OpenAI released GPT-6 Astra, its newest large model. Latent.Space received early access and tested it on practical work. The team spent more than 20 billion tokens across a month of projects. It reports that Astra can pick and train models, label data and read logs. It also deployed and debugged whole systems in single runs. One lead agent directed 20 to 50 smaller agents at once. Latent.Space estimates a cost under $6 per hour at observed speed and top token rates. It compares that to $200-$1,000 a day for a junior engineer. Astra saturated the hardest FrontierMath and ARC-AGI-3 tests. OpenAI's Greg Brockman said AGI is here.
What changed
Earlier models could not hold coherence across billions of tokens or manage fleets of subagents.
What it unlocks
Running one lead agent that starts, monitors and fixes dozens of parallel agent runs.
- 97.6% on hardest FrontierMath
- 99.9% on ARC-AGI-3
- 20B+ tokens tested
- 20-50 parallel agents
What you need to act on it
- OpenAI API access
- budget for high token volume
Sources