Tesla started rolling out Full Self-Driving (FSD) v15 on Saturday, November 21, 2026. The headline claim is "25x better handling of safety critical cases." Tesla's release notes say the upgraded vision encoder improves the critical intervention rate by 25x. Tesla published the multiplier, not the starting rate, so the claim cannot be checked yet. The details below come from Tesla's X announcement and its release notes as reproduced by Electrek.
This is a software update for cars that still need a human driver. Tesla labels the product FSD (Supervised). What changes is the model underneath, a new shared vision encoder, plus a new on-screen visualization and a list of fixes aimed at long-running complaints about signs, potholes and school zones.
TL;DR: FSD v15 claims and what supports them
| Question | Answer |
|---|---|
| When did it start? | November 21, 2026, to a first group of customers (Tesla announcement via Electrek) |
| When is wide release? | No confirmed date found |
| Headline claim | 25x better critical intervention rate from a new vision encoder (Tesla release notes) |
| Is the baseline published? | No. Only the multiplier is public |
| Which cars first? | Tesla did not say. It has told investors that AI4 (HW4) is enough for v15 |
| New features | Reality View, learned navigation, parking suggestions, speed camera slowdown, Bike Rack Mode |
| Independent data? | None yet for v15 |
What is in the FSD v15 release notes?
According to Electrek, Tesla's announcement came on X and a follow-up from its Tesla Super Intelligence account carried the full notes. The notes say v15 "upgrades every part of the end-to-end driving network." The list includes:
- A shared vision encoder. It is used by FSD, Robotaxi and Optimus, trained on millions of internet videos "for improved generalization and data efficiency," and credited with improving the critical intervention rate by 25x.
- Tesla Augmented Reality. Marketed as Reality View. It replaces separate explicit model signals with one end-to-end visualization that uses less total compute.
- Learned navigation. Route choice uses an end-to-end model trained on route preference data from the fleet. Per Electrek, it covers only the start and end of trips for now, and later v15 releases are to cover the whole route.
- Parking. Parking spot choice uses personalized parking history, and the map shows several options at the destination.
- Driver monitoring. Attention monitoring works more often with hats and sunglasses, and strictness is decreased in higher-speed driver profiles.
- Perception. Better detection of road signs, small animals, debris, chains, potholes and deep standing water.
- Speed handling. Changes for each driver profile, school zones, speed limits and bad weather.
- Extras. Autoshift now uses FSD-based gear predictions. The main announcement adds slowing down for speed cameras and a Bike Rack Mode for hitch accessories.
What does "25x better" actually measure?
The phrase on the announcement is broad: "safety critical cases." The release notes narrow it. The figure is the critical intervention rate, meaning how often a driver has to take over in a safety-relevant moment. A 25x improvement means that rate fell to one twenty-fifth of its previous level, if the claim holds.
Three limits stand out.
- No baseline. A 25x gain on a high rate and a 25x gain on a low rate are very different outcomes. Tesla gave only the multiplier. Electrek notes the same: Tesla "still doesn't share the actual number."
- Narrow scope. The notes attribute the 25x to the vision encoder. That is a component, not the whole driving system.
- Unknown method. Tesla has not said which fleet, software baseline, roads or time window the comparison uses.
Electrek ran a back-of-envelope calculation from its own earlier data: if v14 went a couple of thousand miles between critical interventions, a 25x gain would mean around 50,000 miles. That is Electrek's illustration, not Tesla's number, and it rests on crowdsourced data Tesla does not endorse. Treat it as an example of why the baseline matters, not as a measurement of v15.
Why does the shared encoder with Optimus matter?
The notable engineering claim is the shared vision encoder across three products: customer cars, the Robotaxi fleet and the Optimus humanoid robot. Training one perception model on large amounts of video and reusing it across machines is the same direction taken by robot foundation models more broadly. For Tesla it also means that gains in one product could carry to the others, and that a flaw in the encoder could carry too.
Tesla has framed v15 as a major step for a while. Electrek reports that Elon Musk wrote in April that v15 "will far exceed human levels of safety, even in completely unsupervised and complex situations," and that Tesla told JPMorgan in August that v15 is a "step-change" built on seven "core technologies," with about 40% already running in the Austin Robotaxi fleet. Our posts on the Austin Robotaxi expansion to 435 square miles and the Cybercab Dallas public rides cover how that fleet is used today.
Which weaknesses does v15 target?
Electrek points out that the fix list reads like a list of known problems.
- Signs and speed limits. In a September test, the Brussels nonprofit Johanna.be found that FSD exceeded the limit in 55% of the 30 km/h zones it drove through, averaging 44 km/h, per Electrek. In most of those zones the screen showed 50 km/h.
- Potholes. Electrek says Musk said in August that pothole avoidance was "coming soon," and that FSD still drives straight through them.
- School zones. Electrek reports that the US safety regulator NHTSA is reviewing a petition claiming Robotaxis speed through active school zones, in addition to an October 2025 investigation into FSD traffic-law violations.
If v15 fixes these, the improvement would be visible to ordinary drivers. "Improved detection" is not a number, though, and none of these items comes with a measured before-and-after.
Driver monitoring is going the other way
One change reduces a safeguard. Tesla says it is "decreasing strictness" of driver monitoring in higher-speed driver profiles. Electrek notes this comes while stories about sleeping drivers on FSD keep appearing and Tesla is tightening monitoring in Europe at regulators' request. Our reports on Finland's Traficom approval and Italy's provisional approval show how European regulators are handling supervised driving. A system that asks the driver to supervise depends on monitoring to keep the driver engaged, so this item deserves independent testing.
How does this compare with other self-driving rollouts?
Waymo and Zoox run driverless services under permits and report incidents through regulators. Tesla's customer FSD is a supervised driver-assist product, and v15 does not change that label. For context on how regulators treat driverless fleets, see our coverage of the California DMV freeway ban on Waymo and the Waymo Phoenix incident report. Those are different products with different rules, so a v15 number should not be compared with a Waymo figure without matching definitions.
What will confirm or refute the 25x claim?
Several things would settle it:
- Tesla publishing the underlying rate, in miles per critical intervention, with its definition of a critical intervention.
- Third-party fleet data. Crowdsourced trackers and owner reports will accumulate over the next weeks as the wide release reaches more cars.
- Regulator data. NHTSA reviews of FSD behavior and any crash reports involving v15.
- Independent speed-limit tests, like the Brussels one, repeated on v15.
Earlier FSD versions improved in some ways and regressed in others, as Electrek recounts for v14. A fair reading is that v15's list is promising but unmeasured.
What this means for builders and buyers
For people who build with vision models, the lesson is about claims. A multiplier with no baseline cannot support a safety conclusion, whether it comes from a car company or an AI vendor. Ask for the raw rate, the sample size and the definition of the event being counted.
For Tesla owners, expect a staged rollout: a first group now, no confirmed date for wide release, and no confirmed answer yet on older hardware. If you get v15, keep supervising as you did before, and treat the reduced monitoring strictness as a reason to stay alert, not a license to relax.
FAQ: questions people are searching
When is Tesla FSD v15 coming out? It started rolling out on November 21, 2026 to a first group. We found no confirmed date for wide release.
Is FSD v15 really 25x safer? Tesla claims a 25x better critical intervention rate from the new vision encoder. That is not the same as 25x safer, and no baseline or data is public.
Does v15 run on HW3? Tesla did not say. Outlets report no word on HW3, and Tesla has said AI4 (HW4) is enough.
What is new besides the model? Reality View, parking suggestions, speed camera slowdown and Bike Rack Mode.
Related reading
- Tesla Robotaxi Austin expands to 435 square miles
- Tesla Cybercab Dallas public rides: claimed vs verified
- Finland approves Tesla assisted driving
- Italy gives Tesla assisted driving provisional approval
- California DMV bans Waymo from freeways
- Waymo Phoenix stationary cars and speeding truck crash
- Earlier Electrek analysis: Musk says v15 will be safer than humans, as with every version
Details are accurate as of November 21, 2026. Tesla may change release plans; check Tesla's own channels for updates.
