Lemonade's Autonomous Car product, live in Arizona since January 26, 2026 and now in four states, prices Tesla Full Self-Driving miles at half the per-mile rate of human-driven miles inside the same policy.

That one design choice puts a classification variable on the vehicle that can switch several times within a single trip. No personal auto statistical plan carries a field for it.

Key Takeaways

  • A 50% per-mile discount on FSD-engaged miles, against Tesla's own captive insurance arm capping its safe-driving discount near 10%.
  • The classification state changes mid-trip, so earned premium has to resolve at the mile level conditional on operator state, not at the trip or policy-term level stat plans assume.
  • 1,082 expected claims is the classical full-credibility standard at 90% confidence within 10%, and a four-state book months old cannot approach it.
  • One major collision every 5.3 million FSD-supervised miles against roughly one crash every 660,000 miles is the vendor comparison the relativity leans on, and its numerator and denominator use different severity thresholds.
  • Liability has not moved. FSD Supervised is SAE Level 2, the human remains the operator, and the named insured's policy responds whether or not the system was engaged.

How the Split Rate Works

Lemonade connects to Tesla's Fleet API with customer permission, pulling per-mile data identifying whether FSD was engaged, which software version was running, and how the vehicle was operated at each moment of exposure. The model then applies two per-mile rates inside one policy, with FSD miles cut roughly 50%, reflecting Tesla's claim that those miles run about twice as safe.

The product launched in Arizona on January 26, 2026, added Oregon in February and has since reached Indiana and Colorado. It covers households mixing a Tesla with non-FSD vehicles on one policy and stacks with existing safe-driving and bundling discounts.

The discount is aggressive against the market. Tesla's own captive caps its safe-driving-score discount near 10%, and the exposure base is growing fast: FSD mileage went from about 6 million miles in 2021 to 4.25 billion in 2025, with internal projections near 10 billion for 2026, and paid subscribers reached nearly 1.1 million by the fourth quarter of 2025.

Lemonade has room to run the experiment. Fourth-quarter 2025 in-force premium was $1.24 billion, up 31%, on a record 52% gross loss ratio.

A Rating Variable That Resolves Mid-Trip

Personal auto has always priced exposure coarsely. The earned car-year is the unit, and territory, vehicle symbol, driver class and prior-loss surcharge are fixed at issuance. Telematics loosened that only partly: continuous behavioral scoring still resolves to one discount tier for the coming policy period, so the exposure unit never changes.

This breaks the pattern in kind rather than degree. The state flips within a single trip: highway merge under FSD, manual turn onto a side street, FSD re-engaged for the next stretch. Premium has to be earned at the mile level, conditional on who was driving that mile.

Per-mile billing itself is not the novelty. Pay-per-mile carriers have reported banded mileage to state statistical agents for over a decade. What has no precedent is a second dimension on the same mile: which of two structurally different risk-generating processes was in control. A mileage band is a static count; an operator-state field has to be captured, verified against vendor telemetry, and reconciled with a claim file when a loss happens mid-segment. Neither the ISO Personal Auto plan nor the state plans built on it carry that field.

Credibility compounds it. Full credibility at 90% confidence within 10% asks for roughly 1,082 expected claims, and Bühlmann blends thinner data against a complement by Z = n / (n + k). Lemonade's FSD segment is Tesla owners, in four states, on a product months old. Its own book will not produce 1,082 FSD-mile claims before the relativity has to be priced.

So the filed relativity functions as close to a full complement, and the complement comes from the manufacturer. Tesla reports one major collision every 5.3 million FSD-supervised miles against a national average near one crash every 660,000 miles. A May 2026 Reuters investigation found the comparison strained: Tesla counts crashes severe enough to deploy an airbag while the federal denominator counts any crash requiring a tow, a lower threshold that inflates the apparent advantage roughly threefold.

The profession's own attempt ran the other way. A Spring 2023 CAS E-Forum paper projecting on-road liability losses for autonomous driving landed at roughly $0.20 per mile of bodily injury severity against a $0.15 commercial auto baseline once uncertainty was loaded, despite assuming safety gains, on the reasoning that "past experience may potentially be less powerful or relevant in predicting future losses" when the software keeps changing. One credibility-literate framework added a loading; a live product is taking a flat 50% off.

The Liability Did Not Move

FSD Supervised remains an SAE Level 2 system. The human is the legal operator, and under current tort frameworks the named insured's policy responds whether or not FSD was engaged at the moment of loss. The split rate is a classification signal on frequency and severity. It is not a liability transfer, because liability has not transferred.

That narrows the problem usefully and also limits what the product proves. Everything here sits on the classification and credibility side rather than the liability-attribution side that has dominated discussion of autonomous vehicle insurance. When Level 4 personal-use products reach scale and liability genuinely moves toward the manufacturer, insurers will need this same trip-segment machinery plus an attribution layer on top of it.

The rest of the market has scaled telematics without ever reaching this grain. Telematics informs pricing on roughly 90% of GEICO's new personal auto business, and DriveEasy advertises discounts up to 25% with 5% to 15% more typical, while Progressive built the same flywheel through Snapshot over more than a decade. All of it still resolves to one price for one renewal period.

The commercial stakes sit in the gap Lemonade is trying to close. Electric vehicle owners paid an average $3,159 for insurance in 2026, about 42% above the $2,218 gas-vehicle average. Every carrier layering hands-free highway systems into a personal auto book faces the same sub-trip classification problem as those features move from occasional to majority-of-miles use, and the regulatory record on usage-based insurance has nothing in it about an operator-state field.

Further Reading

Sources