Lemonade's Autonomous Car product, live in Arizona since January 26, 2026 and now available in four states, prices Tesla Full Self-Driving-engaged miles at half the per-mile rate of human-driven miles within the same policy (Lemonade, January 2026). That single design choice forces insurers to rate one vehicle against a classification variable that can switch several times inside a single trip, a granularity no personal auto statistical plan was built to carry.

How the Split-Rate Mechanism Works

The mechanics are narrower than the headline suggests, and that is precisely what makes the product a clean test case. Lemonade connects to Tesla's Fleet API with customer permission, pulling per-mile driving data that identifies whether FSD was engaged, which software version was running, and how the vehicle was actually operated at each moment of exposure. The underwriting model then applies two separate per-mile rates within one policy: a standard rate for human-driven miles and a rate cut by roughly 50% for miles logged while FSD was engaged, reflecting Tesla's own claim that FSD-driven miles run about twice as safe as manually driven ones. The product launched in Arizona on January 26, 2026, followed by Oregon in February, and has since expanded into Indiana and Colorado (eMarketer, 2026). It supports households that mix a Tesla with non-FSD vehicles under a single policy and stacks with Lemonade's existing safe-driving and multi-line bundling discounts.

The discount is aggressive relative to what Tesla itself offers through its captive insurance arm, which caps its own safe-driving-score discount near 10% (Kavout, 2026). It also lands into a market that is growing fast on the exposure side: Tesla FSD mileage climbed from about 6 million miles in 2021 to 4.25 billion miles in 2025, with internal projections pointing toward 10 billion miles in 2026, and paid FSD subscribers reached nearly 1.1 million by the fourth quarter of 2025 (Kavout market analysis, 2026, citing Tesla and Lemonade data). Lemonade's own fourth-quarter 2025 results showed in-force premium of $1.24 billion, up 31% year over year, alongside a record 52% gross loss ratio, giving the company both the balance sheet exposure and the loss-ratio headroom to experiment with a rating variable this granular before its book has matured (Lemonade Q4 2025 results, via Kavout). None of that changes what the actuarial literature has mostly treated as a hypothetical: how do you rate a policy when the risk driver itself is not fixed at issuance, not even fixed at the trip level, but toggles mid-drive between a human and a system?

The Classification Problem: A Rating Variable With No Class Code

Personal auto ratemaking has always priced exposure at a coarse grain. The earned car-year is the traditional unit, and the classification variables loaded onto it, territory, vehicle symbol, driver class, prior-loss surcharge, are fixed at issuance and revisited at renewal, not mid-term. Telematics-based usage-based insurance loosened that only partially. Programs built on continuous behavioral scoring, hard braking, phone handling, speed, time of day, still resolve to a single discount tier applied for the coming policy period. The exposure unit does not change; only the price attached to it updates on a renewal cadence measured in months.

Lemonade's mechanism breaks that pattern in kind, not degree. The classification state, human-operated or FSD-engaged, can flip multiple times within one origin-to-destination trip: a highway merge under FSD, a manual turn onto a side street, FSD re-engaged for the next stretch of interstate. Earned premium has to be calculated at the mile level, conditional on real-time operator state, rather than at the trip level or the policy-term level that state statistical plans assume. ISO's Personal Auto stat plan and the state-specific plans built on it were designed to capture classification data at the vehicle-policy grain: one set of class codes per vehicle per reporting period. None of them carry a native field for per-mile, time-varying operator state at the moment of loss. A carrier filing this rating variable has to either shoehorn it into an existing usage-based data call or seek a variance for a genuinely new statistical plan element, and state statistical agents have not yet said which.

It is worth being precise about what is actually new here, since per-mile billing itself is not it. Pay-per-mile carriers have reported banded mileage exposure to state statistical agents for more than a decade; the infrastructure for metering a policy at the mile level, rather than the car-year level, already exists and already clears regulatory review. What has no precedent is a second dimension layered onto that same mile: not just how many miles were driven, but which of two structurally different risk-generating processes, a human or a software system, was in control of the vehicle for each one. A mileage band is a static count. An operator-state field is a real-time attribute that has to be captured, verified against vendor telemetry, and reconciled with a claim file if a loss occurs mid-segment. That is the genuinely unprecedented element, and it is also the one most likely to draw the sharpest regulatory scrutiny during the review process.

Credibility at a Sub-Trip Grain

The credibility question compounds the classification problem rather than sitting alongside it. The classical full-credibility standard used across U.S. property-casualty ratemaking calls for roughly 1,082 expected claims before observed experience is trusted at 90% confidence within a 10% margin of the true underlying value, and Bühlmann credibility blends thinner experience against an industry complement using Z = n / (n + k), where n is the observed claim count. Lemonade's FSD-mile segment is a subset of a subset: Tesla owners, in four states, on a product that had been live for only a matter of months by mid-2026. There is no plausible scenario in which Lemonade's own book generates 1,082 FSD-mile claims before the relativity has to be filed and priced.

A rough illustration sharpens how wide that gap is. Even in the unrealistic case that every one of Tesla's nearly 1.1 million paid FSD subscribers nationally were a Lemonade policyholder, itself far from true given the product's four-state footprint, a typical personal auto book generates on the order of one bodily injury claim per several hundred insured-years, meaning full credibility above the 1,082-claim standard would still take a multi-year accumulation window even at that inflated scale. Lemonade's actual FSD-mile book, confined to Arizona, Oregon, Indiana, and Colorado and only months old as of mid-2026, sits several orders of magnitude below even that hypothetical ceiling. The relativity filed today is not a credibility-weighted blend between carrier experience and an industry complement in the Bühlmann sense; it functions as close to 100% complement, and the complement itself is supplied by the vehicle manufacturer rather than by an independent industry source such as ISO or a state rating bureau.

That means the 50% relativity is not coming from Lemonade's own seasoned loss experience. It is anchored to Tesla's own safety reporting, which claims one major collision every 5.3 million FSD-supervised miles in North America against a national average of roughly one crash every 660,000 miles (Tesla FSD Safety Report, November 2025). A Reuters investigation published in May 2026 found that comparison methodologically strained: Tesla's numerator counts crashes severe enough to deploy an airbag, while the NHTSA and FHWA denominator it is compared against includes any crash requiring a tow truck, a far lower severity threshold that inflates Tesla's apparent safety advantage by roughly threefold. Reuters also noted that Tesla redacts data from the standing general order crash reports it files with NHTSA at a level no other manufacturer applies. An actuary signing off on Lemonade's relativity is being asked to credibility-weight a live rating variable against vendor-supplied safety data that had already drawn public methodological criticism, without a comparably sized book of the carrier's own FSD-mile claims to check it against.

The contrast with the profession's own attempt at this problem is instructive. A CAS E-Forum paper published in Spring 2023, "Projection of On-Road Liability Losses for Autonomous Driving," built a collective risk model projecting autonomous vehicle bodily injury severity from a commercial auto baseline, applying six adjustment factors including a velocity-squared injury scaling term and the removal of impaired- and distracted-driving loading that does not apply to a system. Its credibility-weighted case study landed at roughly $0.20 per mile in projected bodily injury severity, up from a $0.15 per mile commercial auto baseline, once uncertainty was incorporated, despite the paper's own assumption of AV safety gains (CAS E-Forum, Spring 2023). The authors were explicit about why: "past experience may potentially be less powerful or relevant in predicting future losses" (CAS E-Forum, Spring 2023) when the underlying software keeps changing. That is a direct tension worth sitting with. One credibility-literate actuarial framework applied uncertainty loading to autonomous driving and pushed severity up relative to a conventional baseline, while a live commercial product two years later is pricing a flat 50% discount off a single manufacturer's safety marketing data.

Liability Without a Liability Shift

The more consequential nuance is what Lemonade's product does not do. Tesla FSD Supervised remains an SAE Level 2 system: the human is legally the vehicle operator, and under current tort frameworks in Arizona and Oregon, the named insured's policy responds to a claim regardless of whether FSD was engaged at the moment of loss. That stands in contrast to the scenario the profession has spent years modeling in the abstract, in which liability at SAE Levels 4 and 5 shifts meaningfully toward the vehicle manufacturer and software developer as the human is removed from the control loop entirely. Lemonade's split-rate variable is a classification signal layered onto a frequency and severity expectation. It is not a liability-transfer mechanism, because liability has not actually moved.

That narrows the problem in a useful way. The ratemaking challenge here sits entirely on the classification and credibility side, not on the liability-attribution side that has dominated most academic and regulatory discussion of autonomous vehicle insurance. It makes Lemonade's launch a more tractable test case than the full Level 4/5 problem the Casualty Actuarial Society's own automated vehicle task force flagged years earlier as an area where actuarial research remained thin. But it also means the classification and statistical-plan infrastructure this product requires today is only the first layer. When Level 4 personal-use products eventually reach scale and liability genuinely does move toward the manufacturer, insurers will need this same trip-segment classification machinery, plus an entirely new liability-attribution layer stacked on top of it. Lemonade's launch previews the classification problem in isolation, before the liability problem arrives to compound it.

Where This Sits on the Telematics Maturity Curve

GEICO's telematics program offers the clearest benchmark for how far usage-based pricing has already scaled without touching this sub-trip problem. Telematics data now informs pricing on roughly 90% of GEICO's new personal auto business, according to comments from Berkshire Hathaway vice chairman Ajit Jain (Berkshire Hathaway, 2026). GEICO's DriveEasy program, which scores phone handling, braking, acceleration, cornering, time of day, and mileage, has advertised discounts as high as 25%, with 5% to 15% more typical in practice (The Zebra; Insurify, 2026). Progressive built a similar flywheel through its Snapshot program over more than a decade of continuous behavioral scoring. All of that penetration, at genuinely massive scale, still resolves to one price applied for one renewal period. The classification variable does not change mid-term, let alone mid-trip.

Lemonade's FSD product is the first at-scale entrant into a different tier of the telematics maturity curve: a rating variable that resolves multiple times within a single earned-exposure unit rather than once per renewal cycle. The EV insurance market backdrop makes the stakes of getting this pricing right more concrete. Electric vehicle owners paid an average of $3,159 annually for insurance in 2026, roughly 42% more than the $2,218 average for gas-powered vehicles (eMarketer; Insurify, 2026), a gap Lemonade is explicitly betting it can compress for FSD-equipped Teslas by pricing the autonomous-mile discount aggressively enough to become, in the company's own framing, the go-to insurer for software-defined vehicles.

What State DOIs Will Need Before Approving This Rating Plan

Regulators reviewing a rate filing built on this structure are working from the same actuarial memorandum requirements the NAIC's usage-based insurance and vehicle telematics research has documented for over a decade, extended into a genuinely new grain of classification. A defensible filing needs to disclose whose data underlies the relativity, Lemonade's own book or Tesla's vendor-supplied safety claims, and show a credibility weighting schedule for how that mix will shift as the carrier's own FSD-mile experience seasons. It needs a statistical-plan reporting proposal for the new per-mile operator-state field, since no existing ISO or state stat plan carries one today. It needs explicit disclosure of the data-quality caveats already raised in public reporting about the safety claims the relativity leans on. And it needs a trigger, an exposure or claim-count threshold at which Lemonade commits to re-filing the relativity based on its own seasoned experience rather than continuing to price off a manufacturer's marketing data.

None of that is unique to Lemonade or to Tesla. Every carrier layering ADAS and hands-free highway systems into personal auto books, GM's Super Cruise, Ford's BlueCruise, and whatever comparable systems reach meaningful engagement-time share over the next few years, will eventually face the identical sub-trip classification problem as those features move from occasional use to majority-of-miles use. Lemonade has simply filed the first live test case. The statistical-plan and credibility infrastructure the industry builds in response to this filing, not the next one, is what will determine whether the next wave of autonomous-mile products can be priced on the carrier's own experience instead of borrowed vendor claims.

Further Reading

Sources