Spec Drift: Which Humanoid Manufacturers Quietly Downgrade Their Robots Between Reveal and Ship

Figure 03 shipped in October 2025 with a rated payload of 20 kg. Figure 02, the platform it succeeded, was rated 25 kg. The launch deck spent time on the new palm cameras, the wireless foot charging, and the drop from 70 to 60 kg of body mass. It did not mention that payload had dropped by a fifth.

That is spec drift. A published number changes between the announcement and the shipping unit — sometimes with a footnote, more often without one. Every humanoid manufacturer we track has done it at least once since 2023. Some of the changes are engineering trade-offs that make sense on paper. Some of them are marketing rewrites of the same hardware. All of them matter to a procurement team trying to plan a deployment against a reference sheet.

Below is what the last three years of spec drift look like across six manufacturers, and how to check the number you are about to sign a contract against.

Payload is the number most often revised downward

Payload is the single most quoted humanoid spec because it maps directly to warehouse work. A robot rated for 25 kg lifts a full tote. A robot rated for 15 kg lifts most totes but not all of them. Buyers read this number first. Which is why it moves.

Figure 02 launched in August 2024 with a payload rating of 25 kg. Figure 03, announced October 2025, dropped that number to 20 kg. Figure has framed the change as a home-vs-warehouse repositioning, and the 60 kg body mass supports that framing. The 20 kg figure never appeared in the year of Figure 03 pre-announcement materials that positioned it as an "everything upgrade" from Figure 02.

Apptronik Apollo's numbers have been more stable — the 25 kg published rating held between the December 2023 debut and the shipping Mercedes-Benz and GXO deployments. But early materials from 2022 quoted a 25-kg-per-arm rating that Apptronik clarified in later documentation to be total system payload, not per-arm. That is not a hardware change; it is a specification cleanup, and the difference for a purchasing team is not trivial.

Unitree G1 was announced in May 2024 with a payload rated at 3 kg. Field measurements from research labs published in 2025 confirmed the number. The H1-2, released in October 2024, showed a payload figure that moved between preview materials and the shipping product page over the course of a month. Neither release note nor changelog marked the revision.

Runtime numbers are harder to verify and drift more often

Runtime is the second most-quoted spec. It is also the hardest to verify from outside the manufacturer, because "runtime" means different things under different load profiles.

Tesla Optimus Gen 2, unveiled December 2023, quoted no runtime figure. Optimus Gen 3, revealed October 2025 at the We, Robot event, quoted 540 minutes. Tesla stated the number on stage without a load profile. Follow-up material described the 540 minutes as "estimated under typical work load" without defining typical. The runtime has not been independently verified.

Figure 02 quoted a 300-minute runtime at launch. Figure 03, with a 2.3 kWh battery against Figure 02's 2.25 kWh, quotes 300 minutes as well. That is a runtime tie between two robots that differ by 10 kg of body mass and one full set of camera and tactile hardware. Either Figure 03 is materially more efficient than Figure 02, or one of the two numbers is not being measured the same way. Figure has not published the workload used to derive either figure.

Agility Digit shipped 2019-generation units at roughly one hour of continuous work. Current Digit shipping units publish a base runtime near 240 minutes with hot-swap batteries extending shift time. That is a genuine hardware improvement across generations, and Agility documented it in release notes. It is one of the cleaner examples of runtime progress in the market.

Unitree H1 quotes 105 minutes on its 864 Wh battery. Independent testing by robotics groups published in 2025 measured continuous walking runtime near 90 to 95 minutes, closer to but not matching the manufacturer number.

Silent feature removals happen at the software layer

Hardware specs move slowly and leave paper trails. Software specs move fast and often silently.

The most common silent removal we track is ROS support. Multiple manufacturers advertised "ROS-compatible" or "ROS 2 developer preview" at announcement, then quietly removed the language from product pages once shipping began. Unitree G1 launched with strong ROS 2 messaging; a 2025 firmware update changed the surface of what could be accessed through public nodes without a matching public changelog. Community developers on the Unitree forums documented the delta.

Figure launched with public messaging about openness to research partnerships. By early 2025 that language was gone from the site and Figure had shifted to a fully closed Helix VLA stack. No formal deprecation announcement was made.

Apptronik Apollo initially quoted a modular software architecture with third-party integration paths. Current documentation does not describe any published SDK or public developer program. The website language changed between mid-2024 and early 2025 without a release note.

Silent feature removals matter more than payload drift for research groups and system integrators. A payload change forces a spreadsheet update. A software support change forces a stack rewrite.

Price drift is the most public and least examined

Price is quoted more in podcasts and press interviews than in shipping documentation, and it drifts accordingly.

Tesla Optimus was described publicly as targeting under twenty thousand dollars at multiple points in 2022 and 2023. No production Optimus has yet been sold. The 2025 Optimus Gen 3 reveal restated a $20,000 to $30,000 band as a future target with no order book open. The gap between a stated price target and a shipped unit price is not spec drift in the strict sense — but it is the same information asymmetry.

1X NEO opened preorders at $20,000 with a $200 deposit. Subsequent messaging added a $499 per month subscription option that had not been described in the reveal materials. The all-in cost over three years lands closer to $40,000 than to $20,000. Both figures are true. The reveal figure was the one that produced the headlines.

Unitree G1 launched at a $16,000 base price. Some configurations at the same product page now list higher prices for specific compute upgrades or hand modules that were shown as part of the base robot in early demo footage. The exact hand configuration that appeared in the 2024 dance video is currently not the base configuration.

Boston Dynamics Atlas Electric does not publish a price, so it cannot drift. That is its own information problem, but a cleaner one.

How to detect spec drift before it hits your contract

Every buyer who plans to spend more than $50,000 on a humanoid should do four things.

Wayback Machine the product page monthly. Snapshot the shipping product page at the moment you begin evaluation and again at the moment you sign. If any number has changed in between and the vendor did not flag it, that is a warning. The Wayback Machine covers most manufacturer sites; where it does not, a screenshot of the page saved to your procurement folder does the same job. Compare reveal deck against current datasheet, line by line. For any humanoid still in early production, pull the reveal-day press release and the current product page side by side. Diff every number. If a number has moved without an announcement, ask the vendor for a written explanation and put it in the contract as a clarification. Ask for a per-unit spec attestation at signing. Enterprise humanoid contracts should include a per-unit spec sheet attested by the manufacturer at the date of shipping, not at the date of contract signing. This costs nothing to draft and closes the drift window between the sale and the delivery. Most manufacturers will agree because the specific unit specs are already tracked internally for warranty purposes. Watch the hands. Hand DOF, fingertip sensor spec, and hand payload are the most-drifted numbers on humanoid product pages because they are the hardest to test in a lab demo and the easiest to revise between generations. If a vendor's marketing images show one hand and their shipping units carry a different hand, that is worth catching before signing.

What spec drift tells you about the manufacturer

Consistent, documented spec changes across generations are a good sign. Agility Digit, Boston Dynamics Atlas Electric, and Unitree H1 each show clean generation-to-generation notes in their public archives. Buyers can plan around documented change.

Undocumented spec changes tell you two things. First, the manufacturer's internal engineering process treats the spec sheet as marketing rather than as a contract. Second, the sales team either does not know or does not report the changes upward. Neither is a fatal flaw, but both are inputs to how much room you leave in your integration schedule and your budget.

The humanoid market in 2026 is still small enough that reputation moves slowly. Manufacturers who mark every spec change publicly, however awkward the change is, are building trust that will compound. Manufacturers who quietly revise numbers are training their customer base to distrust the datasheet. The buyers who read spec sheets carefully already know which category each vendor is in. The question is whether the vendors know.