Progress

report · 2026-09-03

GPT-6 Astra System Card

OpenAI

Read the original ↗Publication details
On this page

Why it matters here

A task-specific AI self-improvement suite replaces older measures. Debugging, kernels, pretraining and post-training test different abilities.

Read original source ↗

AI-R&D capabilityAI-system improvement

What the charts show

Training improvements at different token budgets

NanoGPT compares GPT-6 Astra and GPT-5.6 Sol on improving a small language-model training setup. The curves show how reported reward varies with the output-token budget.

Mean reward against output tokens in OpenAI’s NanoGPT evaluation. The budget is part of the result; this is not a single comparable score.

Source curves comparing the two models’ mean reward across output-token budgets.
Original OpenAI figure from the system card, retained unchanged. No point values inferred from pixels. Enlarge chart ↗

View original chart and methods ↗
GPT-6 Astra System Card · Section 10.1.3.3, Figure 54

How to interpret this chart
  • Exact point values are not printed. The tracker retains the figure rather than estimating scores from its geometry.
  • This is a task-specific training evaluation, not evidence of an autonomous improvement loop.

Post-training improvements at different token budgets

PostTrainBench Lite compares GPT-6 Astra and GPT-5.6 Sol on post-training work. Its curves show performance across different output-token budgets rather than one standalone score.

Mean reward against output tokens in OpenAI’s PostTrainBench Lite evaluation.

Source curves comparing the two models’ mean reward across output-token budgets.
Original OpenAI figure from the system card, retained unchanged. No point values inferred from pixels. Enlarge chart ↗

View original chart and methods ↗
GPT-6 Astra System Card · Section 10.1.3.4, Figure 55

How to interpret this chart
  • Exact point values are not printed; no numeric results have been estimated from the plot.
  • Results apply to this evaluation and budget, not the entire process of improving a frontier model.
Internal Research Debugging EvaluationResolve real bugs in internal research experiments.KernelGen 1POptimize correct kernels for OpenAI first-party hardware.NanoGPTReduce small-model training time to a target validation objective.PostTrainBench LiteImprove a pretrained model within a five-hour GPU budget.MLE-benchML competition engineering under bounded compute.

What to keep in mind

  • Internal task details and some numeric values remain unavailable in this extraction.