Snapshot of Tesla Full Self-Driving as of April 2026: architecture (HW3/AI4/AI5), safety data controversy, NHTSA engineering analysis on 3.2M vehicles, CA robotaxi regulatory standoff, and unsupervised rollout in Austin/Dallas/Houston.

Autoresearch: Tesla FSD

Generated by /autoresearch on 2026-04-20. Synthesized from 5 successfully fetched web pages plus search snippets (see Provenance). Treat as raw material — review before promoting. Context: threads/autonomous-driving

Summary

Tesla FSD in April 2026 is two different products depending on how you squint. For consumers it is a supervised Level 2 driver assist whose latest versions (v13/v14) introduce a transformer-based "world model" and a hardware gap between HW3 and AI4 that Tesla has now publicly acknowledged. For Tesla's robotaxi strategy, it is the stack behind a tiny but growing unsupervised fleet — Austin, Dallas, Houston as of mid-April 2026 — that regulators in California explicitly say does not qualify as an autonomous vehicle service. Federal pressure is rising: NHTSA has escalated its reduced-visibility probe to an Engineering Analysis covering ~3.2M vehicles, and independent analysts argue Tesla's published safety numbers are methodologically unfit to support its safety claims.

Findings

1. Architecture and hardware (as of v13, with v14 rolling out)

FSD v13 is described by third-party technical write-ups as a shift from v12's "imitation learning" era into an "end-to-end temporal transformer" / world-model era, processing multiple cameras across time with a ~10-second temporal buffer and a 3D voxel representation (TESMAG deep dive). The driving-logic network reportedly tripled from ~2.3GB (v12) to ~7.5GB (v13), forcing Tesla to run aggressive INT8 quantization to keep HW3 viable (TESMAG v13 paradox).

The HW3/AI4 performance gap is now explicit: community data cited by TESMAG shows AI4 averaging ~450 miles between critical disengagements versus ~120 miles on HW3 (TESMAG HW3 vs AI4). V14 is rolling out on AI4, with a "V14 Lite" planned for HW3 (notateslaapp on v14 Lite).

AI5 (the next-gen inference chip) was pushed to early 2027 per Musk in January 2026, and is expected to carry up to 9× the memory of HW4 (~144GB RAM) (notateslaapp on HW3 end-of-line). (Claims about HW3/AI4 disengagement rates and AI5 specs come from enthusiast/aftermarket blogs, not Tesla — treat as community-sourced rather than primary.)

2. Safety data and the NHTSA posture

NHTSA escalated its reduced-visibility investigation (previously PE24031, opened October 2024) to an Engineering Analysis — EA26002 — in March 2026, covering approximately 3.2 million vehicles. The scope is nine crashes in impaired-visibility conditions with one fatality and one injury, plus six additional incidents under examination (Electrek on EA26002). The specific technical concern is that FSD "did not detect common roadway conditions that impaired camera visibility and/or provide alerts… until immediately before the crash occurred." An EA historically resolves within ~18 months into either closure or a recall mandate.

A parallel probe (PE25012) examines 58+ incidents of traffic violations — running red lights, crossing into opposing lanes — that had grown to ~80 documented violations by December 2025. Tesla requested multiple deadline extensions, citing 8,313 records requiring manual review at ~300/day (Electrek on violation probe delays).

Independent safety researcher Phil Koopman argues Tesla's public safety report is methodologically unfit: it compares new FSD-equipped Teslas against 12+ year-old vehicles (rather than modern ADAS-equipped peers), uses a 5-second reporting window versus NHTSA's 30-second standard, excludes crashes without airbag deployment, and has under-reporting bias for severe crashes because telematics can be disabled by heavy damage. Koopman's conclusion: "The threats to validity are so pervasive… we can conclude nothing useful about the practical safety of FSD from this report." He notes IIHS classifies FSD as a convenience feature, not a safety improvement (Koopman substack).

Third-party community tracker data cited by analysts shows the "city miles to critical disengagement" metric falling from ~4,109 (v14.1 peak, October 2025) to ~809 after v14.2 rolled out — a sharp regression, and roughly 37× worse than Waymo's ~30,000 mile threshold before it pulled safety drivers (Sahm Capital analyst summary). Tesla does not itself publish disengagement data, crash-by-severity, miles by road type, or the denominator behind its crash-rate claim.

3. Regulatory status

California: The CPUC deputy executive director Pat Tsen stated on the record that "Tesla is not operating an autonomous vehicle service" in California. Tesla holds a Charter Party Carrier (TCP) permit — the same category as limousine companies — not an AV permit. California defines autonomous vehicles as SAE Level 3+; Tesla's Level 2 FSD doesn't qualify, and Tesla has not applied for a CPUC or DMV AV permit (Electrek on CPUC statement; EVXL on limo permit). Under the TCP classification Tesla avoids safety data reporting obligations that apply to permitted AV operators like Waymo.

Analysts note Tesla would need to log ≥50,000 supervised autonomous miles before it could even apply for driverless testing in California, a bar it has not publicly moved toward (Automotive World on CA timeline).

Federal: EA26002 plus PE25012 (see §2) put Tesla in an unusually exposed position — three concurrent active NHTSA investigations into FSD.

4. Competitive landscape

Mercedes Drive Pilot is the only Level 3 consumer system generally available in the US, allowing hands-off, eyes-off operation but only in mapped, approved freeway conditions at low speed. It uses a redundant LiDAR + radar + camera + ultrasonic + HD-map stack; Tesla remains the only major OEM pursuing vision-only (Auto2x on Mercedes Level 3 pricing; InsideEVs on FSD vs Drive Pilot). The trade: Tesla operates on arbitrary streets as Level 2; Mercedes operates hands-off but only on a narrow ODD.

Waymo is the operational yardstick — ~500,000 paid, fully driverless rides per week across 10 cities, 15M rides completed in 2025, and ~100 vehicles in Austin alone covering 90 sq mi (Electrek on Austin scale gap). Waymo publishes per-trip data; Tesla (under TCP) does not.

5. Robotaxi / unsupervised rollout status

As of April 18, 2026, Tesla operates unsupervised robotaxi service (no human in front seat, remote supervision only) in three cities: Austin, Dallas, and Houston. Dallas and Houston launched April 18 with ~25 sq mi geofences each, covering parts of Willowbrook/Jersey Village (Houston) and Highland Park / central Dallas (tech-insider on Dallas/Houston launch; Teslarati on two new cities).

In Austin, the unsupervised fleet as of late March 2026 was 4–8 Model Ys, with a total Austin robotaxi fleet of ~42 vehicles — only ~12% operating without a front-seat human (Electrek on Austin handful). Unsupervised here still includes remote human monitoring.

Tesla's Q4 2025 guidance projected H1 2026 expansion to seven US cities including Phoenix, Miami, Orlando, Tampa, and Las Vegas (CNBC on 2025 robotaxi boom).

Contradictions and open questions

Safety claims vs. independent analysis. Musk/Tesla marketing claims "10× safer than humans"; Koopman and sellside analysts both conclude Tesla's public data doesn't support that (and may show regression) (Electrek on Musk's 10× claim; Koopman). Unclear which is closer to truth absent Tesla releasing denominator data.
"Unsupervised" definition. Tesla's Austin/Dallas/Houston service has no driver in the car but includes remote supervision. How much intervention remote operators actually perform — and how this maps to Waymo's own remote-assistance model — isn't publicly disclosed.
HW3 long-term viability. v14 "Lite" is planned for HW3, but with the HW3/AI4 disengagement gap widening, whether HW3 owners receive meaningful parity or whether Tesla offers upgrades/retrofits is unresolved.
Impact of EA26002. Whether an Engineering Analysis concludes in recall, mitigation order, or closure is not yet known; outcome materially affects FSD's product roadmap.
Primary-source gap. tesla.com/fsd/safety and two InsideEVs articles returned 403 during this pass, so Tesla's own current safety framing and the latest InsideEVs hands-on coverage weren't captured directly — would strengthen the synthesis.

Provenance

Sub-questions:

What is Tesla FSD (architecture, hardware, software stack) as of 2026?
What is the current safety, disengagement, and accident data?
What is the regulatory status (US NHTSA, California DMV, international)?
How does FSD compare to competitors (Waymo, Mercedes, Chinese OEMs)?
What is the status of robotaxi / unsupervised FSD rollout?

URLs fetched (5 successful, 3 failed):

Electrek: NHTSA upgrades Tesla FSD visibility investigation to 3.2M vehicles — news — primary detail on EA26002 scope and meaning.
Electrek: California regulator confirms Tesla not operating autonomous vehicle service — news — CPUC on-record statement; TCP permit mechanics.
Electrek: Tesla expands unsupervised Robotaxi area, still handful of vehicles — news — Austin unsupervised fleet size (4–8), Waymo comparison.
Phil Koopman: New Tesla FSD safety data — academic/expert blog — methodological critique of Tesla's safety report.
Teslarati: Tesla expands unsupervised Robotaxi service to two new cities — industry blog — Dallas/Houston launch details for April 18, 2026.
~~[Failed: https://www.tesla.com/fsd/safety]~~ — 403, primary source not captured.
~~[Failed: https://insideevs.com/news/781782/tesla-removes-robotaxi-model-y-safety-driver-testing/]~~ — 403.
~~[Failed: https://insideevs.com/news/783339/mercedes-mb-drive-assist-pro-test/]~~ — 403.

Additional citations drawn from WebSearch snippets (not directly fetched — treat as weaker evidence):

Tools used: WebSearch, WebFetch. Generated: 2026-04-20 16:50 EDT