BMW’s humanoid pilot, the most documented one on earth, delivered about 1,250 operating hours in ten months, which is 0.6 of one worker’s year. That is the number to hold onto when a vendor arrives with a video. Figure 02 moved more than 90,000 sheet-metal parts at Spartanburg on 10-hour shifts, Monday to Friday, while the plant built 30,000 X3s, and both BMW and Figure published the hours. Nobody else has. This piece reads every announced humanoid pilot the way a plant manager reads a bid: robot-hours delivered, task, dates, price, and what was disclosed against what was claimed; then prices a robot-hour against a fully loaded worker, scores the vendors on the five OEM Disclosure Scorecard criteria, and behind the wall lays out the pilot terms to sign, the terms to refuse, and the two use cases where the math already works.
Source: Open Factory OEM Disclosure Scorecard, humanoid supplement (compiled from BMW Group, Figure, GXO, Agility Robotics S-4, Apptronik, Hyundai Motor Group, Boston Dynamics, Tesla 10-K, Electrek; as of Sep 2026). CSV: data/pilot-table.csv
Nine pilots, one runtime figure (1,250 hours), one fleet-hours figure (Agility’s 65,000 across nine committed facilities), one price ($8,500 a month, and only because Agility filed to go public), and six rows with “None” in both the hours and the price column.
What BMW Actually Bought
The Spartanburg deployment is the reference case because two parties published numbers that agree. BMW’s 27 February 2026 release says Figure 02 “moved more than 90,000 components and covered approximately 1.2 million steps in around 1,250 operating hours” over ten months of ten-hour shifts, five days a week. Figure’s own 19 November 2025 write-up gives the task spec: a station requirement of 84 seconds total cycle and 37 seconds of load time, parts placed within a 5-millimetre tolerance in 2 seconds, a target of over 99% placement success per shift and a goal of zero interventions per shift. It also says the forearm was “our top hardware failure point” and that Figure 02 is being retired now that Figure 03 exists.
Ten months of ten-hour shifts, five days a week, is roughly 2,100 to 2,200 scheduled shift-hours; 1,250 hours of runtime is 57 to 60% of that, and at that pace a robot delivers about 1,500 productive hours a year against 2,000 for a one-shift worker. That arithmetic is ours, in data/assumptions.csv, and it is generous to the robot because BMW’s “operating hours” may include time the line was down. What neither party published is the achieved success rate against the 99% target, the interventions per shift against the zero goal, or what BMW paid, if anything. The pilot was a vendor demonstration on a live line, and it was a good one; it was not a purchase.
The pace matters more than the parts count. 90,000 parts in 1,250 hours is 72 parts an hour, one every 50 seconds, on a station specified at 84 seconds; the robot kept up with the line when it ran. The question a buyer asks is how often it ran, and 57 to 60% of scheduled shift time is the only answer in the public record. A Fanuc arm in a welding cell is expected to run above 95%, and the Bid-Tab Price Book prices a conventional cell at $100,000 to $250,000 installed for a task like this one, with a 15-year life.
Agility Filed the Only Price Sheet
Agility Robotics is the one humanoid vendor that has been forced to write numbers down, because it is merging into Churchill Capital Corp XI and filed an S-4 on 4 September 2026. The illustrative unit economics for Digit v5 are: a RaaS fee of $8,500 a month per robot including software and maintenance, a one-time deployment fee of about $25,000, a five-year useful life; or an ownership model at about $200,000 per robot, $20,000 deployment, and $36,000 a year for Arc software and maintenance. Agility’s cost to build is $150,000 at launch, targeted at $75,000 at 1,000 units a year and $30,000 at 10,000. The filing states plainly that these are “illustrative” and that actual contract terms “are subject to negotiation.”
The June 2026 investor deck supplies the comparison the sales team will use. It sets a fully burdened human rate of $30.50 an hour, built from “$21.49 published hourly rate for material movers in 2025 per BLS data” with 5% wage inflation and a 35% benefits load, then runs the robot 120 hours a week, two ten-hour shifts, 52 weeks a year, for about 31,000 service-life hours over five years. On those assumptions one Digit replaces about $200,000 a year of labor for about $100,000 a year of RaaS. Every one of those assumptions is a negotiable input, and the deck says so in its footnotes.
The number the deck does not print is the cost per productive hour, and at BMW’s disclosed pace it is $71. RaaS at $8,500 a month plus the deployment fee spread over five years is $107,000 a year. Divide by Agility’s 6,240 hours and a robot-hour costs $17.15. Divide by 4,000 hours, a real two-shift year, and it is $26.75. Divide by 1,500 hours, the only pace any humanoid pilot has disclosed, and it is $71.33, more than double Agility’s own worker rate.
Source: Open Factory Bid-Tab Price Book, automation worked example (compiled from Agility Robotics S-4 Sep 2026, BLS ECEC Jun 2026, BLS OES 2025 via O\*NET; as of Sep 2026). CSV: data/cost-per-productive-hour.csv
The RaaS curve crosses Agility’s $30.50 at 3,508 productive hours a year, crosses a fully loaded material mover at $27.65 at 3,870 hours, and crosses the loaded average manufacturing production wage of $43.39 at 2,466 hours; the ownership model at $80,000 a year crosses $30.50 at 2,623 hours. At 1,500 hours the RaaS robot costs $71.33 and the owned robot $53.33 an hour.
Our worker rates come from public series, not the deck. BLS puts the median material mover at $19.35 an hour and the median team assembler at $21.47 in its 2025 wage data, and the June 2026 Employer Costs for Employee Compensation release puts private-industry compensation at $46.89 an hour against $32.82 of wages, a 1.43 multiplier. Loaded, the material mover is $27.65 and the assembler $30.67, so Agility’s $30.50 is fair for a warehouse and low for a body shop, where the manufacturing production average of $30.37 in August 2026 loads to $43.39. A Spartanburg or Savannah plant competing with datacenter construction for labor pays more than the national median, as we showed in electricians’ wages by metro.
Announced Capacity vs Delivered Units
The gap between what humanoid vendors have announced and what they have delivered is the largest in any equipment class we cover, and it is the reason a plant manager should treat every capacity claim as marketing until it appears in a filing.
Source: Open Factory OEM Disclosure Scorecard, humanoid supplement. CSV: data/announced-vs-disclosed.csv
Figure announced 12,000 humanoids a year at BotQ in March 2025 and reported more than 350 Figure 03 units delivered by 29 April 2026, allocated to “internal research and development groups, data collection, housework efforts, and commercial use-case development.” Agility’s RoboFab is “designed to produce up to 10,000 Digits annually”; the S-4 does not count deployed robots, but it reports 2025 robot sales revenue of $1,550,360, which at the company’s own illustrative $200,000 purchase price is fewer than eight robots, and total 2025 revenue of $1,781,967 against a net loss of $138.1 million. Tesla told investors in January 2025 that “several thousand Optimus robots will be doing useful things by the end of the year” and in January 2026 that Optimus is “not in usage in our factories in a material way,” with start of production “before the end of 2026” and eventual capacity of one million a year, as reported by Electrek from the Q4 2025 call; Tesla’s 10-K filed 29 January 2026 states “We have yet to commercialize Bots.” Hyundai’s CES 2026 release plans Atlas on parts sequencing at Metaplant America “beginning in 2028” and a 30,000-unit-a-year robot facility; Boston Dynamics’ own account of the fall 2025 Metaplant test gives no hours or units. Apptronik has raised $935 million in Series A money, named Mercedes-Benz, GXO and Jabil, and has published no unit, hour or price figure since its March 2024 Mercedes agreement.
The units that exist are Chinese: Interact Analysis counts humanoid production rising from under 2,000 units in 2024 to more than 20,000 in 2025, over 90% of it in China, with about 10% of 2025 output in real-world applications. That is Interact Analysis’s own June 2026 note, which puts Unitree and AgiBot above 5,000 units each and the commercial inflection “post-2032.” The same firm’s forecast of fewer than 100,000 humanoids globally by 2030, as reported by Manufacturing Dive, sits against 621,000 conventional industrial robots installed in 2025 alone, 38,000 of them in the US. Against that base, North American buyers ordered 36,766 arms for $2.25 billion in 2025, about $61,000 each; the arm is the cheap part, and the cell around it is what costs $150,000.
Scoring the Vendors on Disclosure
The OEM Disclosure Scorecard asks five observable questions of every equipment maker we cover: does the company publish lead times, list prices, backlog in dollars, capacity in units, and delivery performance. We applied the same five to six humanoid vendors, source beside every cell, with the same right of reply we gave the sixteen electrical, power and cooling OEMs in December.
Source: Open Factory OEM Disclosure Scorecard, humanoid supplement. CSV: data/disclosure-scorecard.csv
Agility scores 3.0: a dollar backlog ($300 million of Digit v5 orders for 1,000 robots on a three-year RaaS term, from one customer, with warrants issued to that purchaser), a unit capacity (10,000 a year), a partial price (illustrative, “subject to negotiation”) and partial delivery performance (65,000 fleet hours and 98% tote accuracy, no uptime and no robot count). Figure scores 1.5 on capacity and the BMW hours. Unitree scores 1.0 because its web store lists the G1 at $13,500 and the H2 at $29,900 while the H1 line says “Contact us for the real price.” Tesla scores 0.5 for a planned capacity and nothing else; Boston Dynamics and Hyundai 0.5 for a planned 2028 facility; Apptronik 0.0.
No humanoid vendor publishes a lead time, and only one publishes a price a buyer could put in a purchase order. For comparison, Schneider Electric publishes a monthly lead-time guide that feeds our Lead-Time Monitor, Haas prints a list price on its website, and even the $2,850 market reports we criticised in The $2,850 Report That Says Nothing carry a unit count. The humanoid sector in 2026 discloses less than the machine-tool trade did in 1990, which is a reason to write the disclosure into the pilot contract, because the vendor will not volunteer it.
What to Do Monday
Price the robot-hour, not the robot. Take data/cost-per-productive-hour.csv, replace our loaded wage with your own plant’s for the specific position, and read off the productive hours a year the robot must deliver before it beats the person. For a two-shift 3PL floor at $27.65 that is 3,870 hours on Agility’s published RaaS terms, 62% of the 6,240 hours Agility assumes and 2.6 times the pace BMW disclosed. A vendor whose proposal does not commit to logging productive hours has not agreed to be measured.
Ask for the four numbers BMW and Figure published and the two they did not: runtime hours, parts moved, cycle time achieved, scheduled shift-hours, plus interventions per shift and first-pass success rate. Put them in the RFQ as reported metrics with a weekly cadence, the way our capital-equipment RFQ template treats FAT and SAT, and run the price through Quote Check with the RaaS fee, deployment fee and software line separated. If you have signed a humanoid pilot, tell the Equipment Price Benchmark the monthly rate; nobody outside Agility’s S-4 has published one, and the cell opens at five contributors.
Behind the paywall: the six clauses to sign and the five to refuse, with model language, including the 30-day exit that turns a pilot back into a pilot, and the two use cases where a Digit at $8,500 a month already beats the loaded wage on the Bid-Tab Price Book numbers.