OpenAI Confirms It Will Build a Humanoid Robot — What That Actually Means

Twenty months ago, OpenAI's robotics story was one sentence long: it had invested in Figure AI and licensed its models to power Figure's humanoid. It built no hardware of its own. That sentence no longer holds.

On a September 2, 2026 episode of the Sources podcast, Sam Altman said it plainly: "We will definitely do a humanoid. We will do other form factors as well." That's the most direct confirmation yet that OpenAI intends to build and ship its own robot hardware — not just supply the models that run inside someone else's. It's a genuine pivot for a company whose entire public identity has been chatbots, APIs, and coding agents, and it's worth being precise about what's actually confirmed versus what's still open.

TL;DR — what's confirmed vs. still open

What Altman actually said, and how confident to be

Precision matters here because robotics stories about frontier labs tend to snowball from "reportedly exploring" to "confirmed" faster than the reporting justifies. So here's the ladder, rung by rung:

• Confirmed, on the record: Altman's quote itself — "we will definitely do a humanoid" — plus the existence of a named internal robotics division since May 2026.
• Reported, not officially detailed: the division's headcount, specific role postings (actuator design, sim-to-real engineers, data acquisition operations), and its technical roadmap (custom actuators, MuJoCo/Isaac Sim-based sim-to-real pipelines).
• Still unknown: ship date, unit target, manufacturing partner, price, and even the final form factor. Altman's near-term framing points at industrial and infrastructure work — "help skilled workers build things" — with household robots described explicitly as a longer-term ambition ("someday... I think everyone should have a personal robot").

That's a real strategic commitment from the CEO, not a rumor — but it is also nowhere near a product announcement. Treat every downstream headline that implies a launch date as getting ahead of what's actually been said.

The Figure AI breakup that set this in motion

This story doesn't start in September 2026 — it starts with a partnership falling apart 19 months earlier. OpenAI and Figure AI announced a collaboration in February 2024 to combine OpenAI's models with Figure's humanoid hardware, with Figure's robots running OpenAI-powered natural-language reasoning. That relationship ended in February 2025: Figure CEO Brett Adcock said Figure had made "a major breakthrough on fully end-to-end robot AI, built entirely in-house," arguing that embodied AI can't be effectively outsourced and that vertical integration — owning both the model and the body — was the only path to real progress.

That argument turned out to cut against OpenAI staying on the sidelines, not for it. Rather than retreating from robotics after losing its hardware partner, OpenAI absorbed the lesson: if vertical integration is what it takes, build the hardware. The Figure Index robot dataset and Figure's own subsequent momentum (see Figure's robots outnumbering humans at its own facility) show what a lab looks like when it commits fully to owning the full stack — and OpenAI's May 2026 division launch reads as a direct answer to that.

Inside OpenAI's robotics division

Altman confirmed OpenAI's dedicated robotics division on May 31, 2026. It's led by Aditya Ramesh — the researcher behind DALL-E and, later, Sora — and grew out of OpenAI's world-simulation research rather than being spun up from nothing. Since launch, OpenAI has been hiring for specialized physical-hardware roles in San Francisco: actuator design engineers, simulation-realism engineers, and data-acquisition operations managers, among others, building toward "sim-to-real" pipelines on physics engines like MuJoCo and Isaac Sim.

The division has not been without turbulence. Caitlin Kalinowski, who joined OpenAI from Meta's AR glasses group to work on hardware, resigned in March 2026 over concerns tied to OpenAI's Pentagon contract work and what she described as insufficient guardrails around surveillance and autonomous-weapons use — a reminder that "OpenAI builds robots" carries defense and dual-use questions that a pure chatbot company never had to answer this directly. That tension sits alongside broader industry debate on autonomous weapons rules from the UN and Red Cross.

OpenAI's stated goals split into two horizons: near-term, robots that support skilled workers on infrastructure and construction sites; long-term, general-purpose personal robots. Altman's own framing for why humanoid specifically: most of the built environment — door handles, kitchen counters, keyboards, vehicle controls — assumes a human body operating it, so a humanoid form factor is the most general way to drop a capable AI system into that environment without redesigning the world around it.

Not the same story as the Jony Ive device

It's easy to conflate OpenAI's two hardware pushes, but they're distinct efforts. The robotics division is about physical robots that manipulate the world. Separately, OpenAI has been developing a personal AI companion device with former Apple design chief Jony Ive, reportedly in a deal worth roughly $6.5 billion — see explainx.ai's earlier coverage of OpenAI's hardware companion device plans. That device is a screenless, ambient companion gadget, not a robot with actuators and limbs. Both efforts share a root cause, though: OpenAI concluding that the next competitive edge isn't just a better model behind an API, but a form factor it fully controls.

Why this matters for the software-agent world explainx.ai covers

This is where the story stops being robotics trivia and starts being relevant to anyone building with today's agent tooling. The last 18 months of AI product work have been about giving models "hands" inside software — Claude's computer use, background tasks, and Cowork, and OpenAI's own computer-use framing baked into GPT-6 Astra's positioning. The pitch has consistently been: an agent that can click, type, and navigate a screen the way a person does, instead of calling narrow APIs.

A frontier lab moving from "computer use" to "physical use" is the next logical rung on that same ladder, not a separate business. The skills a computer-use agent already needs — perceiving an environment, planning multi-step actions, recovering from failure, operating tools built for humans — are the same skills a humanoid robot needs, just with a body attached to the loop instead of a browser tab. OpenAI's own research groundwork already blurs this line: its world-simulation research underpins both the robotics division and the broader industry move toward world models as a category, alongside NVIDIA's Cosmos and other physical-AI world-model efforts.

For practitioners, the practical takeaways are:

• Agent tooling and embodied AI are converging technically, even if they ship as different products. The orchestration patterns, tool-use loops, and failure-recovery techniques you build for software agents are increasingly the same primitives robotics teams reach for.
• Data collection is the real bottleneck, and it's why labs want hardware. Simulation only gets you so far; owning physical robots means owning the real-world interaction data that trains the next generation of embodied models — the same logic driving robotics data-collection efforts across the industry, and one reason robotics still isn't a solved problem despite rapid model progress.
• Expect more labs to follow this path. Anthropic, Google DeepMind (Gemini Robotics), and NVIDIA are all already active in physical AI in some form. OpenAI committing publicly to hardware raises the pressure on every other frontier lab to state its own position rather than staying quiet about robotics ambitions.
• This is a multi-year bet, not a near-term product to plan around. No practitioner needs to change a roadmap today because of a podcast quote. What's worth tracking is whether OpenAI's division produces an actual prototype, and how it positions relative to Figure, Tesla's Optimus, 1X, and Unitree in the meantime.

The competitive picture

OpenAI re-entering hardware puts it in a field that's already crowded and fast-moving. Figure has kept shipping without OpenAI, Tesla's Optimus program continues its own iteration cycle, and startups like 1X and Unitree have been shipping humanoid hardware at a pace that outstrips most software labs' comfort zone. OpenAI joining as a builder rather than an investor changes the dynamic: it's now a direct competitor to companies it used to fund, and a peer to Google DeepMind's own robotics push rather than purely a model supplier to the category.

None of the industry reaction so far amounts to a confirmed on-the-record statement from a named rival executive reacting specifically to Altman's September 2 quote — most of what's circulating is analyst and outlet commentary framing OpenAI as "late" relative to Figure and Tesla, or noting that OpenAI's simulation-first approach was exactly what pushed Figure to go it alone in the first place. That history is worth remembering as OpenAI's own hardware effort matures: building humanoid robots has repeatedly proven harder in practice than it looks from a research lab, which is a large part of why robotics still isn't solved even with today's model capabilities.

What to watch next

• A named prototype or demo. Until OpenAI shows hardware, "we will definitely do a humanoid" remains a stated intent, not a product.
• Manufacturing and supply-chain partners. Building humanoid hardware at any scale requires actuator, sensor, and assembly partnerships OpenAI doesn't currently have in public view.
• How the Pentagon/defense question resolves. Kalinowski's resignation flagged a real internal tension between OpenAI's robotics ambitions and its defense contracts that hasn't been publicly settled.
• Whether Figure or other robotics-first companies respond. A well-funded former partner turned competitor is a dynamic worth watching in its own right.

Details in this post reflect reporting and statements available as of September 4, 2026. Robotics timelines, executive statements, and organizational structures at fast-moving AI labs change quickly — treat specific dates, titles, and role counts as accurate as of publication rather than permanent facts.

Related reading

• Figure's Robots Outnumber Humans at Its Own Facility
• Figure Index Robot Dataset
• OpenAI's Hardware Companion Device with Jony Ive
• GPT-6 Astra Launch: Benchmarks, Pricing, Rollout
• Claude Background Computer Use, Cowork, and Code
• What Are World Models? Complete Guide
• World Labs Atlas: Multimodal World Model
• Why Robotics Still Isn't Solved
• Gemini Robotics 2: Whole-Body Intelligence
• 1X NEO: 25-DoF Hands and Physical API