Tesla Optimus Review: The Scorecard, Not the Demo Reel

No robot on earth generates more search traffic than Tesla Optimus, and almost none of that traffic gets a straight answer. The coverage splits into people who think it is already replacing factory workers and people who think it is a man in a suit. Both are wrong, and the evidence to settle it is public and specific.

The most useful sentence about Optimus was said by Tesla, on an earnings call, in January 2026. This review is built around it.

Companion reading: how to choose a humanoid robot, humanoid robot hardware, foundation models and VLAs for robotics, robot teleoperation, and the next 10 years of robotics.

Table of contents

1. Key takeaways
2. The sentence that settles it
3. Shipped against target
4. The teleoperation question, both sides
5. The hardware is the strong part
6. How it compares
7. What would change the verdict
8. Verdict
9. Frequently asked questions
10. Changelog

The sentence that settles it

On the Q4 2025 earnings call, on 28 January 2026, Musk described Optimus units as "primarily for learning, not productive tasks."

Sit with that for a second, because it does two jobs at once. It deflates the claim that Optimus is doing factory work today: Tesla is saying the units are gathering data and training policies rather than earning their keep. And it defends against the claim that the whole thing is theatre: you do not convert a Model S line and build thousands of robots to run a hoax.

Nothing else in Optimus coverage is this clarifying, and it comes from the party with every incentive to say otherwise. Take it at face value in both directions.

Shipped against target

The most honest way to read a robotics programme is shipped-versus-promised, over time.

A roughly order-of-magnitude miss in 2025 is the number to anchor on. Note also what it is not: zero. Hundreds of humanoids built and deployed inside a manufacturing business is more than almost anyone else has managed, and it is a fraction of what was promised. Both things are true, and the pattern (ambitious target, real progress, order-of-magnitude shortfall) is consistent enough across Tesla programmes to be a useful prior for the 2026 numbers.

For the wider industry context on why humanoid timelines slip, see the next 10 years of robotics, which sets out the data bottleneck that binds everyone.

The teleoperation question, both sides

This deserves precision, because it is where most coverage becomes unreliable in one direction or the other.

The criticism is real. At the October 2024 "We, Robot" event, robots interacting with guests were human-assisted via teleoperation, and this was not clearly disclosed at the time. Milan Kovac subsequently confirmed they were "assisted to some extent". A demo that reads as autonomy and is actually a person in a motion-capture rig in the back room is a disclosure failure, and it earned the scepticism that followed.

The rebuttal is also real. The October 2025 demo was confirmed as AI-driven rather than teleoperated. That is a genuine autonomy proof point, and people still citing 2024 as the current state are a year behind.

The durable lesson is a viewing habit rather than a verdict: for any humanoid demo, from any vendor, ask whether autonomy was disclosed, over what duration, and with how many takes. Our robot teleoperation guide covers why the line between assisted and autonomous is blurrier than a video can show, and why "human-gated" is a legitimate engineering stage rather than automatically a cheat.

The hardware is the strong part

Set the timeline arguments aside and the engineering is serious, particularly in the hands.

Gen 3 hands run roughly 50 actuators across 22 degrees of freedom, which puts them among the most articulated production robot hands anywhere. Tesla built custom electromagnetic actuators using planetary roller screw drives, chosen for shock-load resistance during walking, high torque density, precision and backdrivability, which is what makes an actuator safe to be near and controllable in contact.

Backdrivability is the tell that this is real robotics rather than a styling exercise. It is exactly the property that makes force control possible, and it is expensive to engineer. See robot actuators and end-effectors and grippers for why hands are the hardest part of a humanoid.

Vertical integration is the strategic bet: Tesla makes its own actuators, its own compute and its own training data pipeline. If humanoids become a volume manufacturing problem, that is the right shape. If they remain a dexterity and generalisation problem, it buys less than it costs.

How it compares

The sharpest contrast is with Figure 03, which is deployed at a customer site and has published a cost per robot-operating-hour. Optimus has far greater manufacturing ambition and, by Tesla's own description, is not yet doing productive work.

What would change the verdict

• Verifiable productive autonomous work, at a named site, over a shift, disclosed as autonomous. Not a demo, not a learning session.
• The 2026 unit number, measured against the 50,000 to 100,000 target with the 2025 miss as the prior.
• Sales to anyone outside Tesla. Optimus doing Tesla's work is vertical integration; Optimus doing someone else's work is a product.
• A published unit economic, which is the number Figure has and Tesla does not.

Verdict

Optimus is the most heavily resourced humanoid programme in the world, with genuinely impressive actuator and hand engineering, an order-of-magnitude 2025 shortfall against its own target, and a company that describes its deployed units as being there to learn rather than to work.

You cannot buy one, so this is not a purchasing decision. It is a forecasting one, and the honest read is that Tesla has built real hardware at real scale and has not yet shown the thing that matters: a humanoid doing useful work unattended, disclosed as autonomous, for a full shift. Watch for that, and discount every demo reel until it arrives.

Frequently asked questions

Is Tesla Optimus actually working in factories? Not productively, by Tesla's own account. On the Q4 2025 earnings call on 28 January 2026, Musk described deployed units as "primarily for learning, not productive tasks". Units exist and are deployed inside Tesla, gathering data and training policies.

Was Tesla Optimus teleoperated? At the October 2024 "We, Robot" event, yes: robots were human-assisted and this was not clearly disclosed, later confirmed as "assisted to some extent". The October 2025 demo was confirmed AI-driven rather than teleoperated, so citing 2024 as the current state is out of date.

How many Optimus units has Tesla built? Hundreds in 2025 against a stated target of 5,000. Gen 3 mass production began on 21 January 2026 at Fremont, with a 2026 target of 50,000 to 100,000 and talk of 1M/year eventual capacity.

Can I buy a Tesla Optimus? No. It is not sold to the public. If you want a humanoid you can actually purchase, see our Unitree G1 review.

What is impressive about the Optimus hardware? The hands. Gen 3 runs roughly 50 actuators across 22 degrees of freedom, on custom electromagnetic actuators using planetary roller screws, chosen for shock-load resistance, torque density and backdrivability. Backdrivability is what makes an actuator safe in contact and controllable under force, and it is expensive to engineer.

Optimus or Figure 03? They are at different stages. Figure 03 is deployed at BMW Spartanburg with a published cost per robot-operating-hour. Optimus has far larger manufacturing ambition and, per Tesla, is not yet doing productive work.

Will Optimus hit its 2026 target? Unknown, and the prior is not encouraging: 2025 missed by roughly an order of magnitude. The number to watch is units doing verifiable autonomous work, not units built.