OpenAI found three problems behind GPT-6 Astra's quality drop

For a week, GPT-6 Astra users said the new model had gotten worse. On September 12, OpenAI confirmed three real problems behind the complaints.

What users reported

GPT-6 Astra launched on September 3, 2026. About a week later, developers began posting examples of the model stopping in the middle of tasks and answering older messages instead of the latest one. In one GitHub report, a user said the model claimed work was done when it was not. At that stage these were user reports, without confirmation from the company.

The three problems OpenAI found

OpenAI’s Tibo Sottiaux posted the findings on September 12.

The first involved skills. Some skills written for previous models were triggering too often or preventing the model from checking its work.

The second was an experiment. An opt-in context management feature could cause early stops or replies to older messages. OpenAI disabled it and estimates that 4,000 to 5,000 users were affected.

The third was in serving. Badly configured engines caused a measured quality drop for a long tail of traffic, and OpenAI removed them. The company also reset usage limits.

What this means for teams paying for coding agents

The three causes sit in different layers. One came from instructions written for older models, one was a product experiment, and one was infrastructure. Together they changed what users received while the product kept the same name.

That is the practical lesson for anyone paying for an AI coding agent. The model tested on launch day is not guaranteed to be the model in use a week later, because settings, experiments and serving infrastructure can change underneath it. In this case, users noticed the decline before the company confirmed it.

Teams that depend on these tools have reason to keep a small set of their own repeatable tasks and rerun them after launches and updates, rather than waiting for the vendor to say that quality has slipped.