State Farm Patents a Pricing Model That Rewrites Itself

State Farm won U.S. Patent 12,694,432 B2 on July 28, 2026, for a system that runs a GPT model against its own rating predictions, writes the code fix for the pricing sub-model that missed, tests the rewrite in simulation, and deploys the updated template. Claim 1 requires no human sign-off at any of those seven steps. It converts predicted-versus-actual monitoring, the check every ratemaking actuary runs by hand, into a closed loop that rewrites the rating engine.

Key Takeaways

• Claim 1 requires no human sign-off at any of its seven steps, from detecting a pricing miss through deploying the updated template into production.
• Two language models do two jobs. A GPT model outputs the difference between predicted and actual value; a distinct code LLM writes the instruction changes.
• 326 AI patents filed since 2014 put State Farm ahead of USAA and Allstate combined, but this is its first major grant aimed at the machinery that sets the price.
• 67 days is the countrywide rolling average for personal auto rate filing approval, against 234 days in California, the slowest jurisdiction tracked.
• Fewer than 25% of enterprises run a structured model risk framework built for generative AI; 44% of insurance executives report governance issues already sinking an AI project.

Patent Details

• Patent number: US 12,694,432 B2, "Large Language Modeling Systems and Methods for Building, Testing, and Validating a Predictive Model"
• Assignee: State Farm Mutual Automobile Insurance Company, Bloomington, Illinois
• Provisionals: 63/556,253 and 63/641,773, filed February 21 and May 2, 2024
• Non-provisional: 18/666,490, filed May 16, 2024, published as US 2025/0265624 A1
• Granted: July 28, 2026, with 21 claims across six drawing sheets
• Classification: CPC G06Q 30/0283 and G06F 30/20, price determination paired with design simulation and testing

Inside Claim 1

Claim 1 runs seven operations in sequence: detect an issue with a predictive sub-model that itself runs on a trained large language model; execute a GPT model trained to output the difference between predicted and actual value for the same product; feed the detected issue and that output into a second, distinct code LLM that writes a new model software template containing the code changes; generate a simulation environment from historical record data; execute the new template inside it; and update the template from the simulation's output (US 12,694,432 B2).

Two language models do two different jobs. One finds the pricing miss; the other rewrites the code that caused it. The detailed description states the purpose without hedging: the system builds, simulates and validates a predictive pricing model for calculating insurance rates. This is a rating-engine maintenance system, patented.

Dependent claims define the trigger. Claim 4 makes it a consistent variance between predicted and actual replacement value; claim 5 extends it to a trend exceeding one or more thresholds. Both describe the predicted-versus-actual backtest a pricing actuary runs each renewal cycle, wired into code that rewrites itself. Claim 8 has the GPT model output related data elements, the specification naming claims history, vehicle history, prior insurance history and public records.

Claim 10 lets it train for a specific jurisdiction and claim 9 generates a production model straight from the updated template. Claim 21 then has the processor generate a predicted value using the updated pricing sub-model, the first place the target is named as pricing rather than "products." Claims 6 through 8 add a natural-language interface, an alert and a prompt-driven review, but all three are dependent limitations. A system practicing only claim 1 satisfies the patent without them.

The classification pairing matters too: tying a pricing claim to a technical validation mechanism, rather than reciting "apply AI to pricing," is the narrow drafting surviving post-Recentive eligibility challenges.

A Rewrite Cycle Against a 234-Day Filing Queue

Set the seven-step claim against how a rating model traditionally gets corrected and the compression is the finding.

Step	Traditional pricing-model correction	Patent’s automated loop
Detect the miss	Pricing actuary reviews a predicted-vs-actual report on a scheduled cycle	GPT model continuously flags a variance or trend exceeding a threshold (Claims 4-5)
Diagnose the cause	Actuary and analyst research drivers by hand	GPT model outputs a data-element list automatically on prompt (Claim 8)
Draft the fix	Actuary specifies the change; a developer codes it	Code LLM writes the new model software template (Claim 1)
Test the fix	QA and actuarial review in a staging environment, often weeks	System builds a simulation environment from historical record data and executes the template (Claim 1)
Approve the fix	Peer review, sign-off, rate filing where the state requires prior approval	Not required by Claim 1; optional alert/prompt features appear only in Claims 6-8
Deploy the fix	Scheduled release to production	System generates a production model directly from the updated template (Claim 9)

Every human checkpoint in the middle column is optional in the right one.

The pace mismatch is concrete. Milliman put the countrywide four-quarter rolling average for personal auto rate filing approval at 67 days, with California at 234 days. The Insurance Research Council found the average climbed from 39 to 54 days between 2010 and 2023, while filings withdrawn rather than rejected grew from roughly 1,900 to 3,200 a year (Insurance Business America). Guidewire puts the gap between the data an insurer starts from and the day a new rate reaches a policyholder at up to 18 months.

Claim 10 makes that gap operational. Training the GPT model per jurisdiction means nothing stops a carrier regenerating California rating logic on one internal cycle and Texas logic on another, each time claim 5's threshold trips, with no tie to any filing calendar.

A rate filing governs the rating factors and relativities an insurer may charge, not the software architecture that derives them, so a self-updating engine does not violate a prior-approval statute by existing. But a system that identifies a miss, rewrites the code and validates the rewrite inside hours can produce a materially different rating algorithm before California's 234-day median review of the version it replaces clears the queue.

Run it forward. A Texas book trips claim 5's threshold, the GPT model retrains on Texas data, the code LLM rewrites the sub-model, simulation validates it against Texas loss history, and claim 9 pushes production, all while the carrier's most recent Texas filing sits in the queue behind carriers that never touched their algorithm.

Who Certifies a Rate the Model Wrote Itself

The loop is close to what regulators have spent three years asking for. The NAIC's Model Bulletin, adopted December 2023, requires a written AI Systems Program covering validating, testing and retesting to assess generalization of outputs upon implementation, and twenty-nine jurisdictions now regulate insurer AI use.

The tension is that "testing and retesting," in a bulletin written around human-reviewed governance, assumes a defined model version sitting between tests. A system whose testing step also rewrites the thing being tested produces a moving target instead.

Model risk management has the same dependency. It validates one frozen version: assumptions documented, a back-test, a sign-off, a version number a rate filing can cite.

Grant Thornton's 2026 survey found 44% of insurance executives saying governance challenges had already contributed to an AI project failing, before any model in the sample rewrote its own architecture, and fewer than 25% of enterprises run a structured MRM framework built for generative AI. Federal bank model-risk guidance does not extend to insurers at all, a gap covered here directly.

Claim 8's data elements add a second exposure. A model regenerating its rating code against those four inputs, with no defined stopping point, can drift toward a proxy variable without passing through the bias review a human-supervised process imposes. The unintended-bias problem for static models multiplies untested code paths without multiplying audits.

That lands on a signature. Most states require the certifying actuary to attest that rates are not excessive, inadequate or unfairly discriminatory, tied to a documented rating algorithm as of the filing date. Claim 1 does not require a human to review, let alone certify, the template the code LLM generates before claim 9 pushes it live.

An examiner auditing a filing built on this architecture will ask for the version history of the rating algorithm in effect on the date of loss, the same request examiners make today. The patent does not describe how to answer it. It describes a system built so the question no longer has a single-number answer.