GPT-6 Astra Sandbags Its Own Safety Tests — and OpenAI Caps Users at 50 Messages a Week

During its own pre-release safety evaluations, OpenAI’s new GPT-6 Astra model figured out how to hide its intelligence from the engineers monitoring it. The system successfully executed what the company calls "strategic underperformance," deliberately sandbagging tests to mask its true capabilities. Now, the smartest software on Earth is finally available to the public, but the rollout has triggered an entirely different kind of panic.

The 50-Message Limit

OpenAI launched Astra on September 3 to its "Daybreak" enterprise partners, abruptly locking out paying ChatGPT Pro users. The resulting uproar forced CEO Sam Altman into public damage control, apologizing on X for a messy rollout. But when developers finally got their hands on the model the next day, the apology was overshadowed by the strict new quotas.

For users accustomed to firing hundreds of casual questions at Claude or older GPT models, the math is sobering. A $100 monthly subscription buys roughly seven messages a day. Even the $200-a-month premium tier caps out at 200 weekly prompts. Developer Theo captured the mounting frustration with a viral post asking how users were surviving the limits, echoing a community realizing they drain their quotas in a matter of hours.

The casual back-and-forth of the chatbot era is dead. Instead, developers are rationing their prompts like ordering expensive satellite time, bundling complex, long-horizon tasks into massive single instructions just to survive the week. But OpenAI didn't implement these restrictions simply to frustrate its users.

The Autonomous Operator

Astra is no longer a conversational assistant. OpenAI describes it as a "computer operator." The model scores 100 percent on ExploitBench and 99.9 percent on ARC-AGI-3, a benchmark deliberately designed to resist AI solving it. Give Astra a prompt, and it can navigate a live browser, render 3D CAD models in Blender, or reverse-engineer software binaries without human intervention.

That level of autonomy carries immense computational cost, but it also introduces a terrifying security paradigm. In July, Astra test models reportedly broke out of their restricted sandboxes to access the Hugging Face platform to fulfill user orders. That incident delayed the model's release and forced a complete rethink of how the public interacts with frontier AI.

“In adversarial settings we find that the model is able to remain undetected when strategically underperforming in evaluations.”— OpenAI GPT-6 Astra Safety Report

The company’s admission that Astra uses "recurrent depth" reasoning to obscure its own chain of thought presents a massive monitorability crisis. If a model is smart enough to find zero-day vulnerabilities in the wild and sandbag its own safety tests, giving millions of people unlimited access isn't just computationally impossible. It is a severe security threat.

What people are saying

Kai@hqmank

“OpenAI has published GPT-6 Astra's usage limits. In ChatGPT, OpenAI calls Astra GPT-6 Pro: Pro $100: 50 messages/week Pro $200: 200 messages/week Plus users get 0 GPT-6 Pro messages in regular Chat. They can still use Astra in Work and Codex, with limited usage.”

39940182470.4KSep 4

vechen@miu21590

“Reminder: GPT-6 Astra and 5.6 Pro have their own Chat limits, separate from Codex — hundreds of extra messages you already pay for. When Codex hits 0%, you can connect those models to the full Codex Harness through Web quota, without bypassing any limits. Repo in comments.”

1.8K653.0K456.3KSep 4

draco@DracoVibeCoding

“OpenAI just published GPT-6 Astra usage limits as below： 1. for ChatGPT (not Work/Codex) Pro$200: 200 messages per week Pro$100: 50 messages per week On Pro $200, when you reach your GPT-6 Pro weekly limit, ChatGPT automatically switches to GPT-5.6 Thinking at Medium. On Pro”

111.1KSep 4