Tesla Says Cybercab Is Level 4. What Comes Next Is Harder.

Disclosure: I own Tesla stock — not enough to change my life in either direction — and I have driven Full Self-Driving (Supervised) since 2022. Both seem worth stating before an essay that measures Tesla against its competitors.

I have been supervising Full Self-Driving since Version 10, which works out to roughly four years of ordinary driving — to work, to the store, through whatever the day happens to contain. The software has improved dramatically in that time. It has also regressed in ways I did not expect. What has never changed is the arrangement underneath it: the car does the driving, and I am responsible for the car.

Tesla has scheduled Cybercab’s Austin launch ↗ for September 3, 2026 — an invitation-only event ↗, with a livestream for everyone else. Tesla’s first responder plan ↗ describes a two-seat robotaxi that is “not typically equipped with a steering wheel or acceleration and brake pedals.” The people inside are not drivers or supervisors. They are just passengers.

That absence makes Cybercab easier to understand than customer-owned Tesla vehicles, even if the technology underneath both remains difficult to evaluate. A Model Y running FSD (Supervised) can do an impressive amount of driving, but someone still sits behind the wheel, watches the road, and remains responsible for correcting it. Cybercab removes that ambiguity from the cabin. If it encounters something it cannot handle, the answer cannot be that the driver takes over.

A successful ride around a selected route would prove something real but narrow: Cybercab can operate on that route under those conditions. The harder questions begin at the route’s edge. What happens on an unfamiliar road, in worse weather, or when the car meets something Tesla did not prepare for?

What the levels actually measure

The SAE scale ↗ assigns roles in the driving task and its fallback. At Level 0, the human does all the driving; warnings and brief emergency interventions do not count as driving automation. At Levels 1 and 2, software can control part of the drive, but the person must keep watching. Level 2 can steer and control speed at the same time. At Level 3, the system performs the whole driving task under certain conditions, while a human remains available to take over when asked.

Level 4 changes the fallback. Within a defined set of roads and conditions — an operational design domain, or ODD — the system is responsible for driving and for reaching what the standard calls a minimal risk condition when it cannot continue. NHTSA’s guidance ↗ describes an ODD in terms of geography, road type, speed, time of day, weather, and other constraints. The passengers do not need to supervise the system or rescue it at the edge of that domain.

Level 5 has no capability-based operating boundary. The system can drive on all roadways under all conditions in which a human could reasonably drive, without keeping someone ready to take over. A company might still limit service for legal or commercial reasons, but the vehicle would not require those limits in order to drive safely. NHTSA’s description ↗ is unusually plain: Level 4 works in limited service areas, Level 5 works universally.

A ride-hailing company does not need Level 5 to build a useful business. A Level 4 service that safely covers the places and times people actually need may be valuable long before it can navigate every mountain road, unmarked construction detour, blizzard, or flooded street. Level 5 is not “Level 4, but more polished.” It asks the system to work without the boundaries that made Level 4 manageable.

Why Cybercab makes Level 5 feel close

That distinction became the center of an argument on X in the days before the launch. Peter Welinder, a vice president at OpenAI, wrote ↗ that affordable Level 5 self-driving “requires AGI and is still at least 5-10 years away,” and laid out a chain of prerequisites: multimodal AGI first, then real-time inference, then an energy-efficient, low-cost edge chip. The pushback ↗ pointed to Cybercabs without steering wheels or pedals in Austin, and to FSD V14 handling ninety-nine percent of some owners’ miles. The reply drew several hundred thousand views within a day.

Cybercab makes autonomy feel imminent because it looks like the final form. There is no steering wheel, no pedals, and no person sitting behind the controls while software drives. But those absences establish something narrower than Level 5. A vehicle can carry passengers without a driver inside its ODD and still be Level 4. Tesla’s own first responder plan says as much, then lists the weather conditions outside Cybercab’s present capability.

No one can settle a five-to-ten-year forecast today. What is worth noticing is that the two claims are not about the same thing. Welinder’s estimate is barely a date at all; it is a dependency, and the dependency is general intelligence cheap enough to run on a chip in a car. The reply answers with driverless vehicles operating in Austin, which is a claim about Level 4. Both can be true at once. Cybercab existing does not make the estimate absurd, because the estimate was never about Cybercab. (Whether unsupervised Cybercabs were carrying anyone in Austin that week is itself unclear; pre-launch rides were reported to be limited to Tesla employees.)

More vehicles, more rides, and larger service areas would demonstrate that Level 4 is scaling. Level 5 requires the capability boundary itself to disappear. Those are different achievements, and only one of them is on the calendar for September 3.

Four systems, four different boundaries

Tesla’s autonomy products share technology and an ambition, but they assign responsibility differently. Four systems are worth separating: FSD (Supervised) in customer cars, Tesla Robotaxi including Cybercab, Waymo, and Zoox.

The version in my driveway is Full Self-Driving (Supervised). The parenthetical tells you who is still responsible for driving.

Under the SAE definitions, FSD (Supervised) in customer-operated Teslas is Level 2. It steers and controls speed, changes lanes, negotiates intersections, and can travel from a starting point toward a destination. But Tesla’s own instructions say the driver must pay attention, be ready to intervene immediately, and remain responsible for the vehicle. Tesla says directly ↗ that “none of these advanced driver assistance features make your Tesla vehicle fully autonomous or replace you as the driver.” The breadth of what the software attempts does not change who is driving in the legal and practical sense.

Tesla Robotaxi is a different product, not merely the same Level 2 feature with the person removed. Before Cybercab’s launch, Tesla offered rides ↗ in limited areas of Austin, Dallas, Houston, Miami, Orlando, and Tampa, using a fleet of Model Ys.

The front seat in those cars is empty in some cases and not others. Tesla launched the Austin service in June 2025 with a human safety monitor riding in the front passenger seat, began operating some vehicles without one in January 2026 ↗, and has scaled slowly since. Even that transition was less clean than the announcement sounded. Electrek reported that some of the first unsupervised cars were trailed by chase vehicles carrying safety monitors, and Tesla’s Head of AI Ashok Elluswamy described the rollout as a few unsupervised vehicles mixed into a fleet that still had monitors, with the ratio increasing over time.

Scale is the other half of it. Independent trackers have counted the driverless portion of the fleet in the low dozens across the three Texas cities, and the number of Robotaxis operating at any given moment across all markets in roughly the same range. Waymo, by comparison, reported a fleet of 3,000 ↗ in February. That is not a verdict on either company, and Waymo has been at this considerably longer. It is the difference between a service and a demonstration, and it is not something a launch event shows.

Tesla does not leave Cybercab’s classification implicit. The first responder plan calls the vehicle an “SAE Level 4 automated driving system” and says that in Autonomous Mode it is “designed to be capable of performing the entire dynamic driving task without any input from a human driver.” The same document draws the line by hardware: a Cybercab that does have a steering wheel and pedals, it notes, “is typically an engineering or test vehicle, and operates at SAE Level 2 autonomy.” The plan describes an ODD covering public roads from city streets to freeways and rural roads, at any time of day, in light or moderate rain, fog, and snow. It excludes off-road use and says the vehicle has limited capability in flooding, heavy rain, snowy or icy roads, and hurricane-strength winds. Off-road driving sits outside the SAE scale anyway; the weather exclusions are what mark the difference between the Level 4 claim and a Level 5 one.

That claimed ODD is broader than the city service areas available to riders. The first responder plan describes where Tesla says the system is capable of operating; the Robotaxi app determines where Tesla actually offers rides. A service can be commercially geofenced even when its vehicles are capable of driving beyond the fence.

The plan also says that after certain first-responder stops, a Tesla representative may remotely move the vehicle or re-engage Autonomous Mode. It does not explain what “remotely moves” entails. The Automated Vehicle Safety Consortium, an SAE industry program, distinguishes remote assistance ↗ — guidance without direct control of the vehicle — from remote driving. Tesla’s document establishes the Level 4 design claim. It leaves the exact role of remote personnel worth asking about.

Cybercab is camera-only, in contrast with the lidar and radar stacks at Waymo and Zoox.

Waymo describes the Waymo Driver as operating at Level 4 ↗, carrying passengers within defined service areas using cameras, radar, lidar, and detailed maps. When a scene is ambiguous, a vehicle can ask remote staff for context. Waymo says those people neither drive the car remotely nor watch it continuously; they offer “advice which the system can decide to use or reject.” In February 2026, Waymo reported about 70 remote-assistance agents on duty worldwide for that fleet of 3,000. The human support is real. The system remains responsible for the drive.

Zoox has the same shape without using the label as prominently. Like Cybercab, its purpose-built vehicle has no conventional driver controls. It uses cameras, lidar, radar, and long-wave infrared sensors ↗, along with high-definition maps prepared before deployment ↗ in a new city. Its public service remains limited to designated areas and conditions; Zoox says it qualifies its vehicles for light rain ↗, not bad weather generally. A vehicle facing an unusually complicated scene can ask the company’s TeleGuidance staff for high-level assistance, and Zoox is explicit about the limit: those tacticians “do not directly drive the vehicle,” and the robotaxi “remains fully responsible for all driving decisions.” Zoox treats that arrangement as part of the safety case ↗ for its defined operating domain.

FSD (Supervised) attempts a broad variety of roads while leaving responsibility with the driver. Tesla Robotaxi includes both monitored and driverless operation; Waymo and Zoox already carry passengers without a driver inside defined domains. The first approach has breadth with human supervision. The others accept more responsibility inside narrower boundaries. Success at one does not prove the other.

What Level 5 still requires

The remaining work falls into two different categories: responsibility and reach. FSD (Supervised) has not crossed the first line. The driver still watches the road, handles failures, and takes over when needed. Removing that obligation would turn the driver into a passenger — which is to say it would turn my car into a different category of thing.

Cybercab, Waymo, and Zoox have crossed that line within an ODD. Their remaining Level 5 problem is reach. Cybercab would have to handle the heavy rain and snowy or icy roads its plan currently excludes. Waymo would have to move beyond roads and weather it has cleared city by city. Zoox would have to stop depending on advance mapping and testing as a condition of capability, and operate beyond its light-rain qualification. Any geographic, road-class, weather, or time restriction required by the technology would have to disappear. A company could still limit service for legal or commercial reasons; the vehicle could not need those limits in order to drive safely.

Scale is a separate question. More vehicles, more cities, and a larger driverless share would make Tesla Robotaxi a more substantial Level 4 service. Waymo’s fleet already shows that Level 4 can operate at meaningful scale. Neither achievement changes the definition of Level 5. A small fleet can possess Level 5 capability; a vast geofenced fleet is still Level 4.

Remote assistance does not have to vanish. Human dispatchers, customer-support agents, emergency coordinators, and roadside crews are not disqualifying; human-driven taxis need all of them too. A remote person may provide context or suggest a path. The automated system must decide whether to use that advice, retain control of the vehicle, and perform its own fallback. Help can remain. Driving cannot.

What to look out for

Cybercab makes Tesla’s claims more measurable because its design removes the manual controls. A steering wheel can blur the difference between software that usually drives and software that is responsible for driving. I have four years of that blur in my own car.

The useful questions about Austin are operational. How many Cybercabs exist? Where may they go, and which roads are excluded? What weather stops service? Do the rides use highways? How often does a vehicle request remote assistance, and what is the remote person permitted to do? When a car cannot understand a scene, does it find a safe place to stop, wait for advice, or require someone to come retrieve it? And when can an ordinary rider — not an invited guest, not an employee — request one?

Those answers will matter long after the launch event has been forgotten. A smooth ride can be both real and carefully bounded. It does not have to be dismissed as a stunt or accepted as proof that autonomy is solved. No launch can demonstrate the absence of a boundary. It can only show where the boundary sits today, which is worth something: it is the only way to tell, the next time, whether it has moved.