The honest answer: you can't buy one

If a commercial operator walks up to the sidewalk delivery robot industry and asks "what's the cheapest one I can put on the street tomorrow," the answer is not a price tag. It is a category rejection. Every credible sidewalk delivery robot maker in 2026 (Starship, Serve, Coco, Kiwibot, Avride) operates a captive fleet under a Robots-as-a-Service model, not a hardware sale. The robot is not the product; the completed delivery is. That reframes the question from "cheapest to buy" into "cheapest to buy from," and the answer there is Starship Technologies, on the strength of the only third-party build-cost estimate that survives sourcing: $3,000 to $7,000 per unit, per an October 2025 Guardian profile of co-founder Ahti Heinla.

That is a build cost, not a purchase price, and it matters only because Starship spreads it across the most deliveries: 10 million cumulative, 3,000+ robots in service, ~125,000 road crossings per day as of April 2026, across 8 countries and 65+ US university campuses. Volume is the whole game in this segment, and Starship has more of it than anyone else in the West.

Runners-up, ranked by what you can actually verify

Because no manufacturer publishes a per-bot or per-delivery rate card, "cheapest" has to be inferred from the unit economics of the operators themselves. Only one of them files with the SEC, which puts the ranking on unusually solid ground for a private-heavy segment.

1. Starship Bot: cheapest by unit build cost and by scale

Starship's $3,000 to $7,000 build cost is an analyst estimate (labeled as such on the model page) but it is the only third-party number in the segment that has been publicly attached to a specific robot. The Guardian's October 2025 piece framed it as the per-unit cost that made the $50M Series C (led by Plural, closed October 15 2025) look reasonable against a claimed path to "12,000+ robots by 2027." Two features earn Starship the top slot for a would-be operator:

• Captive service model. Starship does not sell robots to third parties. It operates them. A university, grocer, or delivery app buys completed deliveries on Starship's fleet, integrated with Uber Eats, Just Eat, Bolt, Foodora, Grubhub, and Wolt. There is no capex line for the operator.
• The lowest published build cost pulls the RaaS floor down. If a Starship bot costs $3K to $7K to build and lasts multiple years in field, the amortization per delivery on a 22-million-autonomous-kilometer fleet is structurally lower than any competitor without that scale. Starship's own count of 3,000+ robots is more than 1.5x Serve's built capacity and roughly an order of magnitude past Coco's disclosed footprint.

Two caveats to keep this from becoming a hype line. First, Starship markets "Level 4 autonomy without active human supervision," but the honest label is operator-supervised L4: the fleet runs autonomously with remote human oversight and takeover, historically around 99 percent autonomous. Second, the "12,000 by 2027" figure is a company projection, not a run-rate, and should not enter any cost model.

2. Serve Gen 3: the only publicly disclosed unit economics in the segment

Serve Robotics does not publish a purchase price either (it is not for sale), but as NASDAQ: SERV it publishes quarterly financials that are cleaner than any competitor's press-release math. From Serve's Q1 2026 disclosures: revenue $3.0M (+578% year-over-year), split roughly $1.958M fleet services + $1.026M software, against 812 daily-active robots as of the end of the quarter. Divide the fleet-services line by daily-active robots and quarterly-annualize and you land in the low-thousands per robot per month in gross fleet-services revenue, before Serve's own opex. That is a floor, not a ceiling: it is what Uber and other platforms currently pay Serve, not what Serve pays to produce.

Two features make Serve the strongest runner-up:

• A public financial trail. Anyone building an operator model can pull the 10-Q, not a marketing deck. The Serve Robotics x Uber Eats deployment partnership is the reference customer contract in the segment and has anchored DoorDash (October 2025) and Shake Shack integrations on top.
• Gen 3 halved unit cost. Serve's own October 16 2024 disclosure at the Gen 3 rollout put the manufacturing cost at roughly half the Gen 2 figure, with 2x speed, 2x range, +6 field-hours per day, and 5x onboard compute (NVIDIA Jetson Orin). That is the direction the whole segment is moving, and Serve is the only one showing the math.

The cap-flag: Serve markets its 2,000-robot number as fleet size; Q1 2026 daily-active is 812. Built-vs-active gets conflated constantly. Any per-robot cost figure should be computed against daily-active, not built.

3. Coco Delivery Robot / Coco Bot: the merchant-anchored option

Coco Robotics raised an $80M strategic round in June 2025 and reports 500,000+ deliveries, 3,000+ restaurant and merchant partners, and active operations across Los Angeles, Chicago, Miami, and Helsinki. Its Coco 2 hardware (February 2026) upgraded to NVIDIA Jetson Orin NX and front-facing lidar. Coco has not published a per-delivery rate or a per-robot cost anywhere primary-sourced. The February 2026 refresh confirmed a 3-mile delivery radius and a 4-grocery-bag capacity, and Coco integrates with Uber Eats, DoorDash, and Wolt.

Coco earns its spot in the ranking not on price (nothing to cite) but on merchant density: a restaurant onboarding to a sidewalk-delivery motion in LA or Miami is more likely to encounter Coco integration than Serve, because Coco built out on the merchant side while Serve built out on the delivery-app side.

4. Kiwibot 4.0: the semi-autonomous teleop option

Kiwibot is the veteran of the low-cost end of the segment; the drive mode is disclosed as "semi-autonomous with teleoperation," and the Kiwibot x Grubhub partnership put the platform on campuses at a price point Kiwibot has repeatedly claimed is cheaper than any human courier. There is no primary-sourced dollar figure to cite here, and Kiwibot's continued reliance on teleop pilots (rather than progressive autonomy claims) is the reason: the labor line moves the unit cost, and Kiwibot has not published either.

5. Avride delivery robot: the platform-attached option

Avride (spun out of Yandex's self-driving group) has an active Uber Eats deployment on the Ohio State campus and a Grubhub campus footprint. No published unit price. It sits in the ranking because it is the only sidewalk robot outside Serve/Starship/Coco/Kiwibot with a plausible route to commercial scale, riding on the same delivery-app integrations that make the rest of the segment work.

---

What appears to beat the answer but doesn't count

Two shapes keep coming up when this question gets asked and both deserve to be dismantled.

"Amazon Scout is the cheapest, it's Amazon." Amazon Scout is discontinued. Amazon paused field testing in October 2022 and the program has not been reconstituted for external operators. A discontinued platform is not an answer to a "what should I deploy" question at any price.

"Nuro is a delivery robot and it's cheaper per delivery than any sidewalk robot." Nuro does deliveries and its per-delivery cost claims (widely reported) are indeed lower than sidewalk economics on paper. But the Nuro R2 and R3 are autonomous vehicles in the road-legal, DMV-permitted sense: they are AVs deployed on public roads under state autonomous-vehicle regulations, not sidewalk delivery robots operating under state-by-state Personal Delivery Device statutes. Different regulatory regime, different capex class (Nuro's per-unit hardware cost is a multiple of Starship's), different failure modes. A commercial operator asking about sidewalk delivery is asking about the PDD regime, not the AV regime. Do not conflate them.

"Also Autonomous Delivery Vehicle or Delivers AI YEBO is cheaper, they're newer." Both are still in pilot stage per the sidewalk-cohort review. "Newer" is not "cheaper" in a segment where scale is what makes the amortization work. A pilot-stage platform without disclosed unit economics is not the cheapest anything; it is unpriced.

"A restaurant robot or busing robot is cheaper." Bear Robotics Servi Plus, Richtech Matradee, and Pudu's line are cheaper per unit than any sidewalk robot, and they run under RaaS, and they are for sale. They are not sidewalk delivery robots. They are indoor-only service robots operating on carpet. Different question.

---

Gotchas for anyone building an operator model

A few standing traps in this segment that trip up cost comparisons.

"Purchase price" and "operator price" are not the same thing. None of the top four sidewalk platforms sell hardware to commercial operators. The Starship $3K-$7K figure is a build cost inferred from a founder interview, not a sticker price. The Serve fleet-services revenue is what Uber pays Serve per active bot per quarter, not what a restaurant chain would pay to lease one. Anyone building a per-delivery cost model has to work from platform-level disclosures (Serve's 10-Q is the only one available) and pilot-level press releases, not vendor spec sheets.

Built capacity vs daily-active is where the padding lives. Starship's 3,000+ figure and Serve's 2,000 figure are cumulative built or in-service capacity. Serve's Q1 2026 filing puts daily-active at 812. Coco's 500,000 deliveries is a cumulative number back to launch, not a run-rate. The industry has learned to publish the flattering ratio; a commercial operator has to work with the tighter one.

Regulatory carve-outs move the cost. Sidewalk delivery robots operate under state PDD statutes (Virginia's 2017 law was the model; 20-plus states have variants). Some cities ban them outright (San Francisco), some cap fleet size, some require permits per corridor. The "cheapest" robot in a state where you can't operate it is not cheapest. Check the Personal Delivery Device statute for your target city before comparing per-delivery quotes.

"Level 4 autonomy" claims are mostly operator-supervised. Every disclosure in this segment marketed as L4 is actually a bot running autonomously with remote human oversight and takeover available. Starship acknowledges around 99 percent autonomous historically; Serve markets 99.8 percent completion but is explicitly remotely monitored; Kiwibot is semi-autonomous with teleop. The teleop headcount is a real cost line and it is not zero. A per-delivery cost model that assumes zero human labor is wrong for every platform in this ranking.

Aggregator valuations are not disclosures. Starship's $50M Series C did not carry a company-confirmed valuation; any figure attached to it is an aggregator inference. The same applies to Coco's $80M round. Neither should enter an operator's cost model.

---

The bottom line

There is no published sticker price for a sidewalk delivery robot a commercial operator can buy in 2026, because none of the credible platforms sell hardware to third parties. The closest thing to a "cheapest" answer is Starship, on the strength of the only third-party build-cost estimate ($3,000 to $7,000 per unit, Guardian, October 2025), the largest deployed fleet, and a captive service model that lets operators buy completed deliveries instead of robots.

Serve is the runner-up on the strength of being the only platform with public financials that let an operator compute unit economics from the 10-Q. Coco and Kiwibot round out the credible set. Everything cheaper is either discontinued, a different regulatory regime, or an indoor robot that does not belong in the comparison.

Sources: Serve Gen 3 registry record; Starship Bot registry record; Coco Delivery Robot registry record; Kiwibot 4.0 registry record; Serve Robotics x Uber Eats partnership; Guardian profile of Ahti Heinla, 2025-10-04; Serve Robotics Q1 2026 SEC filing (NASDAQ: SERV).