What is the Full Self-Driving Computer 4?
The Full Self-Driving Computer 4 (often abbreviated as HW4 or AI4) is Tesla's fourth-generation onboard compute platform designed to power its Autopilot and FSD (Supervised) driver-assistance software. Introduced in early 2023 and iterated through 2026 model-year vehicles, HW4 represents a significant leap in sensor resolution, processing throughput, and neural-network inference speed over its predecessor, HW3.
However, a critical distinction must be made immediately: despite the "Full Self-Driving" branding, this hardware powers a SAE Level 2 driver-assistance system, not a Level 4 autonomous vehicle. The driver must remain engaged, monitor the road, and be prepared to take over at any moment.
SAE Automation Levels — Where FSD Computer 4 Fits
| Level | Definition | Example |
|---|---|---|
| Level 0 | No automation | Basic cruise control |
| Level 1 | Driver assistance (steering OR speed) | Adaptive cruise OR lane-keeping |
| Level 2 | Partial automation (steering AND speed, driver supervises) | Tesla FSD (Supervised), GM Super Cruise, Ford BlueCruise |
| Level 3 | Conditional automation (system drives, driver on standby) | Mercedes Drive Pilot (limited markets) |
| Level 4 | High automation (driverless in geofenced areas) | Waymo One, Zoox |
| Level 5 | Full automation (any condition, any location) | Does not exist commercially |
Tesla FSD Computer 4 operates at Level 2. The system handles steering and acceleration simultaneously but requires continuous driver supervision. This is fundamentally different from Level 4 robotaxi services like Waymo, which operate without a human driver in defined service areas.
HW4 Hardware Specs: Sensors, Compute, and Architecture
Tesla's approach diverges sharply from the lidar-equipped stacks used by Waymo, Cruise, and Zoox. The FSD Computer 4 relies on a camera-only (Tesla Vision) architecture, which Tesla argues is more scalable and cost-effective. Here is the technical breakdown:
HW4 Sensor and Compute Suite
| Component | HW3 (Previous Gen) | HW4 (Current Gen) |
|---|---|---|
| Camera count | 8 cameras (1.2 MP) | Up to 11 cameras (5 MP) |
| Front-facing camera range | ~250 m | ~300+ m (narrow FOV lens) |
| Processor | Dual custom FSD chips (144 TOPS combined) | Dual custom AI chips (~300+ TOPS combined est.) |
| Process node | 12 nm | 7 nm (Samsung) |
| Radar | Removed (post-2021) | None (camera-only) |
| Lidar | None | None |
| Ultrasonic sensors (USS) | 12 sensors (removed late 2022) | None (Tesla Vision for proximity) |
The higher-resolution cameras provide roughly 4x the pixel density of HW3, giving the neural network more data for object detection at distance. Tesla also added a dedicated cabin-facing camera for driver monitoring in newer configurations, and some Model 3 Highland and Model Y Juniper variants carry additional forward-facing narrow-angle lenses for highway perception.
How FSD (Supervised) Actually Works on HW4
Tesla's FSD (Supervised) software, running on HW4, handles several driving tasks simultaneously:
- Lane-keeping and lane changes on highways and surface streets
- Speed management including adaptive cruise and speed-limit awareness
- Intersection navigation including traffic lights, stop signs, and turn execution
- Parking assistance including parallel, perpendicular, and Smart Summon (low-speed, lot only)
- Obstacle avoidance including construction zones and debris detection
The system uses an end-to-end neural network architecture (introduced with FSD v12 in 2024 and refined through 2026) that processes raw camera inputs directly into driving decisions, rather than relying on hand-coded rules. This approach has improved fluidity and human-like behavior but introduces unpredictability — the system occasionally makes decisions that surprise even attentive drivers.
⚠️ Critical Limitations of FSD Computer 4
- Driver must remain engaged at all times. Hands on or near the wheel, eyes on the road. The cabin camera monitors attention.
- Not approved for driverless operation anywhere. No jurisdiction classifies Tesla FSD as Level 3 or above.
- Weather degradation: Heavy rain, fog, snow, and direct sun glare reduce camera effectiveness. The system may disengage with minimal warning.
- Construction zones and unmapped roads: Performance degrades where lane markings are absent, shifted, or conflicting.
- No geofencing safety net: Unlike Level 4 robotaxis restricted to mapped areas, FSD (Supervised) activates on virtually any road — including roads where its capabilities are marginal.
- Phantom braking: Sudden deceleration without apparent obstacle remains a documented behavior, reported to NHTSA by owners.
HW4 vs. Level 4 Robotaxis: A Technology Stack Comparison
Understanding the gap between Tesla's FSD Computer 4 and true robotaxi hardware helps contextualize where each technology stands in 2026:
| Attribute | Tesla FSD (HW4) | Waymo (6th-Gen Driver) | Zoox (Purpose-Built) |
|---|---|---|---|
| SAE Level | Level 2 | Level 4 | Level 4 |
| Driver required | Yes — always | No | No |
| Primary sensors | Cameras only | Lidar + radar + cameras | Lidar + radar + cameras |
| HD mapping | No (vision-based) | Yes (centimeter-level) | Yes (centimeter-level) |
| Geofenced | No | Yes | Yes |
| Redundant compute | Limited | Full redundancy | Full redundancy |
| Operational in heavy rain | Degraded | Yes (within limits) | Yes (within limits) |
| Unit cost (sensor suite est.) | ~$1,500–$2,000 (integrated) | $100,000+ per vehicle | $100,000+ per vehicle |
Tesla's camera-only approach keeps vehicle cost low and enables deployment across millions of consumer cars. The trade-off is reduced redundancy: without lidar's direct distance measurement or radar's all-weather penetration, the system relies entirely on visual inference, which can fail in edge cases that lidar-equipped stacks handle more robustly.
Where Can You Use FSD (Supervised) Right Now?
Because FSD (Supervised) is a Level 2 driver-assist system, it is technically available anywhere a Tesla with HW4 and an active FSD subscription can drive. There are no geofence restrictions. As of 2026, FSD (Supervised) is available in:
- United States: All 50 states (feature availability varies by software version)
- Canada: Nationwide
- China: Rolled out in select cities starting mid-2025, expanding through 2026
- Australia: Limited beta as of late 2025
- Europe: Not yet approved for FSD (Supervised); basic Autopilot only (lane-keeping + adaptive cruise). EU type-approval processes remain a barrier.
Contrast this with Level 4 robotaxi services, which operate only in specific geofenced cities. Waymo One, for instance, is available to the public in Phoenix, San Francisco, Los Angeles, and Austin as of early 2026, with limited expansion planned. Zoox operates a pilot service in Foster City and parts of San Francisco.
Pricing: FSD Subscription vs. Robotaxi Rides vs. Rideshare
Tesla's FSD (Supervised) is not a transportation service — it's a driver-assist feature you use in your own vehicle. The cost structure is entirely different from hailing a Waymo or calling an Uber:
| Option | Cost Structure | Monthly Cost (est.) | Notes |
|---|---|---|---|
| Tesla FSD (Supervised) subscription | $99/month or $8,000 one-time | $99 | Requires Tesla ownership; driver must supervise |
| Tesla FSD purchase (lifetime) | $8,000 upfront | $0 (amortized) | Transfers with vehicle, not to new owner (as of 2026) |
| Waymo One (Phoenix/SF/LA) | Per-ride pricing | $15–$40/ride typical | Comparable to Uber/Lyft; no driver tip needed |
| Uber/Lyft (US average) | Per-ride pricing | $15–$35/ride typical | Human driver; tip expected ($3–$8) |
| Zoox (pilot, Foster City/SF) | Free (invite-only pilot) | $0 | Not commercially available; limited seats |
The economics only make sense if you already own or plan to buy a Tesla. FSD (Supervised) adds convenience on commutes and road trips but does not eliminate the need for a licensed, attentive driver. For consumers who don't own a vehicle, robotaxi services offer genuine driverless transportation — at per-ride costs comparable to traditional rideshare.
Safety Data: What We Know About FSD on HW4
Tesla publishes quarterly safety reports claiming that vehicles with Autopilot engaged record fewer crashes per million miles than the US average. However, these reports have methodological limitations that independent analysts, including those cited by NHTSA, have flagged:
Tesla Safety Claims vs. Independent Context
- Tesla's claim (Q4 2025 report): 1 crash per 4.7 million miles with Autopilot engaged, vs. US average of ~1 crash per 670,000 miles.
- Key caveat: Autopilot is used predominantly on highways (divided, limited-access), which are inherently safer than surface streets. Comparing highway Autopilot miles to all US driving miles is not apples-to-apples.
- NHTSA Standing General Order (SGO) data: Between 2021 and 2025, Tesla reported more ADAS-related crash incidents than any other manufacturer — partly because Tesla has more Level 2 vehicles on the road and more robust telemetry reporting.
- IIHS research: Studies show that Level 2 systems can lead to driver complacency, with attention degradation increasing over longer trips.
For Level 4 robotaxi comparison, Waymo reported in 2025 that its fleet of 1,000+ vehicles completed over 10 million fully driverless miles in 2024 with a crash rate lower than human benchmarks in comparable urban environments, according to data reviewed by the IIHS.
Regulatory Landscape for Tesla FSD and Robotaxis
Tesla's FSD (Supervised) operates under a fundamentally different regulatory framework than Level 4 robotaxis:
- Level 2 systems (Tesla FSD, GM Super Cruise, Ford BlueCruise): Regulated as driver-assistance features under existing Federal Motor Vehicle Safety Standards (FMVSS). NHTSA has authority to investigate and order recalls but does not pre-approve these systems before deployment. Tesla has faced NHTSA investigations into Autopilot/FSD behavior at emergency scenes and in low-visibility conditions.
- Level 4 robotaxis (Waymo, Zoox, Cruise): Require state-level permits to operate commercially. California's CPUC and DMV, Arizona's governor-executive-order framework, and Texas's emerging regulations each set different requirements for testing, insurance, incident reporting, and geographic boundaries. The NHTSA oversees vehicle safety standards, while state agencies control operational deployment.
In 2025–2026, NHTSA has intensified scrutiny of Level 2 systems, proposing new rules that would require more granular driver-engagement monitoring and standardized performance reporting. Tesla has pushed back on some proposals, arguing its camera-based driver monitoring is sufficient.
Is the Full Self-Driving Computer 4 Truly Self-Driving?
No. The FSD Computer 4 powers a sophisticated Level 2 driver-assistance system — one of the most capable available to consumers, but one that still requires a licensed, attentive driver behind the wheel at all times. The hardware is powerful: higher-resolution cameras, faster inference chips, and an end-to-end neural network that handles complex urban scenarios with increasing fluency. But the software it runs, FSD (Supervised), has not achieved Level 3, 4, or 5 autonomy.
Tesla CEO Elon Musk has repeatedly predicted imminent full autonomy — "a year away" since roughly 2016. As of 2026, that prediction remains unfulfilled. The FSD Computer 4 is a strong hardware platform, but achieving true self-driving requires solving edge cases (unprotected left turns in dense traffic, construction-zone ambiguity, sensor degradation in adverse weather) that camera-only perception has not yet fully cracked.
Frequently Asked Questions
Can I buy a Tesla with HW4 today?
Yes. As of 2026, all new Model S, Model 3, Model X, Model Y, and Cybertruck vehicles ship with HW4 (AI4) hardware. Older vehicles with HW3 can be retrofitted, though Tesla has not offered a paid upgrade path consistently — check with your local Service Center for current availability.
Does FSD (Supervised) work without internet connectivity?
Yes. The neural network runs entirely on-device using the HW4 chips. Internet connectivity is required for software updates, map data refreshes, and telemetry upload, but the core driving functions operate offline.
How does HW4 handle snow and heavy rain?
Poorly compared to lidar-equipped systems. Camera-based perception degrades significantly in heavy precipitation, snow-covered lane markings, and fog. FSD (Supervised) may issue a "takeover" request or reduce functionality. Drivers in snowy climates should expect to assume full control frequently during winter months.
Will HW4 eventually become Level 4 or 5?
Tesla has stated its goal is to achieve unsupervised full self-driving via software updates to existing HW4 vehicles. However, no regulatory body has approved this, and significant technical and legal hurdles remain. As of 2026, no timeline has been confirmed, and no jurisdiction permits driverless Tesla operation on public roads.
What's the difference between Autopilot and FSD (Supervised)?
Autopilot is the base driver-assist system included with every Tesla: adaptive cruise control and lane-keeping on marked highways. FSD (Supervised) is the paid upgrade ($99/month or $8,000 one-time) that adds city-street navigation, traffic-light/stop-sign recognition, automatic lane changes, and parking assistance. Both require full driver attention.
Is HW4 worth the upgrade from HW3?
If you're buying a new Tesla, you'll get HW4 automatically. For existing HW3 owners, the camera resolution bump and compute headroom improve FSD responsiveness and object detection at distance, but the core experience — a Level 2 system requiring supervision — remains the same. A retrofit (if available) typically costs $1,000–$2,000 plus installation.
