Tesla Full Self-Driving (FSD) Explained: The 2026 Deep Dive

What it is: Tesla Full Self-Driving (Supervised), commonly called FSD, is Tesla’s most-advanced consumer driver-assistance product. The current production version is FSD v14.3.2, released April 22, 2026, available only on Hardware 4 / AI4-equipped vehicles. It runs as an end-to-end neural network — Tesla replaced approximately 300,000 lines of hand-written C++ rules with a single unified AI model that takes camera frames in and produces driving controls out. SAE classification: Level 2 (driver supervised).
Who it is for: Tesla owners considering the FSD subscription or one-time purchase, anyone tracking the vision-only autonomous-driving bet, investors evaluating Tesla’s AV story, or AI engineers curious about end-to-end neural networks deployed at scale.
Best if: You want the factual current state of FSD, the NHTSA investigations, the Austin robotaxi deployment, and the open questions about whether vision-only autonomy will reach Level 4–5.
Skip if: You’re looking for Tesla-vs-Waymo opinion-piece content. This is a factual primer. Daily AI fundamentals in our free Beginners in AI newsletter.

Tesla Full Self-Driving is the most-deployed AI driving system in the world. Over 6 million Tesla vehicles have FSD-enabled hardware; several million owners have purchased or subscribed to FSD. In June 2025 Tesla launched a limited robotaxi service in Austin running modified Model Y vehicles, and in April 2026 the company released FSD v14.3.2 — a unified neural-network architecture that consolidated three previously-separate software stacks.

Simultaneously, NHTSA — the National Highway Traffic Safety Administration, the federal agency that regulates US vehicle safety — has open investigations covering 2.88 million to 3.2 million Tesla vehicles into specific FSD crash patterns and traffic violations. Both things are true at the same time. Here’s the factual picture as of mid-2026.

What is Tesla FSD exactly?

FSD — officially “Full Self-Driving (Supervised)” — is the premium tier of Tesla’s driver-assistance suite. It builds on top of Autopilot (standard cruise + lane-keep) and Enhanced Autopilot (adds Auto Lane Change, Navigate on Autopilot, Autopark, Summon).

FSD adds capabilities that the lower tiers don’t have:

Traffic and stop-sign control — Tesla recognizes traffic signals and stop signs and responds accordingly.
Autosteer on city streets — FSD operates on city streets, not just highways.
Smart Summon and Actually Smart Summon — the vehicle can drive itself in parking lots to come to the user.
Park Seek — the vehicle finds and pulls into parking spaces.
Navigate on FSD — complete navigation from point A to point B with stops, lane changes, and turns handled by the system.

The critical caveat: FSD is classified by SAE International as Level 2 (driver supervised), not actual self-driving. The driver must keep their hands available, eyes on the road, and remain ready to take over at any moment. Tesla’s product name has been the subject of regulatory action by the California DMV and others for that reason.

What is Hardware 4 (HW4) and why does it matter?

Hardware 4 (also called AI4) is the current generation of Tesla’s in-vehicle AI inference computer. Tesla introduced HW4 in 2023 and made it standard on all new vehicles. FSD v13 and v14 are only available on HW4-equipped vehicles — older Hardware 3 (HW3) vehicles run an earlier FSD branch.

Generation	Years	AI compute (TOPS)	FSD versions supported
Hardware 2 (HW2 / HW2.5)	2016–2019	Nvidia-based, smaller	Older Autopilot; no FSD v12+
Hardware 3 (HW3)	2019–2023	~144 TOPS (Tesla custom)	FSD up through v12.x; v14 “Lite” promised late June 2026
Hardware 4 (HW4 / AI4)	2023–present	~250 TOPS	FSD v13.x and v14.x — current production stack
Hardware 5 (HW5 / AI5)	Rumored 2026–2027	Not announced	Future; alongside Cybercab launch

The HW3 vs HW4 split is the single biggest practical division among Tesla owners. HW3 owners who paid for FSD years ago have not received the v13 / v14 capabilities; Tesla’s response has been to promise an “FSD v14 Lite” for HW3 in late June 2026 that delivers a subset of the v14 behavior. Whether HW3 vehicles ultimately get full feature parity is the subject of ongoing class-action litigation.

What is FSD v14.3.2 and what changed?

FSD v14.3.2, released April 22, 2026, is the culmination of a two-year Tesla architecture overhaul. The headline change: Tesla unified three previously-separate software stacks — Autopilot highway driving, FSD city driving, and Robotaxi unsupervised driving — into a single end-to-end neural network.

Unified neural model. One AI handles parking lots (Summon), city streets, and highways without “switching gears” between stacks. Tesla reports significantly more fluid and consistent driving behavior as a result.
Video-language-model perception. The system processes the world through a model that generates possible futures rather than only forecasting bounding-box trajectories — a fundamentally different perception architecture than rule-and-detection pipelines.
Approximately 20% faster reflexes by Tesla’s own reporting on key reaction-time metrics.
HW4-only. v14.3.2 runs only on AI4 hardware.
Stepwise toward Unsupervised FSD. Tesla’s stated roadmap is for the same unified model to eventually power both customer vehicles and the Austin Robotaxi service without a safety supervisor.

Earlier FSD v12 (late 2023) had already replaced approximately 300,000 lines of hand-written C++ rules with neural-network code. FSD v13 (December 2024) added native camera-resolution end-to-end processing. FSD v14 is the third major iteration of the end-to-end approach.

What is the Tesla Austin Robotaxi service?

Tesla launched a limited robotaxi service in Austin, Texas in June 2025. The fleet consists of modified Model Y vehicles running FSD-Unsupervised software with onboard safety monitors (a human Tesla employee sits in the passenger seat or back during paid rides).

Tesla Austin Robotaxi metric	Through February 2026	Source
Launch date	June 2025	Tesla announcement; NHTSA filings
Software	FSD-Unsupervised (early variant of what became v14)	Tesla communications
Vehicles	Modified Model Y	Tesla communications
Safety monitor onboard	Yes (passenger seat or rear)	NHTSA Standing General Order reporting
Total fleet miles logged	~800,000 miles	NHTSA crash-reporting filings
Reported crashes	14	NHTSA Standing General Order data
Crash rate	~1 per 57,000 miles	Derived from NHTSA filings
Comparison to U.S. urban human baseline	~4x higher per Bloomberg / Electrek analysis	Press analysis; baseline contested

The crash-rate comparison is contested. Tesla’s own Vehicle Safety Report (tesla.com/fsd/safety) reports a different baseline more favorable to Tesla. Press analyses (Bloomberg, Electrek) reach different conclusions depending on which human-driver baseline they use and how they count Tesla-reported crashes. The raw NHTSA-filed numbers are the most-citable data point; their interpretation against a human baseline is genuinely difficult.

What is the NHTSA investigating?

NHTSA investigation	Scope	Status
EA26002 (Engineering Analysis)	2.88 million Tesla vehicles with FSD; crashes where FSD failed to detect roadway conditions impairing camera visibility (glare, fog, airborne debris)	Opened March 2026; ongoing
PE25012 (Preliminary Evaluation, then expanded)	58 (later 80) consumer complaints of FSD running red lights, drifting into oncoming lanes, crossing into wrong-way traffic	Opened December 2025; data deadlines extended to March 9, 2026
Probe escalation to 3.2M vehicles	Broader review of FSD-related crashes across the installed base	March 2026 escalation

NHTSA investigations don’t automatically lead to recalls or regulatory action; they are the federal investigation process that may or may not result in findings. The size and number of open investigations is notable — investigating 2.88M+ vehicles is a significant regulatory engagement — but the outcomes are still pending.

Tesla has also faced friction supplying NHTSA with requested data on traffic violations and crash details. NHTSA extended Tesla’s data-delivery deadline to March 9, 2026 after multiple extension requests. The relationship between Tesla and the federal regulator is one of the most-watched ongoing dynamics in the AV industry.

How does FSD compare to Waymo?

The two companies make opposite philosophical bets. See the full comparison in our AV computer-vision post and our Waymo profile. Summarized:

Dimension	Tesla FSD	Waymo
SAE Level	2 (Supervised); pursuing 4 via Robotaxi	4 (within geofenced operational design domain)
Sensors	Vision-only: 8 cameras, no LiDAR, no radar	29 cameras + multiple LiDAR + 6 radar
HD maps	None — commercial nav maps only	Centimeter-accurate pre-mapping required
Deployment unit	~6 million HW4+HW3 Teslas with FSD-capable hardware	~700 robotaxi vehicles
Geographic coverage (Supervised)	Available wherever owners drive Teslas	N/A
Geographic coverage (Unsupervised / robotaxi)	Austin TX only (early commercial)	10 US metros at SAE Level 4
2026 paid rides per week (robotaxi)	Tesla has not publicly disclosed	~500,000 (Waymo company communications)
Customer-facing experience	Buy a Tesla, supervise FSD yourself	Hail a Waymo One ride via app, no driver
Valuation lever	Tesla market cap (~$1T+); FSD value embedded	Waymo $126B (Feb 2026 round); Alphabet subsidiary

How much does FSD cost?

One-time purchase: $8,000 (US, 2026; pricing has fluctuated between $8,000 and $15,000 over the years). The purchase transfers with the vehicle to subsequent owners under current Tesla policy.
FSD subscription: $99/month (US, as of 2026). Available on a month-to-month basis; can be cancelled anytime.
Tesla insurance integration: In supported states, FSD usage data feeds Tesla’s in-house insurance pricing.
Free trials: Tesla periodically offers free FSD trial periods to all eligible vehicles (typically 30 days).

The one-time purchase vs subscription decision depends on how long you plan to own the vehicle. At $99/month, 81 months (~6.75 years) of subscription equals the $8,000 one-time price. For shorter ownership, subscription is cheaper; for longer ownership, the one-time purchase wins.

What about Cybercab?

Cybercab is Tesla’s purpose-designed robotaxi vehicle: a two-seat coupe with no steering wheel or pedals, unveiled at the “We, Robot” event in October 2024. Tesla has stated production targets in the 2026–2027 window with pricing under $30,000 per vehicle. As of mid-2026 the production timeline has slipped from initial promises — production was not running at scale at the time of this post.

If Cybercab launches at the price and capability Tesla has described, it would meaningfully change the economics of robotaxi service compared to Waymo’s ~$200K+ converted-Jaguar approach. Whether Tesla can hit the timeline and unit economics is the key open question. Earlier Tesla vehicle launches (Roadster, Cybertruck, Semi) all slipped meaningfully from initial timelines.

What does AI in Tesla FSD actually do?

End-to-end neural network. Camera frames go in, vehicle controls (steering, throttle, brake) come out. The intermediate steps — object detection, lane geometry, path planning — are not explicitly programmed; they emerge from training on millions of human-driven miles.
Multi-camera fusion. Eight cameras around the vehicle are fused into a unified 360° representation before being passed to the planning model.
Video language model perception (v14). The system generates possible future scenarios rather than only producing point estimates — a step toward more robust prediction.
Fleet-data training. Tesla trains on data from 6+ million Tesla vehicles. Edge cases that any one driver encountered are captured for training the next FSD update.
Simulation training. Tesla supplements real-world data with simulated scenarios — particularly for rare-but-dangerous situations.
On-vehicle inference. HW4’s ~250 TOPS of compute runs the inference for perception, prediction, and planning in real time without uplink to a Tesla server.

For deeper coverage of the underlying CV technology see Computer Vision in Autonomous Vehicles and Computer Vision in Modern Drones (the same model families power both).

FAQ

Is Tesla FSD actually self-driving?

No, in the strict SAE sense. FSD is officially labeled “Full Self-Driving (Supervised)” and is classified as SAE Level 2. The driver must remain attentive and ready to take over at all times. The marketing name has been the subject of regulatory action by the California DMV and others. The Austin Robotaxi service operates with FSD-Unsupervised software and a safety monitor onboard.

Should I buy FSD or subscribe?

Break-even is approximately 6.75 years of subscription at $99/month vs the $8,000 one-time purchase. Buy outright if you’ll own the vehicle longer than that; subscribe if you’ll own less time, want flexibility, or want to test it first.

Will my HW3 car ever get FSD v14?

Tesla has committed to delivering an “FSD v14 Lite” for HW3 in late June 2026, with a subset of v14 behavior. Full feature parity with HW4 has not been promised; it’s the subject of ongoing class-action litigation. If feature parity matters to you, plan to eventually upgrade to a HW4-equipped vehicle.

Is FSD safer than human driving?

Genuinely contested. Tesla’s own Vehicle Safety Report reports favorable per-mile crash rates vs the broader US human-driver baseline. Press analyses (Bloomberg, Electrek) of Austin Robotaxi crash filings reach less-favorable conclusions when normalized against urban human-driver baselines. NHTSA investigations are ongoing. The honest read in mid-2026: the picture is unresolved.

Can FSD handle snow or heavy rain?

Tesla’s vision-only approach struggles when cameras are obscured — heavy snow, fog, heavy rain, dust, sun glare directly into lenses. NHTSA’s March 2026 investigation EA26002 specifically covers crashes where FSD failed to detect roadway conditions impaired camera visibility. Tesla’s response is to train models on more degraded-vision data; the operational reality in 2026 is that adverse weather is FSD’s biggest weakness.

What is Cybercab and when will it ship?

Cybercab is Tesla’s purpose-built robotaxi vehicle — two-seat coupe, no steering wheel or pedals. Tesla unveiled it October 2024. Production targets were stated as 2026–2027 with sub-$30,000 unit pricing. As of mid-2026, production has slipped from initial commitments. Tesla’s vehicle-launch history (Roadster, Cybertruck, Semi) suggests further slippage is plausible.

Where can I read Tesla’s official FSD documentation?

Tesla’s support page at tesla.com/support/fsd is the official reference. The Vehicle Safety Report is at tesla.com/fsd/safety. NHTSA investigation documents are public at nhtsa.gov/recalls and through the agency’s investigation portal.

The bottom line

Tesla FSD is the most-deployed AI driving system in the world and the most-watched autonomous-driving bet. The v14.3.2 unified end-to-end neural network is a genuine technical achievement; the Austin Robotaxi service is a genuine SAE Level 4 attempt; the vision-only philosophical bet is intact and being tested at scale.

The unresolved questions: whether vision-only autonomy can reach Level 4 safety margins outside the narrow Austin deployment; whether Cybercab ships at the promised price and timeline; whether NHTSA investigations result in material regulatory action; and whether HW3 vehicles ever achieve feature parity with HW4. The next 24 months answer most of these.

For broader context: Waymo Explained, Computer Vision in Autonomous Vehicles, Computer Vision in Modern Drones, Every AI Model Worth Knowing in 2026. Daily AI fundamentals in our free Beginners in AI newsletter.

Learn Our Proven AI Frameworks

Beginners in AI created 6 branded frameworks to help you master AI: STACK for prompting, BUILD for business, ADAPT for learning, THINK for decisions, CRAFT for content, and CRON for automation.

Get Smarter About AI Every Morning

Free daily newsletter — one story, one tool, one tip. Plain English, no jargon.

Free forever. Unsubscribe anytime.

Sources

Tesla, Full Self-Driving (Supervised) Support page — primary reference for FSD versions, hardware requirements, and capability descriptions.
Tesla, Full Self-Driving (Supervised) Vehicle Safety Report — Tesla’s official per-mile crash-rate reporting.
Tesla 10-K filings via SEC EDGAR — SEC EDGAR (Tesla, CIK 0001318605) — primary financial reference including FSD revenue disclosures.
NHTSA Office of Defects Investigation, Recalls and Investigations portal — primary source for EA26002 and PE25012.
NHTSA Standing General Order on Crash Reporting — nhtsa.gov standing general order. Primary source for the 14-crash, 800K-mile Austin Robotaxi figures cited.
NHTSA PE25012, Preliminary Evaluation document — the December 2025 investigation opening.
SAE International, SAE J3016 Levels of Driving Automation — industry-standard definition of Level 2 vs Level 4.
California DMV, Autonomous Vehicles program — California testing permits and disengagement reports.
Electrek and Bloomberg ongoing coverage of FSD crashes, NHTSA filings, and Austin Robotaxi operations — high-authority automotive and financial press reporting.
Tesla AI Day presentations (publicly available on YouTube via Tesla’s official channel) — technical detail on the end-to-end neural network architecture.

What Are Gemini Gems? A Guide

Best AI Prompts for HR

What Is Google Gemini? A Guide