# GPT-6.2 Astra scores 41.2% on Ember-0 robotics benchmark, more than doubling its predecessor

> Source: <https://cryptobriefing.com/gpt-6-2-astra-ember-0-robotics-benchmark/>
> Published: 2026-11-20 20:34:27+00:00

**Key takeaway:** GPT-6.2 Astra reached 41.2% on XDOF's Ember-0 versus 19.6% for GPT-6 Astra, 79/100 on AA-Robotics v0.1, and cut average prediction latency from 15.0 to 10.1 seconds, but 10.1 seconds remains too slow for real-time reaction.

# GPT-6.2 Astra scores 41.2% on Ember-0 robotics benchmark, more than doubling its predecessor

Photo: Steve A Johnson / Pexels

OpenAI's latest Astra model posts big gains in robot control, including the first reported success on deformable-object tasks

[OpenAI](https://cryptobriefing.com/markets/openai/)’s GPT-6.2 Astra has scored 41.2% on Ember-0, a robotics benchmark maintained by XDOF. Two months ago, its predecessor managed 19.6%.

## What the numbers show

XDOF reported the Ember-0 result on November 20, 2026. The previous model, GPT-6 Astra, posted its 19.6% in September 2026.

On a separate benchmark, AA-Robotics v0.1, GPT-6.2 Astra scored 79 out of 100. It also used fewer arm commands than the other models tested.

Speed improved too. Average prediction latency across the Astra series fell from 15.0 seconds to 10.1 seconds between model versions.

## The t-shirt problem

The most striking claim attached to GPT-6.2 Astra is qualitative rather than numerical. It is described as the first model to demonstrate success on deformable-object tasks.

Those include handling a t-shirt and tying a knot. To a human, these are chores. To a robot, they are notoriously hard.

### AI, tech, and the markets they move—in one daily briefing.

Daily. Free. Join 34,000+ readers across crypto, finance, and policy.

A rigid object like a block or a cup keeps its shape, so a model can predict where every part of it will be after a move. Fabric and rope do not cooperate. Every tug changes the shape of the whole object, which means the model has to keep re-reading the scene and adjusting on the fly.

## Background: language models as robot brains

The Astra results land as frontier AI labs increasingly test large language models as controllers for robots. The idea is that a model trained on vast amounts of text and imagery can translate instructions into physical actions.

The GPT-6.x line has been part of that push. The progression from GPT-6 Astra to GPT-6.2 Astra reflects ongoing refinement within the series. Benchmarks like Ember-0 and AA-Robotics v0.1 exist to keep that confidence honest. They put models through standardized tasks so that progress can be measured instead of merely claimed.

## What this means

The headline number deserves both credit and context. A 41.2% score is a large improvement, but it also means the model still falls short on the majority of what Ember-0 measures.

Latency is still the obvious weak point. An average of 10.1 seconds per prediction is far better than 15.0 seconds, but it remains slow for anything that needs to react in real time.

A model that leads on Ember-0 while also posting a 79 out of 100 on AA-Robotics v0.1 puts pressure on rivals to show comparable physical-world performance, not just strong chatbot scores.

**Disclosure:** This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our [Editorial Policy](https://cryptobriefing.com/editorial-policy/).
