Tesla FSD Safety Record 2026

FSD v13: The End-to-End Revolution

Tesla's FSD v13, which began its phased rollout in late 2025, represents the most significant software architecture change in the system's history. The update introduced a fully end-to-end neural network that processes raw camera inputs and outputs vehicle controls directly, eliminating many of the hand-coded rules that governed previous versions. Tesla CEO Elon Musk described it as a "fundamental rewrite" of the driving stack.

Early reports from FSD v13 users indicated smoother driving behavior, fewer phantom braking events, and improved handling of complex intersections. However, the system also introduced new failure modes — including occasional hesitation at green lights, unexpected lane changes on multi-lane roads, and difficulty with construction zones where lane markings conflict with traffic cones.

By mid-2026, FSD v13.4 had been deployed to approximately 1.2 million vehicles in North America. Tesla reports that the system has driven over 3 billion supervised miles since launch, though independent verification of these figures remains impossible as Tesla does not share raw telemetry data with regulators or researchers.

Unsupervised FSD: Austin Launch and Early Results

In June 2025, Tesla launched its first unsupervised FSD service in Austin, Texas — a milestone that had been promised for years. Using a fleet of approximately 50 Model Y vehicles operating without safety drivers, Tesla began offering rides to employees and select testers in a geofenced area of central Austin. By early 2026, the service expanded to a wider Austin footprint and opened to public riders through the Tesla app.

The Austin deployment is significant because it transitions Tesla from an ADAS provider (where the human driver is responsible) to an ADS operator (where Tesla bears responsibility for driving decisions). This shift has profound implications for liability, insurance, and regulatory classification. Incidents involving unsupervised Tesla vehicles are now reported as ADS events in NHTSA's database — the same category as Waymo and the former Cruise robotaxis.

Early data from Austin shows a mix of results. Tesla's unsupervised vehicles have completed thousands of rides with no fatalities. However, several minor incidents have been reported, including low-speed contacts with parked vehicles and one incident where the vehicle stopped unexpectedly in a travel lane, causing a rear-end collision. The sample size remains too small for statistically significant safety comparisons with Waymo or human drivers.

NHTSA PE25012: Escalation to Engineering Analysis

The most consequential regulatory development for Tesla in 2026 is the escalation of NHTSA investigation PE25012 to an Engineering Analysis — the final step before a potential mandatory recall. Originally opened as a Preliminary Evaluation in 2025, the investigation covers approximately 3.2 million Tesla vehicles equipped with FSD across the Model S, Model 3, Model X, Model Y, and Cybertruck.

The investigation centers on camera visibility failures — situations where Tesla's vision-only system fails to detect objects or hazards due to environmental conditions. NHTSA documented 9 specific crashes where FSD failed to react appropriately in conditions including:

• Direct sun glare — cameras overwhelmed by low-angle sunlight, particularly during sunrise and sunset
• Dense fog — reduced contrast making obstacle detection unreliable
• Dust and debris — particulate matter obscuring camera lenses
• Wet road reflections — specular reflections creating false positive and negative detections

One of the 9 documented crashes was fatal. If NHTSA determines a safety defect exists, it could force the largest recall in autonomous driving history — potentially requiring hardware modifications rather than Tesla's preferred over-the-air software updates.

Tesla has 261investigations related to its automated driving systems in NHTSA's database. The PE25012 escalation is the most serious, and its outcome could set precedent for how all consumer-facing autonomous driving features are regulated.

Vision-Only Architecture: The Ongoing Debate

Tesla remains the only major autonomous driving company relying exclusively on cameras for perception. Since removing radar sensors in 2021 and ultrasonic sensors in 2022, Tesla has bet entirely on its "Tesla Vision" camera system — eight cameras providing 360-degree coverage around the vehicle, processed by custom neural networks running on Tesla's FSD computer.

Every competitor uses multi-sensor fusion. Waymo employs cameras, lidar, and radar. Zoox uses cameras, lidar, and radar. Aurora uses cameras, lidar, and radar. The redundancy means that if one sensor type fails — cameras blinded by sun, for example — others can compensate. Tesla's approach offers no such fallback.

Proponents argue that the human visual system proves cameras are sufficient for driving, and that Tesla's neural networks will eventually surpass human perception. Critics counter that human vision has billions of years of evolutionary optimization, 120-degree horizontal field of view with depth perception, and common sense reasoning that no current AI possesses. The NHTSA PE25012 investigation directly challenges the vision-only thesis by documenting camera-specific failure modes.

Crash Data: 2024–2025 vs. 2026 Trends

Tesla's total incident count stands at 3550 reported crashes, with 57 fatalities and 202 injuries. The overwhelming majority — 3532 — are classified as ADAS incidents where a human driver was present and theoretically supervising the system.

The transition in 2025–2026 is notable: as Tesla deploys unsupervised vehicles in Austin, the ADS incident count (18to date) is expected to grow. This creates a natural experiment — the same underlying technology operating with and without human oversight. If the unsupervised ADS fleet shows lower incident rates per mile than the supervised ADAS fleet, it would support the argument that human "supervision" actually introduces risk through inattention and delayed takeover. If ADS rates are higher, it would validate the need for human backup.

Early indications are mixed. Tesla's Austin fleet has had fewer severe incidents but has also operated in a limited, well-mapped geographic area — a constraint that Waymo critics have long cited as the key difference between controlled robotaxi deployment and real-world consumer driving.

Regulatory Outlook

The regulatory landscape for Tesla FSD in 2026 is the most complex it has ever been. Several developments are converging simultaneously:

• NHTSA PE25012 resolution — Expected findings in late 2026 or early 2027. A mandatory recall could affect 3.2 million vehicles and potentially require hardware additions (radar or lidar).
• Texas AV framework— Texas has enacted relatively permissive AV regulations that enabled Tesla's Austin launch. However, any serious incident could trigger legislative reconsideration.
• California expansion ambitions— Tesla has signaled interest in expanding unsupervised FSD to California, but the state's DMV has indicated it will require a separate permitting process. Given the Cruise precedent, regulatory scrutiny will be intense.
• Federal AV legislation — Congress continues to debate national AV standards. Any federal framework would supersede the current state-by-state patchwork but remains politically stalled.

The outcome of the PE25012 investigation will likely be the single most important factor shaping Tesla FSD's future. A finding of no defect would validate Tesla's approach and potentially accelerate unsupervised deployment. A recall would force a fundamental rethinking of the sensor architecture — and could cost Tesla billions in retrofit expenses.

Key Takeaways

2026 is a pivotal year for Tesla FSD. The system is simultaneously achieving its most ambitious milestone (unsupervised driving) while facing its most serious regulatory challenge (the PE25012 Engineering Analysis). The data will ultimately determine whether Tesla's vision-only, software-first approach can deliver on the promise of safe autonomous driving — or whether the industry consensus around multi-sensor redundancy was right all along.