Tesla FSD Version History

FSD 13.2 refines the neural planner with improved urban intersection handling and smoother highway merges. Fleet-wide miles between interventions continue to trend upward.

• Improved handling of complex multi-lane intersections
• Better merge behavior on highway on-ramps and off-ramps
• Reduced phantom braking incidents in urban environments
• + 2 more features

FSD 13.1 is an incremental improvement over the V13 architecture, focusing on edge-case handling and reducing intervention rates in dense city traffic.

• Improved lane-change initiation in dense traffic
• Better handling of temporary traffic control devices (cones, flaggers)
• Reduced speed profile oscillation on winding roads
• + 2 more features

FSD V13 is a major architecture update featuring a significantly larger neural network model and expanded training data. Elon Musk described it as '5-6× better' than V12.5.

• New, significantly larger end-to-end neural network model
• Expanded training dataset — billions of miles of fleet data
• Improved handling of rare edge cases and novel traffic scenarios
• + 3 more features

FSD 12.5 is a significant iterative improvement to the V12 neural planner, bringing improved city street navigation and reduced false positives in obstacle detection.

• Improved city street driving — better gap acceptance and yielding
• Reduced phantom braking on highways and expressways
• Better handling of roundabouts and traffic circles
• + 2 more features

FSD 12.4 advances the end-to-end V12 neural planner with targeted improvements in roundabout handling, unprotected turns, and lane selection on multi-lane arterials.

• Better roundabout entry, navigation, and exit logic
• Improved unprotected left turn gap acceptance
• More accurate lane selection on wide multi-lane roads
• + 2 more features

FSD 12.3 is the first wide release of the V12 end-to-end neural architecture to the broad Tesla fleet. V12 eliminates over 300,000 lines of C++ code, replacing explicit rules with a neural network trained end-to-end.

• End-to-end neural network from camera pixels to vehicle controls — no explicit rules
• Elimination of 300,000+ lines of hand-coded C++ decision logic
• Training on millions of human driver video clips for imitation learning
• + 3 more features

FSD 11.4 is a major single-stack release merging the highway and city driving stacks into one unified neural network pipeline. This unified model handles all driving scenarios without mode-switching.

• Unified highway and city driving stack — one neural network for all roads
• Elimination of the mode-switching boundary between highway and city logic
• Improved consistency as vehicles transition between road types
• + 2 more features