The 3.5/7.5/11.5 Patent Fee Clock Starts at Grant, Not Filing

More than half of patented American invention quietly dies on a calendar almost nobody reads. Roughly 54% of issued US utility patents never see their first maintenance fee paid — a bill that comes due at the 3.5-year mark counted from grant, not filing — and the pattern is visible in the USPTO's own fee records. Teams that anchor budgets to the filing date meet the obligation a year late, inside a grace window already collecting surcharges.

Auto-renewal looks like diligence and behaves like negligence. Flipping every asset onto autopay converts each ambiguity into a paid liability by default, funding patents that no register ties to a named owner or a live value line. The professional default runs the other way: deliberate, documented lapse unless the license register produces both. Where the evidence is missing, the correct calendar entry is an explicit no — signed, dated, and auditable.

From fiscal 2026, the pricing turns annual: the USPTO can revisit maintenance-fee rates every year, with each proposed schedule exposed through a 30-day Federal Register comment window before it locks. Autopay portfolios absorb those increases silently; governed portfolios re-underwrite them line by line. By the 11.5-year window, roughly 91% of grants have already gone dark — proof that lapsing is normal, and that the only failure worth auditing is the unsigned kind.

The 3.5/7.5/11.5 Machine

The clock starts at grant, not filing — and most maintenance-fee errors trace back to getting that anchor wrong. According to the USPTO, utility-patent maintenance fees fall due at 3.5, 7.5, and 11.5 years measured from the patent's grant date, payable any time inside years 3–3.5, 7–7.5, and 11–11.5, with a six-month grace period pushing the final payment opportunity out to years 4, 8, and 12. Key the windows to the application date instead — the classic docketing shortcut — and every alert in your system aims at the wrong year, because the filing-to-grant gap is rarely trivial.

(The small- and micro-entity lifetime figures are the straight sum of their three gates.) Run the 3x test against your own tier's column, not the large-entity one — a micro entity clearing 3x is a very different bar than a large one.

Two scope rules decide who belongs on the triage list at all. Maintenance fees attach to utility patents only — design and plant patents run their full term with no annuities, so they never touch this machine regardless of value. And entity status is not a lifetime credential: small and micro discounts must be re-verified against current revenue and headcount criteria at every window, or the discounted rate is invalid. A team that qualified as small at grant and has since grown owes large-entity money at the next gate; paying the stale tier doesn't bank savings, it accrues a deficiency the USPTO will chase later.

Execution lives in the docketing stack. Vendor platforms — Clarivate, Anaqua, Dennemeyer — fire reminder cascades at T-12, T-6, and T-3 months before each gate, plus a 30-day grace alert, and they ship from the factory with "auto-pay all" enabled. That default is the whole problem: it silently converts what should be a per-patent triage decision into a blanket renewal habit, paying invoices no license register ever approved and running no value check at all. The machinery renews everything by inertia at precisely the moments the fee ladder was built to demand judgment. Before the next cascade fires, pull the auto-pay setting for the coming window's batch and make the exception list explicit — one checkbox currently stands between your portfolio and deliberate lapse as the default.

Roughly half of all issued US utility patents never receive a single maintenance-fee payment, and only about a third stay paid-up through full term. Those figures come from USPTO maintenance-fee payment records surfaced through the PatentsView database, and they reframe the triage question entirely: deliberate lapse is not an aggressive cost cut — it is the modal outcome the market already produces on its own. A default-to-lapse policy stops fighting the base rate. And the curve never bends back up, because a patent that slips past the six-month grace period is gone permanently; the USPTO operates no restoration path for maintenance-fee expiries, so "catching up later" is a feature the fee system simply does not have.

Why is it legitimate to read a lapse curve as a value curve? Because Mark Schankerman established the econometrics at RAND in the late 1990s: his renewal models show the lapse hazard spiking sharply at each fee step, with observed renewal decisions serving as a proxy for the patent's private value to its holder. The mechanism is revealed preference — when the price of continued exclusivity steps up, the holders who pay are the ones whose expected returns clear the new price. The fee ladder forces every owner to re-underwrite the asset three times, and the aggregate answers form the survival curve.

The Half-Life Evidence

Peer-reviewed research by Gaetan de Rassenfosse and Bruno van Pottelsberghe on patent fee systems confirms the design intent holds in practice: rights-holders systematically shed low-value patents as renewal fees escalate. The stepped ladder functions as a self-selection filter, not a revenue trap. Your renewal decisions are made inside a system engineered to separate keepers from ballast — the defensible default aligns with the filter rather than against it.

One input to that filter moved on October 1, 2025. The Unleashing American Innovators Act of 2022 authorizes the USPTO to adjust fees annually for inflation beginning in fiscal year 2026, ending the era of predictable step-changes. Every flat-fee assumption built into existing annuity spreadsheets — the ones that model the three fee gates profiled above as fixed rungs — is now stale. The 3x value multiple in the decision rule must be recomputed against each current-year fee schedule, because the denominator now moves every fall.

The correlation triage teams actually exploit comes from the renewal literature built on Schankerman's framework: payment propensity tracks forward citations and patent-family size. That yields a two-field screen to run before any window. Pull forward-citation counts and foreign-family membership for each asset approaching its 3.5-year mark. A US-only patent with zero citations and no foreign family members by year 3 is empirically the profile the market itself most reliably lets die — the register review becomes pro forma, and the lapse default writes itself unless a live license line overrides it.

The immediate action is mechanical: add citation-count and family-size columns to the register extract you already pull ahead of each window, and let zero-zero entries route straight to lapse. The survival data says the market made that call long before your committee meets.

A patent never earns its own renewal — a register entry does. At every 3.5-, 7.5-, and 11.5-year window described above, the defensible triage question in 2026 is not "is this patent valuable?" but "can the license register produce two specific lines?" If it cannot, the asset lapses on schedule, and no volume of attorney conviction overrides that.

The gate is a two-line test, and both lines must resolve to named entries. Line one: a named internal owner — a person whose name resolves in the HR system, not a department, a cost center, or an "IP team" alias — accountable for the asset. Departments fail structurally: when a reorg dissolves the group, the attestation dies with it and nobody notices until the fee notice surfaces. Line two: at least one live value line — a paying licensee under an active agreement, a currently shipped product covered by a live claim, or a documented freedom-to-operate dependency signed off by counsel. A register entry reading "licensed under the 2019 agreement" passes only if that agreement is still live and royalties still post; "product team confirms coverage" fails, because that is an assertion, not an entry. One line without the other is a fail: a healthy royalty stream bolted to an orphaned asset is precisely the profile that auto-renews for years while nobody can name who last decided.

Owner-or-License

Line two carries a number: renew only when attributable annual value — royalty income plus counsel-quantified FTO avoidance, summed line-by-line from the register — reaches 3x the specific fee falling due at that window. The multiple is deliberately harsh because it prices the asymmetry between the two errors. An unnecessary renewal wastes money recoverably: decline the next gate and the loss stops growing. A wrong lapse is irreversible: the USPTO operates no restoration path once a maintenance fee expires past its grace cliff. A 1x or 2x asset looks affordable and is still a bad buy once you price the cash and the docketing attention it consumes for another cycle.

Scored honestly, the three dispositions compare like this:

The winner is explicit: owner-signoff renew sitting on top of a lapse default. Auto-renew wins on convenience and loses everywhere it matters — it converts a governance decision into a vendor subscription and leaves no artifact an auditor can interrogate. Pure lapse is clean but discards the minority of assets that genuinely clear the bar. The hybrid keeps lapse as the factory setting and makes renewal an earned exception with paperwork attached.

Then put the decision on a clock that ends before the safety-valve temptation begins. Register review opens 120 days before each window; owner attestations close at 90 days out; counsel files either the renewal order or the lapse memo at 60 days out. That leaves the entire grace period unused by default — which is the point. The six-month grace window reads like slack and behaves like a toll road to a cliff: paying late buys a surcharge for the identical outcome, and letting the window close kills the patent permanently, with no restoration form to file afterward. Reserve those months exclusively for a scheduled licensing negotiation that genuinely needs the time — never as buffer for a missed internal deadline.

One lane bypasses the table entirely. Standard-essential families carrying FRAND commitments, and patents embedded in interoperability or open-source pledges, renew on obligation grounds regardless of the 3x math, because lapsing them would contradict the entity's own declared licensing position and hand competitors a ready-made argument. Sequence matters here: verify obligation status against the relevant IPR declaration — the ETSI IPR database for cellular families, for instance — before the two-line test is applied, because an essential patent that fails the value test is still not yours to abandon. Skipping that check is how organizations accidentally de-pledge themselves mid-standard-generation.

The concrete move: open the register now and add two mandatory fields before the next 120-day review opens — owner-of-record as a resolvable person, and value-line type restricted to licensee, shipped-product, or FTO-dependency. Any asset carrying an empty owner field, or an assertion where an entry should be, goes into the lapse pile that day, not at the window.

Every figure in this guide traces back to a single instrument: the USPTO's public maintenance-fee transaction record. According to that record's own structure, it captures only that a fee was paid, on which date, at which entity-size rate — nothing else. It cannot distinguish a deliberate lapse ratified in a governance meeting from a lapse caused by a docketing reminder that fired into a decommissioned inbox. Those two events produce identical rows, which means the half-life curves covered above blend strategy and negligence into a single line. Read them as a floor, not a forecast.

The contamination runs in both directions. The cohort that stays paid through full term includes patents renewed by blanket portfolio policy — auto-renewal across an entire family with no register entry behind it — so survivors overstate how many renewals were actually earned. And because payment history is all anyone observes, the counterfactual stays invisible: no record shows the licensing revenue a lapsed patent would have produced had someone paid. Any threshold calibrated on observed payments inherits both biases.

What the Data Doesn't Tell You

Variance across cases is wider than any aggregate curve admits. The same final gate that is nearly automatic for a monoclonal antibody whose regulatory exclusivity typically outlasts the fee schedule can be a genuine coin flip for a networking patent sitting in an active standards working group, and a clear lapse for a consumer device already designed around. Entity status moves the arithmetic too — reduced rates and the surcharge structure mean the same nominal gate carries different break-even economics depending on who holds the title. The direction of the correction is usually predictable; the magnitude is not.

In four of the five rows the fix is identical: convert an open-ended renewal into a dated, event-triggered payment. Even a register that clears the owner-plus-value test set out above can mislead at these boundaries — the check is necessary, but at the edges it needs a second input.

One limit deserves its own warning, because it inverts a comfortable assumption. The post-due-date grace window is not a safety net. Paying inside it repurchases the patent only at a surcharge — a penalty for lateness, not insurance — and once that window closes unpaid, the patent dies permanently; the USPTO operates no restoration path for maintenance-fee expiries. That asymmetry is precisely why the default-to-lapse rule is built the way it is: a wrong lapse costs one bounded, known amount, while reflexive renewal compounds at every gate. Where the rule strains is confined to the table above — and even there, the defensible move is a conditional payment tied to a dated event, never an open-ended bet.

The practical close: before each gate, run a delta reconciliation. Pull every license contract and product-line change signed since the last register review and match them against the patents entering a window. Anything in the delta set gets a named-owner review even if the standing register reads empty; everything else defaults to lapse, exactly as the rule prescribes. The data supplied the rule — register discipline is what keeps it honest at the edges.

Blend every field together and the corporate lapse rate becomes the most misleading number in the portfolio review. Slice the same USPTO payment record introduced earlier by technology class and the average dissolves: pharmaceutical and biotech estates renew near-universally through the 11.5-year gate because the patent is the product, while software and electronics books routinely shed the majority of their filings by the first 3.5-year window. If your portfolio spans both worlds, the blended figure tells you almost nothing about the single window decision sitting in front of you this quarter.

The deeper problem is what the record cannot contain. Because the USPTO operates no restoration path for a maintenance-fee expiry, the file holds zero observations of "lapsed, then regretted" — the counterfactual value of every abandoned patent is unobservable, so any lapse rule calibrated purely on historical renewal curves structurally ignores tail-value regret. Practitioners tolerate this gap largely because of a comfortable myth: that the six-month grace period is a safety net waiting to catch a missed window. It is not. Paying late costs a surcharge for the privilege, and missing the grace cliff kills the patent permanently — which also means the record contains almost no "lapsed, then rescued" rows to learn from, either.

What the Curves Hide

The counter-evidence cuts against clean triage as well. Citation analyses of USPTO data — the renewal-and-citation literature descended from Ariel Pakes and Mark Schankerman's modeling — find that abandoned patents keep accumulating citing activity for years after expiry. "Nobody cites it yet" at year 3 is therefore a noisy predictor, and some of the highest ex-post-value patents sat squarely in the lapse cohort when the decision was made. The curve records who paid; it never records who should have.

A staleness confound hides inside corporate lapse patterns, too. Departed inventors, renamed product lines, licenses that expired on paper but continued informally — each converts an invention-value signal into a bookkeeping artifact. Anyone who maintains a license register recognizes the failure mode: a product line gets folded into a new division, the register mapping breaks, and the renewal question routes to nobody. Observed renewal data measures administrative hygiene at least as much as invention value.

That leaves the 3x rule's softer input. License-or-product value can be observed on the register; freedom-to-operate avoidance value is asserted, never observed — no ledger records the lawsuit that didn't happen. Counsel's estimate carries wide error bars, so take any borderline candidate whose memo lands exactly on the threshold and run a ±50% sensitivity sweep on that single input:

Most borderline renewals flip to lapses the moment the assertion moves down even one notch, so run the sweep before accepting any borderline yes. If a renewal survives only at the top of counsel's range, the register-governance default — deliberate lapse — stands, and the decision was never a legal judgment call to begin with.

Rule 1 — Default to lapse, and configure the queue to prove it. The triage decision is made the day someone sets up the docketing system, not at the window: every approaching 3.5-, 7.5-, and 11.5-year date should render as "pending lapse," never as "fee due." A fee-due task invites a payment; a pending-lapse task demands a written reason to stop one. Conversion to renewal fires on a single trigger — a named-owner attestation entered in the license register at least 90 days before the due date. That threshold equals one full quarterly reconciliation cycle, so an attestation can cite only evidence compiled this quarter, never folklore carried over from the last docketing migration.

Rule 2 — Apply the 3x gate mechanically. Attributable annual value — royalties actually deposited per the AP ledger plus counsel-quantified freedom-to-operate avoidance — must reach three times the specific invoice sitting in front of you, recomputed fresh from this cycle's register entries. Not last cycle's invoice; not the sponsor's recollection in the review meeting. Counsel's FTO figure moves with claim scope and design-around cost, so require the conservative end of their range to clear the gate. A marginal pass is a fail.

The April 2019 Connector Pair

Rule 3 — Never buy indecision twice. Kill the comfortable myth that the six-month grace period is a safety net; it is a toll booth. Under the USPTO's maintenance-fee rules, late payment repurchases the same patent at a surcharge, and a miss past the grace cliff ends the patent permanently — the office runs no restoration path for maintenance-fee expiries. So the surcharge may be paid exactly once per window, and only against a scheduled, named event: an active licensing negotiation with an identified counterparty documented in the register. Never as a standing think-it-over buffer across the portfolio. A portfolio-wide grace habit is indecision bought twice — once in surcharges, once in a queue that stops ever forcing a decision.

Rule 4 — Weight the three cuts differently. The 3.5-year gate carries the smallest invoice of the three, which makes it the cheap portfolio-shaping cut: owner evidence is nearly all you have, so anything without a named owner goes. By 7.5 years, three years of shipment, license, and citation data exist — the evidence-based cut, where the 3x gate finally runs on actuals rather than projections. The 11.5-year gate is the strategic-asset cut, reserved for patents still tied to shipping products or active license lines; past litigation glory earns nothing there.

Rule 5 — Reconcile the register to reality quarterly. Diff the license register against current shipped-product BOMs and the AP ledger's actual royalty payments. Every entry must survive a three-way match: named owner in the register, the patented technology on a current BOM line, money actually moving through AP. An entry failing any leg is dead weight at the next window, whatever its provenance.

Run Rule 5 this week, before the next gate lands: pull the register, execute the three-way match, and stamp every patent inside one quarter of a window as pending lapse. From that point, each 2026 window asks only the question the register has already answered — the queue decides, not whoever happens to be holding the invoice under deadline pressure.

Close the loop while the artifacts are warm. The renewal order, the lapse memo, and the surcharge receipt go into the license register with dates attached — including the $18,000-to-$24,000 step-up that justified the grace premium. When the 11.5-year window opens in October 2030, the reviewer inherits a clean decision trail and re-runs the same 3x test against fresh register lines, instead of re-deriving the entire case from scratch.

How to Choose Well

Rule 1 — Default to lapse, and configure the queue to prove it. The triage decision is made the day someone sets up the docketing system, not at the window: every approaching 3.5-, 7.5-, and 11.5-year date should render as "pending lapse," never as "fee due." A fee-due task invites a payment; a pending-lapse task demands a written reason to stop one. Conversion to renewal fires on a single trigger — a named-owner attestation entered in the license register at least 90 days before the due date. That threshold equals one full quarterly reconciliation cycle, so an attestation can cite only evidence compiled this quarter, never folklore carried over from the last docketing migration.

Rule 2 — Apply the 3x gate mechanically. Attributable annual value — royalties actually deposited per the AP ledger plus counsel-quantified freedom-to-operate avoidance — must reach three times the specific invoice sitting in front of you, recomputed fresh from this cycle's register entries. Not last cycle's invoice; not the sponsor's recollection in the review meeting. Counsel's FTO figure moves with claim scope and design-around cost, so require the conservative end of their range to clear the gate. A marginal pass is a fail.

Rule 3 — Never buy indecision twice. Kill the comfortable myth that the six-month grace period is a safety net; it is a toll booth. Under the USPTO's maintenance-fee rules, late payment repurchases the same patent at a surcharge, and a miss past the grace cliff ends the patent permanently — the office runs no restoration path for maintenance-fee expiries. So the surcharge may be paid exactly once per window, and only against a scheduled, named event: an active licensing negotiation with an identified counterparty documented in the register. Never as a standing think-it-over buffer across the portfolio. A portfolio-wide grace habit is indecision bought twice — once in surcharges, once in a queue that stops ever forcing a decision.

Rule 4 — Weight the three cuts differently. The 3.5-year gate carries the smallest invoice of the three, which makes it the cheap portfolio-shaping cut: owner evidence is nearly all you have, so anything without a named owner goes. By 7.5 years, three years of shipment, license, and citation data exist — the evidence-based cut, where the 3x gate finally runs on actuals rather than projections. The 11.5-year gate is the strategic-asset cut, reserved for patents still tied to shipping products or active license lines; past litigation glory earns nothing there.

Rule 5 — Reconcile the register to reality quarterly. Diff the license register against current shipped-product BOMs and the AP ledger's actual royalty payments. Every entry must survive a three-way match: named owner in the register, the patented technology on a current BOM line, money actually moving through AP. An entry failing any leg is dead weight at the next window, whatever its provenance.

Run Rule 5 this week, before the next gate lands: pull the register, execute the three-way match, and stamp every patent inside one quarter of a window as pending lapse. From that point, each 2026 window asks only the question the register has already answered — the queue decides, not whoever happens to be holding the invoice under deadline pressure.

What to do next

Frequently Asked Questions

If my company misses a maintenance fee payment, can we still catch up and pay it later?

No — a patent that slips past the six-month grace period is gone permanently, because the USPTO operates no restoration path for maintenance-fee expiries.

When exactly do the three maintenance fee payments come due?

Utility-patent maintenance fees fall due at 3.5, 7.5, and 11.5 years measured from the patent's grant date, payable any time inside years 3–3.5, 7–7.5, and 11–11.5, with the six-month grace period pushing the final payment opportunity out to years 4, 8, and 12.

Do design patents or plant patents ever owe maintenance fees?

No — maintenance fees attach to utility patents only, while design and plant patents run their full term with no annuities regardless of value.

What does it cost to keep a utility patent paid up through its entire term?

All three gates combined total $6,720 for a large entity, $2,688 for a small entity, and $672 for a micro entity.

Once we qualify for small- or micro-entity rates, do those discounts apply for the patent's whole life?

No — entity status is not a lifetime credential, so small and micro discounts must be re-verified against current revenue and headcount criteria at every window, and paying a stale tier accrues a deficiency the USPTO will chase later.

Can I lock in current maintenance fee amounts in my budget spreadsheet going forward?

No — beginning fiscal 2026 under the Unleashing American Innovators Act of 2022, the USPTO can revisit maintenance-fee rates every year, with each proposed schedule exposed through a 30-day Federal Register comment window before the yearly adjustment locks.

Quick answers

Also worth reading: M&A Recordal: USPTO 2 Weeks vs EPO 6 Months—No Mere Formality: M&A Recordal: USPTO 2 Weeks · API Triangulation and Clearance Velocity Data for IP Teams: API Triangulation and Clearance Velocity