modes. The new version is also accessible in Arena Mode's battle groups, inviting users to download the latest Windsurf version to experience Claude Sonnet 4.6.

GLM-5 and Minimax M2.5 are two new models introduced in Windsurf, available for users with limited-time promotional pricing, where GLM-5 is offered at 0.75x credits and Minimax M2.5 at 0.25x credits. Both models can be accessed by all users and are featured in Arena Mode's Frontier Arena Hybrid Arena battle groups, allowing users to test their performance against other frontier models on real-world tasks. Users are encouraged to download the latest version of Windsurf to explore these new additions.

GPT-5.3-Codex-Spark has been made accessible in Windsurf via Arena Mode's Fast Arena Hybrid Arena battle groups, providing a unique deployment not available in the standard model selector. This release allows users to experience the capabilities of GPT-5.3-Codex-Spark exclusively through these specialized battle groups. To explore this feature, users need to download the latest version of Windsurf.

Arena Mode's leaderboard offers a distinct evaluation platform for AI models, differing from traditional web-based arenas by focusing on task distribution and user engagement without penalizing models for completing tasks quickly if they maintain quality. With over 40,000 votes collected weekly, the leaderboard provides statistically significant validations, such as Claude Opus 4.6 outperforming Opus 4.5, alongside surprising results where models like Gemini 3 Flash and Grok Code Fast surpass their pro counterparts. Notably, no model achieves over an 80% win rate, indicating a balanced competition. Arena Mode aims to contribute to the community by continuously updating with new evaluations, such as the upcoming GPT-5.3-Codex launch, to enhance understanding of AI performance dynamics.

Opus 4.6 (fast mode) is now accessible in Windsurf as a research preview, offering up to 2.5 times higher output speeds while maintaining the same intelligence as the standard Opus 4.6. Users can experience limited-time promotional pricing for self-serve access until February 16, with a rate of 10x credits for operations without thinking and 12x credits for those with thinking. To explore the enhanced performance, users are encouraged to download the latest version of Windsurf and try out Opus 4.6 (fast mode).

Opus 4.6 has been released in Windsurf, offering limited-time promotional pricing for self-serve users with a 2x credit rate for non-thinking tasks and 3x credits for tasks requiring thinking. This new version is accessible as one of the models in Arena Mode's Frontier Arena, allowing users to test its performance against other frontier models on real-world tasks. To experience Opus 4.6, users are encouraged to download the latest version of Windsurf.

One year after the initial launch of the predictive code tool Tab, the Windsurf team has introduced Tab 2.0, featuring significant improvements in its underlying model and context engineering pipeline, resulting in a 54% increase in characters per prediction while maintaining acceptance rates. Understanding that users have diverse preferences for autocomplete behavior, Windsurf tailored the tool to allow users to adjust the "aggression" level of predictions, balancing the risk of more extensive code suggestions with the likelihood of acceptance. This customization emerged from extensive testing and user feedback, revealing differing opinions on prediction boldness. By refining the context engineering and data pipeline, the team reduced the system prompt prefix by 76%, enhancing performance and efficiency. Tab 2.0 aims to optimize the amount of code generated per call, prioritizing user efficiency over merely achieving high acceptance rates, and represents a step forward in AI-assisted coding by catering to individual user preferences.