Self-driving cars in 2026: where do we really stand?

Here is the critical point to understand: Tesla FSD is officially classified as Level 2, that is, driver assistance, not autonomy. Tesla often communicates as though it were more, but the regulatory classification is unambiguous. Waymo, by contrast, genuinely operates at Level 4: the car has no human driver on board and no one is supervising.

Waymo: the market's silent leader

Waymo is probably the most under-reported AI success story of 2026. Alphabet's autonomy subsidiary (founded in 2009, spun out of the Google X programme in 2016) now operates commercially in 5 American cities: Phoenix, San Francisco, Los Angeles, Austin, and, in early 2026, Atlanta. The strategic goal: to reach 3,500 vehicles in the fleet by the end of 2026, with expansion planned into 10 new cities (Dallas, Denver, Detroit, Houston, Las Vegas, Miami, Nashville, Orlando, San Antonio and Tampa).

The figures documented by Waymo itself:

• More than 80 million kilometres driven without a driver, mainly in Phoenix
• 83% fewer accidents involving airbag deployment compared with human drivers in Phoenix and San Francisco
• 81% fewer accidents involving injuries
• 64% fewer accidents reported to the police

These figures come from Waymo, so they should be treated with methodological caution. But they are corroborated by several independent studies and by the regulatory authorities. The conclusion is clear: Waymo is statistically safer than an average human driver in the areas where it operates.

Tesla: the opposite bet and its limits

Tesla did the exact opposite. Musk's strategy is to rely on cameras alone (without lidar, unlike Waymo) and to scale through the quantity of data. Tesla has more than 4 million vehicles on the roads equipped with Autopilot or FSD Beta, which send their data back to the company. That is more real-world data than any other player in the world.

This approach has two genuine advantages. First, it is cheaper per vehicle (no $50,000 lidar). Second, it benefits from a massive volume of training data. But it also has hard drawbacks: cameras alone fail in several conditions (heavy rain, low sun, fog) where lidar works perfectly.

The safety figures published by Tesla in November 2025 deserve close attention.

• Tesla claims that drivers on FSD cover around 5 million miles between major collisions (with airbag deployment) in North America.
• And 1.5 million miles between minor collisions.
• By comparison, the NHTSA (National Highway Traffic Safety Administration, the US federal road safety agency) states that across all drivers, 505,000 miles are driven per major collision and 178,000 per minor collision.

On paper, Tesla is therefore 10 times safer than average. Except that this comparison is misleading. Drivers using FSD do so mainly on motorways (where accident statistics are naturally better), and with sustained attention since supervision is mandatory. The fair comparison would be with an attentive human driver on the same types of road, and there the gap is much smaller. The raw figures do not lie, but they do not tell the whole story.

Worse, since the launch of the robotaxi programme in Austin in June 2025, Tesla has already reported 7 collisions involving its 2026 Model Y to the NHTSA up to 15 October 2025. These vehicles were equipped with the new ADS (Automated Driving System) which is not yet widely deployed. No serious accidents, but the warning signal is there.

The death of Cruise: a lesson for the whole industry

One of the major events in the sector since 2024 is the effective disappearance of Cruise, the autonomous driving subsidiary of General Motors (a historic American carmaker). Cruise had reached more than 10 million driverless miles before 2024, operated in San Francisco and Phoenix, and was considered one of the three global leaders alongside Waymo and Tesla.

Everything collapsed in October 2023, when a Cruise vehicle dragged a female pedestrian 6 metres after an accident in San Francisco. The company initially concealed the severity of the incident from Californian regulators. This deception led to the suspension of its California permit, mass departures of executives, and ultimately to the announcement in December 2024 that General Motors was permanently shutting down Cruise as a commercial service in order to integrate the technology into its own vehicles.

The Cruise lesson is instructive for understanding the sector. As we explored in our article on major AI failures, transparency with regulators has become an absolute imperative. Concealing an accident can destroy in a matter of days a company that cost 10 billion dollars to build.

The other players that really matter

Beyond the three historic leaders, several players are gaining ground in 2026.

Mercedes-Benz Drive Pilot. The German carmaker is the only one to have obtained Level 3 approval in the United States (in addition to Germany). In practice, on certain motorways in California and Nevada, the driver of a Mercedes Drive Pilot can look away from the road and do something else (watch a film, read). This is a real advance, but limited to 95 km/h and to very specific conditions. No one else has obtained equivalent approval.

Zoox (Amazon). Amazon's autonomous driving subsidiary, which is developing a vehicle designed specifically for robotaxi use (no steering wheel, symmetrical design). First deployments limited to Las Vegas and San Francisco in 2025. The company has an interesting "from scratch" approach but remains in the pilot phase.

Pony.ai, WeRide, Baidu Apollo (China). Chinese players now operate large-scale robotaxi fleets in several Chinese cities (Beijing, Shanghai, Guangzhou, Shenzhen). Baidu Apollo Go claims more than 7 million commercial rides completed. China probably has the densest robotaxi fleet in the world in 2026, even if the Western press says little about it.

Wayve (United Kingdom). A London startup developing an "end-to-end" AI learning approach, without prior HD mapping. Raised 1 billion dollars in 2024, partnerships with Nissan and Microsoft. A promising approach for markets where Waymo's millimetre-accurate mapping is not economically viable.

Ford BlueCruise and GM Super Cruise. The two main Level 2 hands-free systems in North America, deployed on millions of consumer vehicles. Not real autonomy, but genuine driver assistance on motorways. Ford has announced a move to Level 3 targeted for 2028, which looks modest compared with Tesla's promise.

Why robotaxis still cost more than an Uber

A point often glossed over: Waymo robotaxis are currently more expensive than an Uber or Lyft driven by a human. According to data from Obi (a price comparison platform for ride-hailing), a Waymo ride in San Francisco or Phoenix generally costs 15 to 30% more than an equivalent Uber.

This pricing is explained by the enormous costs of the fleet. A Waymo vehicle costs around 200,000 dollars to equip (Velodyne lidar, multiple cameras, onboard computers, redundant telecommunications). Plus the operating cost (teleoperators who can intervene remotely if there is a problem, maintenance teams, specific insurance). Over a vehicle's amortisation period, these costs remain higher than the salary of a human driver in most current configurations.

The economics should gradually reverse with scaling. At 3,500 vehicles by the end of 2026, then 10,000 and more in 2027-2028, unit costs will fall through economies of scale. It is from that point that robotaxis could become 30 to 50% cheaper than a human ride-hailing service, transforming the economics of urban mobility.

What remains genuinely difficult

Despite the progress, several scenarios remain hard even for Waymo, the best player.

Extreme weather conditions. Snowstorms, torrential rain, thick fog: sensors (lidar and cameras) can be overwhelmed or blinded. Waymo still operates mainly in cities with mild climates (Phoenix, Austin, LA). Expansion to Detroit, Denver and Nashville in 2026 will test its ability to handle harsh weather.

Construction zones and temporary roadworks. Roads change. A closed lane, a police officer directing traffic, a diversion marked by orange cones: these situations remain extremely difficult for AI. Waymo engineers call them "the long tails" (the rare but numerous exceptions taken together) that still require frequent human intervention.

Unpredictable behaviour from other road users. Cyclists who change direction without signalling, pedestrians who cross outside crossings, aggressive drivers who ignore the rules: AI handles poorly whatever is not in its model. Humans adapt intuitively, AI reacts more slowly.

The Morocco, India and Vietnam test. Cities where driving is "organic" (few traffic lights, lots of motorbikes, constant negotiation between vehicles) are currently out of reach for every autonomous approach. That is probably where the most significant progress remains to be made.

What impact for you in France and Europe?

Europe is lagging behind on the subject, unsurprisingly. Driverless robotaxi deployments are authorised in France only on a very restricted basis (pilot programmes in Paris during the 2024 Olympics, Lyon and Bordeaux for limited trials). No commercial service operates in France yet. Mercedes Drive Pilot Level 3 is available in Germany but not yet commercially in France.

Several European players are developing interesting approaches:

• Stellantis (the Franco-Italian car group created by the 2021 merger of PSA and Fiat Chrysler) is developing Level 3 systems for its 2026-2027 premium vehicles.
• Renault and its partnership with Wayve mentioned above.
• Volkswagen developed its own system with MobilEye (an Intel subsidiary based in Israel, a world leader in chips for autonomous vehicles), but scaled back its ambitions in 2025 after costs spiralled.

For a French person in 2026, here is what you will actually see:

• No commercial robotaxis in French cities before 2028 at the earliest.
• Level 2 and soon Level 3 systems on motorways, mainly from Mercedes, BMW and Audi, and gradually from Stellantis.
• Autonomous shuttles at some airports and campuses, but at reduced speed and on fixed routes.

The honest verdict for 2026

After a decade of promises, here is the factual assessment.

The autonomous car really does exist. Waymo proves it every day in 5 American cities. The service works, it is used by hundreds of thousands of people, and it is statistically safer than a human driver.

But it is geographically limited. The dream of an L5 car that drives everywhere, in all conditions, like a human, has still not been realised. And it probably will not be before 2030, or even 2035.

Tesla has probably promised too much for too long. FSD remains a Level 2 system despite the marketing. The Tesla robotaxi exists but with 30 vehicles, far from the millions promised. As we explored in our article on faked AI demos, the gap between communication and reality has become a problem for the company's credibility.

China could surprise. With Baidu, Pony.ai and WeRide operating at large scale, China is probably further ahead in real-world deployment than the Western press reports. One to watch in 2026-2027.

Europe is lagging behind. No European champion features in the global top 5. As we explored in our article on France's lag, the autonomous car is another field where Europe risks becoming a customer rather than a producer of the technology that will transform its citizens' mobility.

The autonomous mobility revolution is now real, but it is unfolding differently from what we imagined in 2015. Slower. More geographical. More dependent on partnerships with cities. To follow what comes next, you can also read our article on the environmental cost of AI, which touches on the subject too, since autonomous fleets consume enormous amounts of energy and bandwidth, and our analysis of AI regulation, which explores similar regulatory challenges.