In Austin, the announcement arrived as a post: Tesla had begun running some robotaxis without a safety monitor inside the car. The sentence was short. The caveats took longer.
Only a few vehicles were described as operating that way. Other cars in the fleet still had safety monitors. California, where Tesla also wants to operate ride-hailing service, had not granted the company permission to run driverless tests or carry passengers without a human ready to steer or brake. Tesla shares rose that day anyway.
This is the self-driving market in 2026. It exists. It disappoints simple language.
Waymo, owned by Alphabet, is now the company most closely associated with working robotaxis in the United States. Baidu's Apollo Go has delivered millions of fully driverless rides in China. Zoox, Amazon's autonomous vehicle unit, has a strange little carriage of a vehicle and an Uber partnership. Mercedes-Benz sells a Level 3 system that allows the driver to look away from the road under narrow freeway conditions. Tesla sells Full Self-Driving (Supervised), a name that still requires the parenthetical.
The public was promised a cliff. Instead, it got a staircase.
The empty seat
Tesla's achievement in Austin matters because the company has spent years insisting that autonomy would emerge from the same camera-heavy consumer fleet it already had on the road. It has also missed its own timelines. In 2019, Elon Musk said he was very confident that Tesla vehicles would be able to function as robotaxis by the end of the following year. By January 2026, the company was still describing its Austin service as a limited fleet, with some vehicles running without in-car monitors and others still supervised.
The distinction is not semantic. A robotaxi pilot in a defined service area is one kind of claim. A consumer car that can drive itself anywhere, under any condition, with no human fallback is another. No automaker is selling the second version to the public.
Tesla's own support page for FSD (Supervised) says the system can make lane changes, navigate intersections, steer through turns, and drive on many road types. It also tells the owner to remain attentive. The company says the current features require active driver supervision and do not make the vehicle autonomous.
That has not stopped Tesla from building a business around the software. On its first-quarter 2026 call, Tesla said FSD adoption had reached nearly 1.3 million paid customers globally, with most growth coming from subscriptions. The same call placed FSD at the center of the company's sales strategy: the vehicle as a delivery mechanism for the software.
China complicates the story. In May, Tesla confirmed that FSD (Supervised) was available in China after years of regulatory delays. Local rivals had not waited. Xpeng, Xiaomi, Baidu, Pony.ai and WeRide had all pushed driver-assistance or robotaxi systems into a market where municipal approvals, state planning and intense EV competition move together. Chinese regulators have treated Tesla's supervised system as Level 2 automation. The name is grand. The legal category is not.
A fleet that acts like a transit system
Waymo's scale is now large enough to look mundane in some neighborhoods. That may be the most meaningful marker yet. Reporting in March said the company was providing 500,000 paid rides a week across 10 U.S. cities. By June, Waymo was described as operating in 11 U.S. markets, with plans for London and Tokyo. Its own technical writing says its vehicles have driven nearly 200 million fully autonomous miles.
The company's newest vehicle, Ojai, shows where the industry has moved. It is a pale-blue autonomous van based on a Zeekr platform from China's Geely. Waymo adds the autonomous system in Arizona. The vehicle carries 13 cameras, six radar units and four lidar sensors. It still has a steering wheel.
That last detail says something. Waymo has long argued that autonomous driving needs redundancy: cameras, radar, lidar, custom silicon, audio receivers for sirens, detailed maps and a cautious expansion process. Tesla has argued that vision and scale can do the job more cheaply. The split is technical, but it is also commercial. Waymo is trying to build a reliable driver. Tesla is trying to convert millions of privately owned cars into an autonomy network.
Waymo's rollout of Ojai also revealed a slower administrative reality. In California, the first rides were free while the company waited for permission to charge fares in the new vehicles. Regulators asked about minors riding alone, natural disasters, major disruptions and power outages. A robotaxi does not merely drive. It becomes part of a city's failure plan.
The software recall era
The closer robotaxis get to routine service, the less their failures look like science fiction. They look like municipal paperwork.
In June, Waymo recalled almost 3,900 U.S. robotaxis after some entered closed freeway construction zones. The notice followed 13 known incidents in Phoenix and the San Francisco area. The company restricted freeway operations while working on a software remedy. A month earlier, Waymo had filed another voluntary recall after vehicles drove into flooded areas or standing water.
There were other troubles: failures around school buses, performance during San Francisco power outages, vehicles halted in traffic. None of this means the technology has failed. It means the industry has moved into the phase where unusual events are not lab problems. They are public records.
Baidu offers the same lesson at larger Chinese scale. Apollo Go delivered 3.2 million fully driverless operational rides in the first quarter of 2026, according to Baidu's investor release. Weekly rides peaked above 350,000 in March. As of April, cumulative public rides had exceeded 22 million; by May, Apollo Go's global footprint reached 27 cities. Baidu said the fleet had logged more than 330 million autonomous kilometers, including more than 220 million fully driverless kilometers.
Then came Wuhan. In April, CNBC reported that Apollo Go robotaxis had stopped mid-traffic, trapping passengers and contributing to at least one highway collision, according to social-media videos and a Wuhan traffic police statement. The preliminary finding was system malfunction. Baidu did not immediately comment for that report.
Scale does not erase edge cases. It manufactures them.
The middle level no one can sell in a slogan
The Society of Automotive Engineers still provides the cleanest taxonomy, even if the consumer market keeps trying to smudge it. Levels 0 through 2 are driver-support systems. The human remains responsible for the driving task. Level 3 is conditional automation: the system drives under specific conditions, but the person must be ready to take over when asked. Level 4 is high automation inside a defined operating domain. Level 5 is the car that can go anywhere a human can, in any condition.
Level 5 remains absent.
Mercedes-Benz Drive Pilot is the rare U.S. Level 3 example. It is certified for major freeways in California and parts of Nevada, but only when the conditions line up: clear lane markings, approved freeway segments, heavy traffic, speeds under 40 mph, daylight, clear weather, no construction zone and a driver visible to the in-car camera. The system can let the driver look away. It does not let the driver disappear.
That is the strange middle. Level 3 asks a machine to drive until it cannot, then asks a human to return on command. Level 4 robotaxis avoid that handoff by limiting geography. Tesla's consumer FSD avoids it by never legally taking full responsibility from the driver.
The chip behind the promise
The original case for Tesla's autonomy was that the company had the data. The newer case adds silicon.
Tesla's current AI4 hardware runs the software in newer vehicles. On Tesla's Q1 2026 call, Musk said the company had taped out AI5, calling it an edge inference chip and tying it to Optimus, data-center use and the broader autonomy program. The company also discussed a research chip fab at Giga Texas, a possible multibillion-dollar project meant to speed iteration across logic, memory, masks and packaging.
The hardware discussion also exposed a harder fact for older owners. Asked how Hardware 3 vehicles would reach unsupervised FSD, Musk said they did not have the capability. He pointed to memory bandwidth, saying Hardware 3 had only one-eighth of Hardware 4's. Tesla said it would offer upgrade paths or discounted trade-ins for FSD buyers whose cars cannot reach the new target.
For years, autonomy was sold as software waiting to arrive. In 2026, it is also cameras, memory bandwidth, chip yields, service-center throughput, regional validation, insurance terms and permissions from local regulators. The car has become the smallest part of the argument.
The shape of the market
Waymo wants a driver it can place into different vehicles and cities. Tesla wants its fleet to learn faster than competitors can map. Baidu is leaning on Chinese city scale and global partnerships. Zoox is betting on a purpose-built vehicle, with no steering wheel or pedals, designed around ride-hailing rather than private ownership. Uber is placing itself between riders and whoever can supply the autonomous trip.
That may be where the market settles for a while: robotaxis in constrained areas, advanced driver support in privately owned cars, Level 3 as a narrow luxury feature, and Level 5 as a phrase more useful in arguments than in procurement.
The public question is no longer whether a car can drive itself somewhere. It is where, under what rules, with which hardware, on whose insurance, and who answers when it stops.
Tesla's brand name remains the loudest phrase in the sector. Its disclaimer remains shorter: It does not make your vehicle autonomous.
