Cost-exchange dynamics between offensive drone systems and counter-drone defense, and whether the asymmetry is actually closing.
Since February 2026, Iranian retaliatory strikes have forced the United States and its regional partners to burn through an estimated 800 PAC-3 and THAAD interceptors, missiles that cost the Army $4.2 million apiece, defending against drones and missiles that, in the case of the Shahed-136, cost roughly $35,000 to build. The exchange rate has not been close. A loitering munition with a moped engine and a GPS chip is being traded against a missile that took a defense contractor years to design and validate, tens of times over, in a matter of weeks.
Russia proved the asymmetry could bleed a defender dry over years. Iran proved it could do the same to the United States in thirty-nine days. The economics become stark once you set production cost against cost-to-kill. At a baseline Shahed-136 cost of $35,000 against a PAC-3 MSE cost of $3.7 million, the production-cost ratio is roughly 106 to 1. And that interceptor is not a guaranteed kill: Patriot batteries have repeatedly fired multiple missiles at a single incoming drone, a pattern that has strained inventories badly enough to force a pause in U.S. arms shipments to allies, including Taiwan. On June 11, 2026, President Trump invoked the Defense Production Act to accelerate munitions output after the 39-day war with Iran (which began February 28, 2026) left the U.S. having expended an estimated 2,000 Patriot, THAAD, and SM-series interceptors. Lockheed Martin, the sole producer of the PAC-3, has committed to scaling annual production from roughly 620 missiles in 2025 to 2,000 by the end of 2030. If that target is reached, sustaining 2,000 PAC-3s a year would cost the U.S. an estimated $7.4 billion annually: 2,000 missiles at $3.7 million each, by this report's own calculation. Redirected at Shahed-136 unit costs, that same funding could field more than 211,000 drones. The PAC-3's economics make sense against a target like a $45 million Su-57 fighter. Against a disposable loitering munition, they barely make sense at all.
The cost-asymmetry problem first came into focus during the opening stages of the Russo-Ukrainian War. Drone usage was sparse through the spring and summer of 2022, limited to small, isolated reconnaissance operations by both sides. That changed in the fall of 2022, when Russia began launching large-scale Shahed-136 ("kamikaze drone") strikes on Ukrainian cities, with the first reaching the Kyiv region on October 5, when a wave struck Bila Tserkva. Russia soon began domestic production of the drone under the designation Geran-2, scaling its use into the near-daily and weekly waves that have defined the war since.
In those early waves, Ukraine's only real answer was expensive: Shaheds and Geran-2s were shot down using MiG-29 fighters and PAC-2/PAC-3 interceptor missiles costing millions of dollars apiece, which is the same cost-exchange problem now visible in the 2026 Iran conflict, years earlier and at smaller scale. As the war ground on, that math became untenable, and Ukraine built a cheaper answer from scratch: domestically produced interceptor drones like the Wild Hornets' Sting, which costs roughly $2,500 per unit and has destroyed thousands of Shahed and Geran-series drones since its first combat use in 2025. Ukraine is where the cost-exchange problem first became impossible to ignore, and where the first real solution to it was built.
This report tracks how that ratio has moved from 2022 to 2026, and builds a breakeven model to find the point where intercepting every drone stops making economic sense.
What follows is built in five stages. The next section lays out this report's methodology — how cost-exchange ratios are defined, which sources are treated as authoritative, and where the data runs thin. From there, the report tracks cost figures across multiple systems on both sides — the Shahed-136 as a fixed offensive benchmark against a range of defense systems, including kinetic interceptor missiles, dedicated interceptor drones, gun-based point-defense, electronic warfare, and emerging directed-energy systems — year over year from 2022 to 2026. A breakeven model then turns that data into two thresholds: the point at which intercepting a drone costs more than the drone itself, and the point at which it costs more than the thing being defended. The final sections state what those findings imply for procurement and force design, and where the data behind them runs out.
The asymmetry gap is closing, however almost none of the closing came from the systems the question is typically asked about. Measuring cost-per-kill rather than its listed price, which accounts for interceptors spent on misses, a PAC-3 engagement runs about 130 times the cost of the drone it destroys. Across the five systems with enough data to trace a trend from 2022 to 2026 only the PAC-3 got cheaper. What altered the asymmetry gap was not the cheapening of existing systems but the introduction of new, more cost-effective ones. Systems fielded in the last two years reach cost-per-kill figures in the hundreds and low thousands of dollars, three to four orders of magnitude below the interceptors they sit beside, and they do it by making the consumable cheap rather than the platform cheap. Adding the value of what is being defended narrows the verdict even further: missile interceptors remain economically coherent against ships, aircraft, and armor, and stop being coherent against the light vehicles and infrastructure that drones actually hunt. The failure is not that interceptors are expensive, instead the failure is the pairing of defenses with the posed threat, and that pairing gets worse as the threat gets cheaper.
The cost asymmetry problem this report investigates is not new. It has become one of the most frequently invoked frames in defense talking points since the start of the Russo-Ukrainian War, and especially since the 2026 Iran conflict pushed it into mainstream reporting. But a survey of how major-defense outlets have actually treated the problem reveals a consistent trend: rich in illustrative comparison, thin on systematic data. What follows is a short survey of existing defense-research commentary, to show precisely where it ends, and why this report picks it up from there.
CSIS's Wes Rumbaugh offered the most direct challenge to the cost-exchange framing in a 2024 report on the cost and value of air and missile defense intercepts. His argument: examples like a $2 million missile interceptor downing a $2,000 Houthi drone make for great headlines, but they're misleading, since missile interceptor costs have historically decreased over time on an inflation-adjusted basis. Rumbaugh also pointed out that cost-per-shot ignores the value of what's actually being defended: an interceptor that costs more than the drone it kills can still be worth it if what it's protecting is worth far more than either.
CSIS's Benjamin Jensen and Yasir Atalan take a more descriptive approach in their report on Russia's Shahed campaign. They document Russia's tolerance for high drone loss rates exceeding 75 percent, treating attrition as an acceptable cost of gradually overwhelming Ukraine's air defenses through sheer volume. Ukraine's scaling of low-cost drone interceptors and point-defense systems helped maintain a more favorable cost-exchange ratio, but at a steady cost: depleting Western-supplied missile interceptors that aren't easily replaced.
RUSI research fellow Robert Tollast's commentary on the Iranian drone threat argues the cost-asymmetry problem isn't new at all. It goes back to 2017, when Iran-aligned forces in Syria deployed a Shahed-129, a precursor to the Shahed-136, against U.S. forces, and the U.S. shot it down with an F-15 fighter jet. That's a real cost mismatch in its own right, given the high hourly cost of flying that aircraft. The Houthi blockade of the Red Sea made the problem more concrete and expensive. Lacking cheap interceptors, U.S. and British forces fired expensive missiles at cheap Houthi drones, with the British Royal Navy in one instance firing Aster 15 missiles, worth roughly $1.3 million each, at low-cost Shaheds. Tollast's broader argument is that despite nearly a decade of warning signs from Syria, Yemen, and Ukraine, the U.S. and UK have mostly responded to the Iranian threat with high-cost, low-quantity solutions rather than the cheap, scalable interceptors already proven out in Ukraine. Even where the UK has shown interest in co-producing Ukraine's Octopus interceptor, the piece argues hardware alone won't close the gap without enough radar coverage and networking to support it.
War on the Rocks provides a contrasting argument from author Vitaliy Goncharuk, a Ukrainian-American entrepreneur who chaired Ukraine's AI Committee and now runs a company that builds autonomous targeting systems. Goncharuk's argument runs on the idea that the propeller-based drone interceptor model employed in Ukraine is becoming obsolete as Russia develops turbojet-powered drones to circumvent the lower speeds of propeller-based defenses. Russia's fitting of its Geran-2 drones with turbojet engines gives the drones a speed upgrade from 90 to 460 mph, and an altitude upgrade from 6,500 feet to 29,000 feet. Iran has built its own answer on the defensive side, the 358 missile, priced at around $90,000, which can hit Shahed-class drones and larger aircraft alike. Goncharuk argues the West is making the same mistake twice, pouring money into propeller drones that can't keep pace with a faster generation of threats. His own fix is a cheap, AI-guided autonomous interceptor missile, which is worth reading carefully given that his company builds exactly that kind of targeting system. Bias aside, his core point still lines up with what other sources in this section describe: the threat is evolving faster than the cheap counter-drone response built to stop it.
MWI's report by Matthew Revels and Eric Uribe is the most discernible angle of any source in this section, framing the problem as doctrinal, not economic. They argue that the U.S. military should not try to copy Russia or Ukraine's drone tactics of attrition, even though both have proven effective against cheap loitering munitions. Their reasoning derives from Michael Horowitz's adoption-capacity theory, which states that a military only adopts an innovation if it fits the organization's existing way of fighting, not just because the innovation works for someone else. Adoption depends on cost and fit, and the authors argue drone-heavy attrition warfare fits Russia's military culture, but not the Army's. Russia's drone use modernizes an artillery-centric, attrition-based doctrine it has followed for decades. The U.S. Army's doctrine centers on airpower-enabled maneuver, the kind of fast armored advance that defined the Gulf War's "Left Hook." Revels and Uribe argue that copying Russia's drone playbook would actually undercut the Army's real advantage rather than restore it. Their alternative is counterinnovation: building counterdrone air defense, electronic warfare, and deception capable of neutralizing the threat rather than mimicking it. For this report, it's a useful warning not to automatically look at a favorable cost-ratio on paper and declare it a good fit for every military's system.
Of the sources surveyed, Rumbaugh's analysis comes closest to tracking an evolving trend rather than a moment, with his multi-year look at interceptor unit costs. But it only tracks one side of the equation, and doesn't account for the cost-exchange of other effective alternatives. It shows interceptor costs falling over time, but never sets that trend against the falling cost of the drones interceptors are built to stop. Jensen and Atalan describe only one campaign in depth, and Tollast traces a decade of warning signs without connecting them into a single measure. Goncharuk identifies one technological shift, but his own proposed alternative, a cheap AI-guided interceptor missile, comes with no real numbers behind it, since he admits it doesn't yet exist at scale. None of them builds a cost-exchange ratio that moves across years, compares more than one defense category at a time, or asks the question this report is built to answer: at what ratio does intercepting a drone stop being worth the cost. That question is where this report begins.
When comparing the costs and capabilities of different systems, like a drone or a missile, the numbers must be constructed through the same process. A favorable cost ratio means very little if one side of it was measured loosely and the other wasn't. This section lays out the parameters this report follows to make comparisons fair: how a system's cost counts as confirmed, which sources get trusted over others, and how a single missile fired once gets compared fairly to a jammer that can be used a thousand times. The consistency of sourced data, applied across multiple systems and years, is what was missing from the commentary surveyed in the last section. The rest of this section is how this report tries to close that gap.
This report will hold the offensive side of the equation constant, using the most prevalent example of an adversarial UAS: the Shahed-136 / Geran-2 drone. Each defensive system in the data section will be measured against this same example, to keep the comparison controlled and to anchor a realistic breakeven point. A report that compared one drone to a PAC-3 and a different drone to a jammer wouldn't be measuring defense cost-effectiveness on a consistent basis, and would further complicate the breakeven point for different defense systems. The Shahed-136 is also the right system to fix it to, since it's the same drone nearly every source surveyed in the last section already used as its own reference point. Anchoring to it keeps this report's numbers internally consistent, and comparable to the commentary that came before it.
This report ranks sources in four tiers. Official government and procurement data comes first: defense budget documents, Pentagon contract announcements, and similar primary records. Where official figures aren't public, the report turns to think tank analysis from outlets like CSIS, RUSI, and RAND, organizations with a track record of citing primary sources themselves. Below that is defense trade press, outlets that often report figures faster than think tanks but with less methodological transparency. Company claims, a manufacturer's own stated price or capability, sit last, since they carry an obvious incentive to favor their own product. When two sources at the same tier disagree, this report defaults to the more recent figure and notes the discrepancy if it's large enough to matter. Every cost figure in this report is converted to 2026 dollars using the U.S. Consumer Price Index, so a number reported in 2022 and a number reported in 2025 can sit next to each other honestly instead of just being placed side by side. This is a general adjustment, not a defense-specific one. Actual procurement costs for any single system can rise or fall for reasons unrelated to broader inflation, supply chain constraints, single-source manufacturer pricing, or a new variant entering production, and CPI normalization doesn't capture those system-specific shifts.
Expendable systems are scored against the chance that a single shot succeeds. A missile or interceptor drone that's destroyed in the act of stopping a threat falls into this category, cost divided by the probability that one shot does the job. Reusable systems are scored differently, since the system itself survives the engagement. Its acquisition cost gets spread across every kill it's expected to achieve over its service life, then added to whatever it costs to fire or operate it once. What counts as expendable or reusable comes down to one question: is the system consumed in a successful engagement, not what kind of object it is. An interceptor drone looks reusable but isn't, it's destroyed the same way a missile is. A gun system looks disposable but isn't, the gun survives, only the ammunition gets spent.
This formula also already accounts for systems that need more than one shot to score a kill. A lower probability of kill per shot produces a higher cost-per-kill automatically, without needing a separate term for it. What this metric does not claim to measure is the full strategic value of air defense, only the comparative cost of achieving a kill across different systems. A system can be cost-inefficient by this measure and still be worth deploying, if what it protects is valuable enough. This report is built to answer a narrower question than that, not whether air defense is worth it, but which approach to it costs less for the same result.
Every system in this report's data section needed the same 4 inputs to fit into the cost-per-kill formula: unit or acquisition cost, the probability of a kill per engagement, classification as expendable or reusable based on whether the system is consumed in a successful engagement, and for reusable systems, the number of engagements expected over its service life.
Three of those inputs got more complicated to source depending on the category. A system like Anduril's Roadrunner can abort or miss without being destroyed, so its real unit cost isn't a flat sticker price, rather it's a blended figure that accounts for how often the unit survives to be used again. Gun-based systems and automated weapon stations don't fire one round per engagement, they fire a burst, so their marginal cost has to be built from rounds expended per kill multiplied by cost per round, rather than taken as a single published number. That round count is measured across a whole successful engagement, so the rounds spent on bursts that miss are already inside the figure, in the same way dividing by probability of kill prices re-engagements for an expendable system. Non-kinetic systems like jammers don't engage one target at a time the way a missile battery does, so estimating how many engagements a system sees in a year requires knowing its effective range or coverage area first.
Not every system needed every input, as truly expendable systems like a PAC-3 missile only needed the first three inputs. Reusable systems like the Gepard-style gun needed all 4, plus the rounds-per-burst calculation. It's worth noting that the cheaper and more flexible a defense system gets, the more inputs it actually takes to price it fairly.
This report tracks five categories of defense system, each measured against the same Shahed-136 benchmark. Kinetic missile interceptors, like the PAC-3, represent the most expensive and most proven approach. Interceptor drones, like Ukraine's Sting and Anduril's Roadrunner, are the cheaper answer that emerged once the cost of the first category became unsustainable. Gun-based point-defense systems, like Gepard-style mounts, fire flak ammunition from a tracked platform. Automated weapon stations, like Ukraine's Sky Sentinel and the U.S. Navy's Phalanx, engage autonomously, no operator aims the weapon, whether that autonomy comes from an AI-driven camera system or a radar fire-control loop. The fifth category, non-kinetic systems, covers both electronic warfare and directed-energy weapons under one heading, though the two aren't at the same level of maturity. Jammers are mature and already widely deployed. Directed-energy systems are included as a forward-looking category specifically because the data behind them is thinner than anywhere else in this report.
Ten systems sit across these five categories. Each is tracked year by year wherever the underlying data supports it. Where a system entered service partway through this report's 2022 to 2026 window, it's tracked only from the year it became operational, not backfilled with an estimate for years before it existed. A system with no public cost figure anywhere is left out of the data section entirely, rather than included with a guessed number standing in for a real one.
Kill-probability figures are often estimated or classified rather than directly measured, and even official sources can reflect favorable test conditions instead of real combat performance. This report cannot independently verify any kill-probability figure it cites. Service-life and engagement-rate assumptions for reusable systems are estimates built from the best available public reporting, not measured facts. That uncertainty is not evenly distributed. Where a figure rests on a budget exhibit and a bounded kill probability, as with the kinetic interceptors, the plausible range is roughly ten percent either side of the number shown. Where it rests on a modeled service life, the range is far wider: plausible values for the gun-based systems span a factor of four to five, and Leonidas spans a factor of ten. Every figure in this report is a point estimate drawn from that range, and the reusable ones should be read as an order of magnitude rather than a precise cost.
This report leaves some real costs out by design. False-positive engagements, a jammer or gun system firing at debris, decoys, or false targets, aren't modeled, since reliable data on wasted engagements doesn't exist across enough systems to compare fairly. Detection and cueing infrastructure, the radar and sensor networks that let any of these systems find a target in the first place, isn't included in any per-system figure either, since that cost is usually shared across an entire defense layer rather than attributable to one interceptor. A favorable cost-per-kill ratio doesn't guarantee a system is available at the volume a real attack requires. Lockheed's multi-year timeline to scale PAC-3 production, covered earlier in this report, makes the point directly: cost-efficiency and deployability at scale are not the same finding.
The benchmark itself carries a limitation. Every figure in this report answers what it costs to stop a Shahed-136 specifically, not any drone. A system built and tuned against a slow, propeller-driven threat may perform very differently against the turbojet-upgraded variants already discussed in this report's second section. Directed-energy systems carry the thinnest data of any category for the same underlying reason, the technology is newer than the rest of this report's window, and is included here as a forward-looking category rather than a fully proven one.
Cost-per-kill is not the same as total strategic value. A system can be expensive relative to the threat it stops and still be worth deploying, if what it's protecting is worth far more than either. This report is not built to discuss whether air defenses are worth having, rather it's built to find what approach to it costs less for the same result.
The section that follows applies everything above to ten real systems across five categories. Each gets a sourced unit cost, a kill-probability estimate, and a cost-per-kill figure built from the formula in this section, expendable or reusable, depending on the system's category. Every figure is normalized to 2026 dollars and tracked year by year wherever the data allows it.
This section applies the cost-per-kill formula from the last section to twelve systems across five categories, using unit costs and kill-probability figures sourced from public budget documents, contracts, and vendor disclosures through 2026. Every figure below is normalized to 2026 dollars using the CPI factors described in §03, and each is measured against the same fixed benchmark, the Shahed-136 class attacker cost of $20,000 to $50,000. What follows is the data itself, category by category, and what it shows about how far each approach still sits from parity with the threat it is built to stop.
The chart below plots cost-per-kill for every tracked system at its actual sourced year, on a log scale, in 2026 dollars. Most systems here have only one or two dated cost figures rather than a full five-year run, so each system appears as an isolated point where the record is thin, and a connecting line only where a real multi-year series exists: PAC-3, THAAD, SM-6, Sting, and Leonidas. The bright orange line marks the Shahed-136 breakeven, the roughly $35,000 cost of the attacking drone itself. A point's distance below that line is a rough measure of how far a system has already closed the cost gap with the threat it is meant to stop. Use the category buttons to isolate one part of the picture at a time.
Each system’s Source column names the primary document behind its unit-cost figure, with a colored marker for where that document sits in the sourcing standard set out above. Green marks Tier 1: official Army or MDA budget exhibits, the only figures here that are procurement-grade. Yellow marks Tier 2, derived from a specific contract award or test record. Red marks Tier 3, meaning OSINT, vendor claims, or figures modeled from adjacent data. Hover any Source cell for the full provenance and caveats.
Two things stand out immediately. The cheapest systems in this dataset, camera and radar-directed weapon stations and directed-energy weapons, already sit one to two orders of magnitude below the benchmark. The most expensive, the kinetic missile interceptors that have anchored air defense doctrine for decades, sit two to three orders of magnitude above it. Nothing in the middle bridges that gap smoothly. The rest of this section walks through each category in turn, then compares all five directly in Block 8.
Three systems make up this category: the Patriot's PAC-3 MSE interceptor, the Terminal High Altitude Area Defense (THAAD) interceptor, and the Navy's SM-6 (RIM-174). All three are true expendables, consumed in a successful intercept, and all three are backed by government budget-grade sourcing across the full 2022 to 2026 window.
| System | Unit Cost | Class. | P(Kill) | Source |
|---|
PAC-3 is the cheapest of the three and the only one trending down in real terms, from roughly $5.2M per kill in 2023 to $4.55M in 2026. THAAD peaked in 2024 near $16.3M before easing back toward $14.2M. SM-6 moved the other direction, climbing from about $8.1M in 2022 to $11.7M in 2026 as Navy inventory has shifted toward the newer Block IB variant. Even at its cheapest, this category remains roughly 100 to 400 times the cost of the drone it is built to stop.
Three systems make up this category: Ukraine's Sting interceptor, Anduril's Roadrunner-M, and Fortem's DroneHunter F700. Sting and Roadrunner-M are both expendable in practice, destroyed or expended on a successful kill, while DroneHunter is a net-capture platform modeled here on an assumed 20-sortie airframe life.
| System | Unit Cost | Class. | P(Kill) | Source |
|---|
Sting is the standout of the category, and of the dataset: real 2025 and 2026 quotes put it at roughly $2,500 to $2,900 per kill, already below the Shahed-136 benchmark itself. The step from $2,100 to $2,500 across those two years sits within normal reporting noise rather than a real cost trend. Roadrunner-M is a different system entirely at roughly $587,000 per kill, its cost driven by an airframe that is expended, not recovered, on a successful hit. It is scored on its point-defense kill probability of 0.90, the same basis used for the kinetic interceptors; its published area-defense threshold of 0.70 would put the figure near $754,000. DroneHunter's figure, near $9,900, carries low confidence and depends heavily on an assumed sortie life that has not been independently confirmed.
Two systems make up this category: the Gepard, a 35mm Oerlikon-armed tracked gun reintroduced to Western inventories via Ukraine, and Rheinmetall's Skyranger 30. Both are reusable platforms, so their cost-per-kill blends acquisition cost amortized over expected kills with the marginal cost of every round fired to achieve one kill, misses included.
| System | Unit Cost | Class. | Exp. Kills / Life | Source |
|---|
Gepard comes in around $17,700 per kill, Skyranger closer to $37,000. Both sit inside or just above the Shahed-136 benchmark band, roughly break-even with the threat rather than a clear win. That is a meaningfully different outcome than the interceptor drones above, which sit well under the band, and it is the category where the cost-exchange argument is closest to a toss-up.
Two systems make up this category, spanning very different levels of maturity. Epirus's Leonidas is a fielded high-power microwave system with a real two-year contract-derived cost step. LOCUST X3, a laser system unveiled in 2026, carries a single OSINT-estimated figure and no disclosed unit price.
| System | Unit Cost | Class. | Exp. Kills / Life | Source |
|---|
Leonidas moved from roughly $825 to $900 per kill in 2023 and 2024 to about $1,090 to $1,120 in 2025 and 2026, a real step tied to a new contract tier, not a gradual drift. LOCUST X3 lands close by at roughly $1,000. Both figures are dominated by amortized acquisition cost rather than the cost of firing the weapon itself, since a laser or microwave pulse costs pennies of electricity per shot. This is the cheapest category in the dataset by a wide margin, and the one where cost asymmetry with the Shahed-136 has already inverted most decisively.
Two systems make up this category, grouped by autonomy rather than caliber: Ukraine's AI-and-camera-directed Sky Sentinel, and the U.S. Navy's radar-directed Phalanx CIWS. A hypothetical illustrative comparison, the ACS Bullfrog, is shown alongside them on the chart as a labeled estimate, not a tracked figure.
| System | Unit Cost | Class. | Exp. Kills / Life | Source |
|---|
Sky Sentinel is the cheapest system in the entire dataset at roughly $185 per kill, a .50 caliber platform whose acquisition cost is low enough that amortization barely matters. Phalanx, by contrast, runs closer to $23,000, a much larger and older system whose 20mm rounds and higher unit cost put it in the same rough band as the gun-based category above. The spread inside this single category, over two orders of magnitude between its two tracked systems, is the widest of any category in the dataset.
The Bullfrog is included to show where this category may be heading rather than where it currently sits. Amortizing a $400,000 turret across a modeled service life of roughly 1,500 engagements gives about $267 per kill in platform cost, and adding roughly $11 per kill in standard NATO ammunition, the cost of a ten-round burst, produces a modeled figure near $278. The unit cost is the midpoint of a $350,000 to $450,000 range listed by Joint Interagency Task Force 401, the Defense Department body that has contracted for the system, which is firmer sourcing than the vendor range this report previously carried. The service life is the soft input: expected kills over the life of the mount remains unresolved, and the figure moves to about $411 at a thousand kills and about $211 at two thousand. What the estimate is useful for is the shape of the argument it implies. At this unit cost the acquisition dominates completely: $267 of the $278 is the turret, and the ammunition is the remaining eleven. A cheap consumable is not sufficient on its own, and the question for this category is not whether bullets beat missiles, which they plainly do, but whether the mount can be amortized over enough engagements to matter. Whether the Bullfrog specifically achieves $278 matters far less than whether that structural logic holds, and this report treats it as a hypothesis about the category rather than as evidence about the system.
Laid side by side, the five categories split into three tiers rather than a smooth ladder. Non-kinetic systems and automated weapon stations, at roughly $185 to $1,100 per kill, already cost 20 to 100 times less than the Shahed-136 they are built to stop, a clear inversion of the cost asymmetry that motivated this report. Gun-based platforms, at roughly $17,700 to $37,000, sit inside or just above the benchmark band, rough parity rather than a clean win. Skyranger is the closest call in the dataset: its figure sits just above the Shahed line, but the plausible range behind it runs from well below to roughly twice above, so the data does not resolve which side of parity it falls on. Kinetic missile interceptors, at $4.5 million to $14.2 million, remain 100 to 400 times more expensive than the threat, essentially unchanged in kind since 2022 even where individual unit costs have eased. Interceptor drones split the category itself: Sting sits with the cheapest systems in the dataset, while Roadrunner-M's expendable-on-kill design puts it well over an order of magnitude above its own category-mates, near $587,000.
The reusable categories also divide along a line the labels do not show. For gun-based systems the ammunition dominates: roughly 74 percent of Gepard’s cost-per-kill is 35mm rounds rather than the vehicle, and 57 percent of Skyranger’s is programmable AHEAD ammunition. Directed-energy systems invert this, spending almost nothing per shot and carrying nearly all of their cost in acquisition spread across service life. That distinction sets how much confidence each figure deserves. Where ammunition dominates, cost-per-kill is anchored to an observable price per round. Where amortization dominates, it rests on a modeled service life no system here has yet demonstrated.
The ranking has not reordered so much as it has stretched. The cheap end has gotten cheaper: camera-guided and directed-energy systems that barely existed as fielded programs in 2022 are now the least expensive way to kill a Shahed-136 by a wide margin, while the expensive end has held roughly flat in real terms. The gap between the cheapest and most expensive system tracked here spans five orders of magnitude, and closing it further is less a question of whether cheaper systems work than of whether they can be fielded at the scale and reliability a sustained drone campaign requires.
Cost-per-kill answers which system is cheaper to use once. It does not answer how many of them a defender needs, how fast they can be produced, or what happens when the cheapest systems run out of rounds mid-barrage. The next section builds a breakeven model on top of the figures established here, translating cost-per-kill into a genuine question of scale: given a sustained attack at a given rate, which combination of these systems actually holds, and for how long.
The previous section took the cost-per-kill of a system against a single drone. Measured against the Shahed’s $35,000, most of those numbers look indefensible. A PAC-3 intercept runs roughly 130 times the price of the thing it destroys, and the cheaper categories look reasonable only by comparison. However, the cost ratio is not the only thing driving decisions regarding the elimination of a target. Comparing an interceptor to the drone it kills leaves out the reason anyone fires it, which is not the drone but whatever the drone was flying toward. A missile that loses badly on cost-exchange can still be the correct shot if it saves something worth far more than both. This section builds two thresholds against which that judgment can be made. The first is parity with the attacker, the point at which a defender stops paying more per kill than the attacker pays per drone. The second is parity with the defended asset, the point at which a shot costs more than the damage it prevents. Systems can fail one threshold and pass the other, and where each one falls decides whether cost-asymmetry is a crisis or an accounting issue.
The model that follows is illustrative. It is simple algebra applied to the cost figures in the previous section, not an operations-research treatment of air defense. It is a threshold calculation, built to show where the economic boundaries sit and how far current systems fall from them, rather than to predict the result of any particular engagement. The assumptions are worth stating before the model rather than after it, because three of the four remove costs a defender would actually pay. That makes every defender figure here a floor rather than a best estimate. Where defending still looks irrational under a model built to flatter it, the finding is harder to argue with than it would otherwise be.
The break-even model is deterministic. Every input is a single point value, so it produces one number per pairing rather than a distribution, and it says nothing about the variance a real engagement would show. It assumes no saturation. Magazine depth is treated as sufficient, which is the assumption that lets swarm size cancel out of the arithmetic below. A defender who runs out of interceptors faces a second set of economics, which the first threshold treats separately. The model assumes no false positives. Interceptors spent on decoys, birds, and misidentified contacts are real expenditures that produce no kill, and each one raises the true cost per kill. The model also excludes the kill chain. The unit costs carried forward from the previous section price rounds and platforms, not the radars, command systems, crews, and power generation without which none of them fire.
The first threshold is the one the previous section already implies. A defender spends the unit cost of an interceptor divided by its probability of kill, because the shots that miss are paid for too. An attacker spends the price of one drone. The exchange ratio is the first figure divided by the second, and parity is the point where it equals one.
The arithmetic has a property that is easy to miss and worth stating directly. If an attacker sends a thousand drones instead of one, the defender needs a thousand times as many kills, and both sides of the fraction scale by the same factor. The exchange ratio does not move. A PAC-3 battery firing at a single Shahed loses the trade by the same margin as a PAC-3 battery firing at a swarm of a thousand, which means the ratio is a property of the interceptor and the threat it is matched against, not of the size of the raid.
This runs against the intuition that mass is what breaks a defense, and while that intuition is not wrong on its own, the model is describing a different mechanism. Swarms defeat defenses by exhausting magazines and overwhelming the sensors and crews that assign shots, not by shifting the price of any individual intercept. That is a saturation problem, and this model sets saturation aside. The consequence is that the first threshold cannot be crossed by scaling an attack. A defender whose cost-per-kill sits above the benchmark stays above it at every raid size, and the only way across is a cheaper interceptor, a higher probability of kill, or a more expensive attacking drone.
Setting saturation aside is a real concession, and it is worth being specific about what is lost. A Patriot battery carries between 72 and 128 ready interceptors across its launchers, and at a kill probability of 0.85 an attacker needs on the order of 113 drones to empty one. Those drones cost roughly four million dollars. The magazine they strip is worth close to four hundred million. That leverage, near ninety-five to one, is larger than the per-shot exchange ratio this section measures, and it is the dynamic straining air defense in Ukraine and Israel today. The reason it is not modelled here is not that the equipment figures are unavailable. Magazine depths are published, and the counts above come from Congressional Research Service reporting; Rheinmetall publishes 252 ready rounds for the Skyranger 30, and nightly raid sizes are reported by the Ukrainian Air Force. What is missing is the term that turns a magazine into a defense: how many fire units cover a given area, how they are layered, how many shots doctrine assigns per target, and what reload looks like in place. Those are operational planning factors which are not public. Without them, the arithmetic above describes one battery rather than whether a city holds, which is a different question from the one this section set out to answer.
Consider an Arleigh Burke destroyer, the most-produced large surface combatant in the US Navy, at roughly $2.7 billion for a Flight III hull. A single Shahed leaking through to that ship does not have to sink it to be worth stopping. Firing an SM-6 at the drone costs about $11.7 million per kill and loses the exchange by a factor of well over three hundred, which by the first threshold is among the worst trades in the dataset. Nobody who has thought about it for a moment believes the ship should be left undefended on those grounds. The exchange ratio is answering a question about the drone, and the officer deciding whether to fire is asking a question about the destroyer.
This is the objection Wes Rumbaugh raised at CSIS, noted in §02. Cost-per-shot ignores the value of what is being defended, and an interceptor that costs more than the drone it kills can still be worth firing if what it protects is worth far more than either. Rumbaugh made the argument without putting numbers to it, which is fair enough for a piece making a different point. The previous section measured one side of that trade and left the other unquantified. What follows is an attempt to close it.
That gives a second threshold. Interception pays for itself when the cost of the kill comes in below the damage the drone would otherwise do, which is the value of what is being defended multiplied by the chance the drone actually damages it. Write that chance as phit. The test is per drone rather than per raid, so swarm size cancels here for the same reason it cancelled above.
The two probabilities in this section are easy to run together and describe different things. Probability of kill belongs to the interceptor and is already inside every cost-per-kill figure carried forward from the previous section, since that is what the unit cost was divided by. The term phit belongs to the attacking drone and to whatever it is aimed at, and it asks what happens after the defense misses. A Shahed that reaches a destroyer may hole it above the waterline, may put a radar out of service, or may hit water fifty meters off the beam. Using the interceptor’s kill probability here would price the defender’s accuracy twice and say nothing about the threat.
No source in this dataset supports a value for phit. It varies by warhead, by target, by how the asset is hardened, and by where on the asset the drone arrives, and none of that is published in a form that would survive the sourcing standard applied to everything else here. It is a declared parameter rather than a measured one. What follows sets it at 0.5 and lets the reader move it, because every threshold in the chart scales by the same factor when it changes. Moving phit shifts where the boundary sits; it does not reorder the systems or change the distance between them.
The chart below places every system from the previous section on a single dollar axis. Each marker sits at the minimum defended-asset value that justifies firing that system once, which is its cost-per-kill divided by phit. The white droplines mark real assets at their sourced values. An asset line falling to the right of a system means that system pays for itself defending it, and the slider moves every marker together, so the ordering holds whatever value of phit a reader prefers.
Read from the top and the first thing that stands out is how much of the chart is settled. A destroyer, a hangar of fighters, and a fuel depot all sit far enough to the right that every system in the dataset clears them, which is why they are not drawn here. Nothing in counter-drone procurement is contested at that end. The argument lives in the band between a Humvee and an Abrams, roughly two and a half decades of asset value, and that is where the ladder separates.
At the Abrams line, eleven of thirteen systems pay for themselves. PAC-3 clears it with room, needing an asset worth about $9.1 million against a tank worth $12.5 million. SM-6 and THAAD do not, and the gap is not marginal: THAAD needs something worth $28.5 million before the shot breaks even, more than twice the value of the thing it is protecting. Firing a THAAD interceptor at a drone threatening a tank loses money even after the tank is fully counted. Step down to a Bradley at $4.35 million and PAC-3 drops out too, leaving every kinetic missile interceptor on the wrong side of the line. Step down again to a Humvee and Skyranger joins them, which puts the cheapest gun-based system in the dataset outside the envelope for the vehicle a drone is most likely to find.
What survives the whole descent is the bottom of the ladder. Leonidas justifies itself against anything worth more than about $2,200, Sky Sentinel against anything worth more than $370, and the Bullfrog estimate against $556. Those figures are so far below every asset on the chart that the second threshold stops constraining them at all. For the cheap end of the dataset the question of whether a shot is worth taking does not arise, and what remains is whether the system works and whether enough of them exist.
| Defended Asset | Value | Min. Justified Cost-Per-Kill | Source |
|---|
Put the two thresholds together and the systems sort into three groups rather than a ranking. Automated weapon stations and non-kinetic systems clear the second threshold so far below every asset on the chart that it stops constraining them: Sky Sentinel justifies itself against anything worth more than $370, and Leonidas against anything worth more than about $2,200. Interceptor drones sit just above them, with Sting at roughly $5,900 and the DroneHunter near $20,000, still under every defended asset here. For all three categories the economic question is settled before it is asked, and what remains is engineering and production.
Missile interceptors are the group where the second threshold does real work. On a unit-for-unit basis they remain defensible against high-value targets, which is the case the destroyer makes and the case Rumbaugh was right to insist on. PAC-3 pays for itself defending a tank. SM-6 pays for itself many times over defending the ship it is carried on. The failure is not the interceptor and not the price of the missile. It is the pairing, and the pairing goes wrong in a specific and increasingly common way: an expensive interceptor matched against a cheap and numerous threat aimed at something ordinary.
That is also where this model stops being the right tool. Everything above holds shot by shot, and the case for firing a missile at a drone threatening a destroyer survives every value of phit a reader might choose. Against a swarm the calculation changes, and it changes for a reason this model sets aside. A magazine that runs out cannot be argued back into existence by the value of what it was defending, and an attacker who can strip one for four million dollars will keep doing so whether or not each individual intercept was justified. The second threshold tells a commander whether to take a shot rather than telling them whether they will have shots left.
Cost-per-kill alone condemns almost every defensive system in this dataset, and that verdict is too broad to act on. Adding the value of what is being defended narrows it to something a procurement officer can use. Cheap reusable systems are justified against effectively anything, so the decision about them is a question of whether they work and whether enough exist. Missile interceptors are justified against high-value assets and only those, so the decision about them is a question of what they are guarding and what else could guard it. The cost-asymmetry problem is not that interceptors are expensive. It is that the threat has become cheap enough, and numerous enough, that the assets worth an expensive interceptor are a shrinking share of the ones under attack. That share may shrink from the other direction too. The systems at the bottom of this ladder came disproportionately from firms that built them as products and sold them at a price rather than from programs funded through years of development against a government requirement, and if that shift holds, less of the fielded inventory will sit above the threshold at all.
The question this report set out to answer was whether the cost-asymmetry that defines drone warfare is closing as counter-drone technology matures. The answer is that it is closing, however almost none of the closing came from the systems the question was originally asked about. Across the five tracked series, only one system got meaningfully cheaper per kill between 2022 and 2026: PAC-3, which fell about thirteen percent in real terms. Systems such as THAAD and SM-6 rose, while Sting and Leonidas, the two cheap systems with enough data to trace, ended the period costing more per kill than they started it. What moved wasn’t the system costs themselves, but the barrier to entry. Sky Sentinel at $185, the Bullfrog estimate at $278, and LOCUST at about $1,000 are all recent arrivals, and they sit two to four orders of magnitude below anything that was fielded when the war began. The asymmetry is narrowing by entry rather than by learning, and that distinction decides what a procurement office can actually do about it.
The first consequence is that buying more of the same interceptor does not improve the exchange. Cost-per-kill for a missile is unit cost divided by probability of kill, and neither term moves much with volume once a program is in serial production. The five tracked series are the primary evidence: PAC-3 is the only one that fell, and it fell thirteen percent over four years while the threat it was built against stayed at roughly $35,000 a unit. A procurement office that responds to cost-asymmetry by ordering more PAC-3 is buying capacity, which is a real thing to want, but it is not buying a better exchange rate. Nothing in this dataset suggests that any incumbent interceptor is on a path to becoming cheap.
The second consequence is that the two thresholds ask different questions of different systems, and a buying decision is easier once it is clear which question applies. For the cheap end of the spectrum the economics are already clear. Sky Sentinel, Leonidas, LOCUST and the Bullfrog estimate clear the asset threshold against anything worth defending, so the open questions about them are whether they work under contested conditions, whether they survive electronic warfare, and whether enough of them can be built. Those are engineering and industrial questions rather than economic ones, and they are the questions worth spending review time on. For missile interceptors the question is narrower and more uncomfortable: what specifically is this system guarding, and is there anything cheaper that could guard it instead that maintains the same level of defense. While a PAC-3 being used to defend a tank is economically coherent, the same PAC-3 being used to defend a substation or a JLTV loses money at every reasonable value of phit.
The argument that cheap mass beats exquisite systems has become the standard pitch from the newer defense firms, and this dataset supports it at the floor without reservation. The three cheapest entries here were all built by companies that funded development themselves and sold a finished product, and all three sit far enough below every defended asset that no reasonable choice of phit puts them on the wrong side of the line. Where the pitch is strongest is exactly where it is least interesting to argue about, as nobody actually needs a model to conclude that a $185 intercept is worth taking.
The middle of the spread is where the claim gets tested, and there it is more complicated than the pitch usually allows. Roadrunner-M is built on the product model by the company most associated with it, and it still needs an asset worth roughly $1.2 million before a shot pays for itself, which puts a Humvee and a JLTV outside its cost-effective range. Skyranger 30 is a modern gun system from a traditional manufacturer and lands at $36,988, close enough to the Shahed line that the data cannot say which side of parity it falls on. Being new does not automatically make a system cheap, and being a legacy prime does not automatically make one expensive. What separates the floor from the middle is not the business model on its own but whether the consumable is cheap: a rifle cartridge, a burst of .50 caliber, or a pulse of electricity. Systems whose consumable is another airframe or another guided round stay expensive no matter who builds them or how they are sold.
For force design the useful version of the claim is much narrower than the standard marketing version. Cheap systems are not a replacement for interceptors, because the assets that justify an interceptor still exist and still need defending. They are a way to stop spending interceptors on everything else. A force that fields enough cheap reusable systems to cover its ordinary assets preserves its expensive magazines for the ships, aircraft, and command nodes that actually warrant them. That is a layering argument rather than a replacement argument, and it is the one the numbers in this report support.
Unit price is one factor, albeit a very important one but not the full scope of what makes an effective manufacturer. Where the thing is built and how fast it can be replaced is another important piece and it acts on the constraint this report set aside rather than the one it measured. TYTAN Technologies builds the METIS interceptor with a 3D-printed airframe, holds a Bundeswehr development contract awarded in October 2025, delivered more than a thousand units to Ukraine’s National Guard under German funding in April 2026, and is standing up a Bavarian plant aimed at 3,000 interceptors a month. The production rate stands out as a true differentiator to exquisite systems with more complex and expensive parts. A magazine that refills at three thousand units a month is a different object than one that refills at the pace of a missile line.
Firestorm Labs pushes the same idea further by moving the factory instead of the inventory. Its xCell is a pair of twenty-foot containers holding industrial polymer printers and assembly stations, able to turn out a complete airframe in nine to twenty-four hours and roughly fifty Group 2 airframes a month from a single unit. In August 2026 an xCell produced more than a thousand parts and twelve assembled airframes aboard the USS Essex while the ship was under way to RIMPAC. Firestorm’s own aircraft are strike and reconnaissance platforms rather than interceptors, so they do not belong in the cost-per-kill tables, but the manufacturing case is indifferent to what gets printed. The relevant claim is that a defended position can hold a design file instead of a stockpile.
Neither company is evidence that additive manufacturing is cheaper per unit, and this report does not have the data to make that claim. What both suggest is that the two hardest constraints on cheap defense may be loosening at once. Cost-per-kill falls when the consumable is cheap, and magazine depth stops being fixed when the consumable can be produced near where it is spent. An attacker who strips a Patriot magazine for four million dollars in drones has destroyed something worth close to four hundred million, and the replacement crosses an ocean. The same attacker stripping a printed interceptor magazine has destroyed something far cheaper, and the replacement is a design file and a container on the same airfield.
The finding is more robust to bad numbers than it looks. The weakest figures in this report are the expected-kill counts behind the cheap reusable systems, and Leonidas carries an explicit order-of-magnitude caveat. If every one of them is twenty-fold optimistic, Leonidas moves from about $1,088 to roughly $21,800, LOCUST from $1,005 to about $20,100, and Sky Sentinel from $185 to $3,700. All three are still under the Shahed line, and all three still clear every defended asset in the previous section. A reader who distrusts the vendor claims entirely can discount them by an order of magnitude and reach the same conclusion, since the economic argument does not depend on the optimistic figures being right.
What would change the conclusion is not the price of these systems but whether they would get to fire in the first place. There are three things that would have a meaningful impact on the conclusion. The first is effectiveness being mismeasured in kind rather than in degree. Epirus reports a single pulse affecting dozens of drones, and that figure is the difference between $1,088 per kill and something far higher, but it presumes drones that respond to the pulse. Hardened airframes, shielded electronics, or a frequency the emitter does not cover would not make Leonidas twenty percent worse, they would make it a different system. The same applies to any rounds-per-kill figure drawn from engagements against one drone type.
The second is survivability, which cost-per-kill does not capture at all. The metric counts what a system kills and ignores what kills the system. A hundred-thousand dollar turret is a soft target, and if a $35,000 drone can destroy one, the exchange inverts: the defensive system becomes the cheaper thing to trade away. Nothing in this report measures how long a Sky Sentinel or a Bullfrog survives once located, and an attacker with any reconnaissance capability will locate them. The third is electronic warfare. Sting is itself a drone and needs a link, and the automated stations need optics and fire control. Most of the effectiveness data here comes out of Ukraine, an environment with a specific and evolving jamming picture, and a system that works there is not guaranteed to work against an opponent who has studied it. Cheap systems are cheap in part because they are simple, and simple systems tend to have fewer ways to keep working when someone attacks the thing they depend on.
The cost-asymmetry problem is not simply that interceptors are too expensive. It’s that these exquisite systems were built to defend against targets worth a missile, and are instead being heavily used against systems that are not. Every system in this report that closes the gap does so by making the consumable cheap rather than by making the platform cheap, and none of the systems built the older way got meaningfully cheaper in four years of the most intense drone conflict on record. The cost-asymmetry correction is coming from platforms outside of the traditional field, one entrant at a time, and the open question is no longer whether cheap defense is economically sound but whether it can be made to survive contact.
Three limitations sit beyond the ones already stated in the methodology. The first is provenance. A meaningful share of the effectiveness data here originates with governments and manufacturers describing their own performance, and figures coming out of Ukraine and Russia in particular serve a purpose beyond informing analysts. Kill counts are claimed more often than they are audited. Nothing in this report treats a self-reported figure as verified, and the tier ratings in the tables are the only honest guide to which numbers rest on a budget document and which rest on somebody’s press release.
The second is time resolution. Only five of the twelve tracked systems carry a multi-year series, and seven are single points anchored to one sourced year. The finding in the previous section that no incumbent got meaningfully cheaper rests on those five, which are also the five oldest programs in the dataset. Newer systems have not existed long enough to show a cost trajectory at all, so this report can say what the floor is today but cannot say whether the floor is still falling. The third is the breakeven model itself, which remains deterministic, ignores saturation, and turns on a value for phit that no source in this report supports.
Two extensions would do the most to close these gaps. The first is a saturation model built on magazine depth rather than unit economics, which is now closer to possible than it was: ready-round counts are published or derivable for most of these systems, and nightly raid sizes are reported out of Ukraine. What that model needs and this one lacks is force allocation, how many fire units cover a given area, and that remains the missing term rather than the missing math. The second is a single-program cost trajectory traced in detail, following one interceptor from first procurement to current unit price, to establish whether the flat cost curves found here are a property of these four years or a property of how these systems are bought.