GPT-5 Is Here, and Most Coverage Is Missing the Point

The math and science improvements matter too. OpenAI claims GPT-5.2 solved an "open theoretical problem" and can generate "reliable mathematical proofs." For robotics, this translates to better physics reasoning, more accurate motion planning, and potentially fewer instances of robots doing that thing where they confidently attempt to put a square peg in a round hole.

The Enterprise Framing Is a Tell

OpenAI is positioning this entire release around "the new era of work" and "enterprise AI." That's marketing, obviously. But it's also revealing.

The robotics companies I talk to are increasingly enterprise customers. Figure is targeting manufacturing. Apptronik is going after logistics. Even Boston Dynamics has pivoted hard toward commercial applications. When OpenAI talks about "agentic workflows" and "automation," they're describing exactly what these companies need.

I don't have confirmation that any major humanoid companies are using GPT-5 yet (it literally just came out), but it would be surprising if they weren't at least evaluating it. The previous generation of models was already being tested for robot reasoning. This one is, on paper, substantially better.

What We Don't Know Yet

Here's where I have to be honest: there's a lot we can't assess from blog posts and benchmark scores.

Latency remains unclear. A model that takes 10 seconds to respond is useless for real-time robot control. OpenAI mentions efficiency improvements but doesn't provide the specific numbers that would matter for embodied applications.

Reliability is another open question. Robots can't afford hallucinations the way chatbots can. A customer service bot that makes something up is embarrassing. A humanoid that hallucinates a staircase is dangerous. The system card for GPT-4.5 discussed safety testing, but I haven't seen equivalent documentation for GPT-5 yet.

And then there's the cost question. Running frontier models is expensive. Running them continuously on robots operating in the real world is very expensive. Whether the performance gains justify the compute costs for embodied AI applications is something we just don't know yet.

The Bigger Picture

I think what's happening here is a gradual convergence that most people aren't tracking.

Language models are getting better at reasoning about physical space. Robotics companies are getting better at integrating language models into control systems. The hardware is maturing. The context windows are expanding. The latency is dropping.

None of these developments is individually revolutionary. But taken together, they suggest we're approaching a threshold where general-purpose AI models become genuinely useful for general-purpose robots.

GPT-5 isn't a robotics product. OpenAI isn't a robotics company (despite the persistent rumors). But this release might matter more for humanoids than any actual robotics announcement this year.

Or I could be wrong. It's too early to say with any confidence. The benchmarks are impressive, but benchmarks have a way of not translating to real-world performance. The capabilities sound relevant, but sounding relevant and being useful are different things.

What I do know is that the robotics engineers I've been texting are paying attention. That probably means something.

Mark Kowalski · 1 hour ago · 5 min