September 12, 2025 — Cursor has released a new version of its Tab model, the autocomplete system that predicts your next coding action. The upgraded model delivers 21% fewer suggestions while achieving a 28% higher accept rate, meaning developers now see more relevant completions and fewer distractions.

What’s New in Cursor Tab

Cursor Tab is triggered every time a developer types a character or moves the cursor, running on more than 400 million requests per day. In the past, one challenge was the number of “noisy” suggestions — predictions that didn’t align with what the developer wanted. These unnecessary completions slowed down workflow and reduced trust in the system.

The new Tab model tackles this problem using online reinforcement learning (RL):

• On-policy training: Instead of training on static datasets, Cursor frequently rolls out new model checkpoints, gathers live acceptance/rejection feedback, and retrains based on real user interactions.

• Policy gradient optimization: The model is optimized to maximize a “reward” score, encouraging suggestions that are likely to be accepted and penalizing those that are not.

• Smarter decision-making: Beyond predicting code, the model now decides when not to show a suggestion — reducing noise while improving flow.

Why This Matters for Developers

• Cleaner experience: Developers see fewer irrelevant completions.

• Higher trust: When a suggestion appears, it’s more likely to be correct.

• Improved productivity: More accepted completions mean less time wasted pressing Escape or undoing incorrect code.

Industry Context

Unlike other AI coding tools that rely heavily on static datasets or paid labelers and update only during major releases, Cursor’s approach is dynamic and continuous. New models are deployed several times a day, with fresh training loops completing in just 1.5 to 2 hours — much faster than industry norms.

The Bottom Line

Cursor’s new Tab model, powered by reinforcement learning, is now the default for all users.