Tesla FSD 14.3.1 Highway Exit Fix: What Changed and What to Do

FSD 14.3.1 Is Rewriting How Your Tesla Handles the Highway

Tesla's Full Self-Driving software has just received one of its most substantive upgrades in recent memory. FSD v14.3 began rolling out on April 7, 2026, followed by the point release v14.3.1 on April 14 — and early real-world testing is producing results that stand out even among seasoned FSD observers.

The headline behavior: starting with FSD v14.3, the car now consistently moves into the right lanes exactly 2.5 miles before a highway exit. No more last-second scrambles across three lanes. No more missed exits. Just a calm, predictable blinker activation the moment the exit crosses the 2.5-mile threshold.

▶ Watch Video on X

Sawyer Merritt, who has logged extensive FSD miles, reports he has been able to consistently repeat this behavior across multiple drives — a key distinction from the occasional lucky runs that FSD users have come to manage expectations around.

📊 What Changed in FSD 14.3 and 14.3.1

Change	Type	Models
Highway Exit Lane Positioning — Initiates right-lane moves 2.5 miles before exit; blinker activates at threshold	User-Confirmed	All FSD-capable
20% Faster Reaction Time — Rewritten AI compiler and runtime using MLIR	Official	All FSD-capable
Upgraded RL Training Stage — Reinforcement Learning improvements across a wide variety of driving scenarios	Official	All FSD-capable
Enhanced Vision Encoder — Better rare/low-visibility scenario handling, 3D geometry, traffic sign recognition	Official	All FSD-capable
Mitigated Unnecessary Lane Biasing — Reduced minor tailgating and erratic lateral positioning	Official	All FSD-capable
Emergency Vehicle & Rare Scenario Handling — Improved response to school buses, right-of-way violators, emergency vehicles	Official	All FSD-capable
Animal Handling — Focused RL training on small animals with added proactive safety rewards	Official	All FSD-capable
Parking Decisiveness — Improved spot selection, maneuvering, and location pin prediction	Official	All FSD-capable
Parking UI Arrow Shortcut — Tapping the small arrow in parking UI now directly opens parking options menu	Official (14.3.1)	All FSD-capable

"Absolutely Sublime" — The LA to SF Test

While Merritt's data is precise and repeatable, it's the qualitative feedback from long-distance runs that's turning heads. Whole Mars Catalog — one of the most prolific FSD testers on X — stopped mid-drive between Los Angeles and San Francisco to post a single verdict on FSD 14.3.1:

The LA-to-SF corridor is one of the most demanding FSD test routes in North America — heavy traffic through the Central Valley, complex interchange geometry, and variable road conditions. "Absolutely sublime" from a driver who has logged hundreds of FSD hours on that exact route carries real weight.

This aligns with Elon Musk's earlier characterization of FSD v14.3 as the "big piece of the puzzle" — language that suggested a more fundamental architectural shift rather than incremental tuning. The MLIR-based compiler rewrite and upgraded RL training pipeline appear to be delivering on that framing. For more on FSD's evolution, see our FSD coverage.

🚦 Owner's Action Plan

1. Check your software version now. Go to Controls → Software on your touchscreen. You're looking for software version 2026.2.9.7 (FSD 14.3.1) or 2026.2.9.6 (FSD 14.3).
2. Request the update if available. If an update is pending, tap Install Now. If none is showing, tap Check for Updates. FSD 14.3.1 is actively expanding to Model S and Model X vehicles as of April 14, 2026 — all models should see it in the coming days.
3. Prioritize a highway drive to test exit behavior. Once updated, take a route with at least one highway exit. Watch for the right-lane merge to begin approximately 2.5 miles out. The blinker should activate the moment you cross that threshold.
4. Verify your FSD profile settings. Go to Autopilot → Customize Autopilot and confirm your lane change preferences aren't set to manual confirmation — automatic lane changes let the new exit logic work as intended.
5. Report anomalies via the thumbs-down button. If the 2.5-mile exit behavior doesn't trigger consistently, use the in-car feedback button immediately after the event. Your data directly feeds the RL training pipeline that produced this improvement.

Who Has It Right Now?

FSD v14.3 initially rolled out to the original FSD beta tester cohort starting April 7. The v14.3.1 point release followed on April 14 and is currently expanding more broadly, with Model S and Model X specifically confirmed in the latest wave. If you're on FSD (Supervised) and don't see it yet, you're likely in the next wave — rollouts at this stage typically complete within 1-2 weeks. Keep an eye on all software updates as the rollout progresses.

📰 Deep Dive

The 2.5-mile exit threshold behavior is more significant than it might appear on paper. Previous FSD versions were frequently criticized for leaving exit lane changes too late — sometimes cutting across multiple lanes within the final half-mile, creating uncomfortable or unsafe situations. A consistent 2.5-mile buffer gives the car time to execute gradual, natural-feeling merges that blend with surrounding traffic rather than forcing gaps. The fact that Merritt has been able to repeat this across multiple drives — not just observe it once — suggests this is a trained behavior rather than a lucky edge case.

The architectural changes underneath this release are worth understanding. The move to an MLIR-based AI compiler and runtime isn't just an optimization — it's a rewrite of how the neural network's decisions get translated into vehicle actions. A 20% faster reaction time at highway speeds means the car is processing and responding to its environment roughly 60-80 milliseconds faster than before. At 70 mph, that's approximately 6-7 additional feet of reaction distance. Combined with the upgraded vision encoder's improved 3D geometry understanding, the system is both seeing the road more accurately and acting on what it sees more quickly.

The reinforcement learning upgrades are the less visible but arguably more important piece. RL training is how Tesla teaches FSD to handle situations that are rare enough that supervised learning data is sparse — small animals, right-of-way violators, unusual vehicle types. The fact that Tesla specifically called out "harder examples" and "proactive safety rewards" in this area suggests they're deliberately pushing the model into the long tail of edge cases that have historically caused FSD disengagements. If the LA-to-SF experience is any indication, that investment is starting to pay off at scale.