Tesla’s autonomy proposition has two very different engineering and commercial realities. Full Self-Driving (FSD) Supervised is an expanding, revenue-bearing driver-assistance product; Cybercab is an intended driverless service whose ability to scale rests on regulatory compliance and demonstrated operation. Treating them as one achievement obscures the central fact: FSD remains SAE Level 2 partial automation, for which the driver must monitor the road and be ready to intervene. 2,11 Tesla’s own name for the product—“Full Self-Driving (Supervised)”—states the limitation plainly. 23
That boundary matters because the company’s autonomy, AI and robotics narrative is increasingly material to how Tesla is viewed, while the legal allocation of responsibility remains anchored in the human driver for its consumer system. 22,15 The proof is in the performance, not the promise. A broad supervised capability can generate subscription revenue and useful real-world experience now; it does not, by itself, establish that a control-free vehicle can operate a commercial transport service without a responsible driver.
FSD’s expanding reach does not remove the supervisory burden
FSD Supervised has genuine product breadth. It is described as supporting point-to-point driving under driver supervision, including city driving, parking and backing out. 26 Tesla is continuing to iterate the software, including the phased rollout of FSD Supervised v14.3.10, wider availability in update 2026.27.300, and v14.2 Lite deployments to some Hardware 3 Model S and Model X vehicles. 13,8,17 The Lite versions are characterized as distilled models for less powerful hardware and less detailed cameras, but associated commentary also reports lower confidence and more variable performance. 38
The operating record explains why continuous supervision remains the safety valve. Users report smooth trips and successful operation in tight spaces, including an instance in which FSD v12 responded after another driver signaled the Tesla to proceed. 3,38 Yet other reports describe interventions, near-misses at roundabouts and traffic lights, strikes involving curbs or objects, parking-garage looping, and persistent difficulty with unusual edge cases. 3,38,24,38 Some users specifically perceive a mismatch between the system’s confidence and its distance judgment, which they link to near-misses. 38
Tesla’s vision-only architecture is central to this trade-off. Tesla uses vision rather than sensor-fusion or sensor-fallback systems. 26 Reported constraints include reduced performance in darkness, sun flare and rain, and weaker sensing in fog and snow than radar- or LiDAR-equipped alternatives. 26 Hyundai, by contrast, says it plans initially for supervised Level 2+/2++ capability and considers LiDAR necessary for Level 3 autonomy. 26 This is not a verdict on either architecture; it is a reminder that a scalable sensor strategy and a redundant safety strategy are not necessarily the same thing.
Safety claims require a measurement system equal to the claim
Tesla reports that supervised FSD 14 has one serious crash every 3.5 million miles with non-professional supervisors in the driver’s seat, while early insurer-reviewed data has been described as indicating lower accident frequency among FSD users. 35,19 These claims are commercially relevant, but they are not independently dispositive. Critics argue that drivers may choose FSD for easier roads and disengage it in difficult conditions, creating selection bias. 28 Tesla’s classification counts FSD as engaged when it was active at any point in the five seconds before a collision; commenters contend that crashes after a takeover or disengagement may then be recorded as manual driving. 27 Publicly available data is therefore insufficient to establish causation in crashes involving Tesla automated systems. 10
The July 2026 crash record shows the distinction between system engagement and fault. An Electrek report characterized July as Tesla’s worst month on record for driver-assist crashes, reporting four fatal Autopilot or FSD crashes in which Tesla data indicated one of the systems had been engaged. 10 But NHTSA’s 2021 Standing General Order requires reporting for Level 2 systems engaged within a specified pre-crash window, so engagement alone does not establish causation. 10 In another reported crash, police did not attribute the event to Autopilot or FSD. 25
Speed compliance is a more immediate and operationally specific concern. A three-day July road test by Johanna.be found the supervised system exceeded limits in 20 km/h zones and on 55% of tested 30 km/h Brussels segments, averaging 44 km/h on segments where it exceeded the limit. 16,1 The group attributed part of this behavior to frequent speed-limit misreading. 29 Germany’s KBA has also flagged the system for exceeding posted limits and characterized the matter as a precedent case for a Level 2 system. 30 Under the present approval model, drivers remain responsible for traffic-law compliance and must intervene when the system errs. 16 That is precisely why a supervised product cannot be treated as evidence of driverless operation.
Cybercab faces a more basic question: can the vehicle lawfully operate as designed?
Cybercab is Tesla’s proposed step from supervised assistance to a robotaxi service, in which Tesla intends to control the vehicle, the autonomous-driving technology and the network. 6,14,37 Its design creates the central regulatory issue. NHTSA is examining the steering-wheel- and pedal-free vehicle and has issued a Special Order requiring Tesla to establish, under oath, compliance with applicable Federal Motor Vehicle Safety Standards. 7,31 The agency has requested information on the claimed SAE automation level, operating locations, operational design domain, certification labels and technical basis for each safety-standard determination. 20
This is not a semantic dispute over whether Cybercab should be called Level 3 or Level 4. NHTSA has also asked whether temporarily attached human controls can enable driving, whether those controls supported certification, and how their subsequent removal complies with the Safety Act’s prohibition on making a safety device inoperative. 20 These software boundaries are the new interlocking signals: a vehicle intended to operate without conventional controls must demonstrate not only that its automated function works, but that the physical vehicle satisfies the legal and safety framework governing its sale and operation.
The Austin record demonstrates activity, but not yet broad driverless-network proof. Tesla began carrying public riders there on September 3 after launching a robotaxi service in June 2025 with safety monitors in vehicles. 12,33 NHTSA opened audit query AQ26002 on September 4 following commercial deployment of a small number of Cybercabs in Austin. 4 Reported sightings of Cybercabs with steering wheels, including in Houston and Washington, D.C., introduce a significant configuration tension; commenters describe these as engineering rather than production versions. 4,18,36 If a control-equipped, supervised test vehicle is necessary for particular operations, it should not be conflated with the control-free production concept under review.
The service also appears bounded by a constrained operational design domain. Cybercabs reportedly cannot use Austin highways including MoPac and US-183 and avoid at-grade railroad crossings. 12 One rider reported a roughly 70-minute trip for a journey expected to take about 10 minutes because of a detour through East Austin. 12 A live service with such restrictions may still be valuable as validation, but it is not yet evidence that the network can reliably offer competitive trip times across ordinary urban infrastructure.
Emergency handling is part of the same certification case. NHTSA has asked whether occupants can move the vehicle through touchscreen controls and under what conditions. 20 The Austin Fire Department has identified the absence of steering-wheel and brake-pedal controls as a concern for emergency handling. 34 Reports further state that first responders must contact Tesla Robotaxi First Responder Support if the touchscreen is unavailable. 5 These reports do not establish an operational failure; they establish that fail-safe recovery and responder access require evidence, not assumption.
Commercial value depends on crossing the responsibility boundary
FSD can be monetized today through subscriptions commonly cited at $100 per month. 21 Its strategic value is real, but its responsible-driver model both enables consumer deployment and limits its conversion into a driverless offering. Tesla has attempted to license FSD without a confirmed third-party agreement, while reports say Ford did not license it because Tesla would not accept crash liability and legacy manufacturers wanted Tesla to bear that risk. 9,32,9
Cybercab offers a larger theoretical opportunity through vehicle manufacturing, a ride-hailing network and software economics, but those revenues remain contingent on reliable operation at scale, regulatory approval, economic manufacturing and rider acceptance. 37 The immediate task is therefore practical rather than promotional: establish a permitted operating domain, validate recovery and emergency procedures, demonstrate sustained unsupervised performance within that domain, and show that the resulting service is commercially useful.
Tesla is making visible progress in supervised automation and has begun to test the boundaries of robotaxi deployment. But safety engineering is what happens between the edge cases. Until the evidence shows that Cybercab can satisfy federal compliance requirements and operate dependably without the human supervision that defines FSD today, the prudent conclusion is that Tesla has a valuable autonomy option—not yet a proven driverless network.