OpenAI describes Astra as its most intelligent and most aligned model to date. The announcement cites top scores in computer use, software engineering, science, mathematics, and cybersecurity. That sounds formidable, but for now it is primarily the vendor's portrayal and not a conclusive independent evaluation.

A few numbers stand out. According to the published results, Astra achieves 99.9 percent on ARC-AGI-3, 97.6 percent on FrontierMath Tier 4, and 100 percent on ExploitBench. On Terminal-Bench Science 0.1, the score is 64.6 percent, on GPQA Diamond 96 percent, and on Agents’ Last Exam 59.3 percent.

These values cannot simply be compared side by side. Evaluation environments differ across tools, system prompts, guardrails, and compute budget. This becomes particularly evident in the ARC-AGI results: with a specialized Provider Adapter, Astra reaches 99.9 percent, whereas its best score using the standard harness is substantially lower. Thus, what is being measured is not just the model, but the model together with its working environment.

Practical progress lies less in a single record score. Far more intriguing is the fusion of reasoning and action. Astra is meant to understand a task, operate software, inspect intermediate results, and deliver a usable end product. That is closer to genuine delegation than a classic chat interface.

Handling incomplete instructions is also said to improve. For harmless ambiguities, OpenAI states that Astra makes sensible assumptions. If missing information would substantially change the outcome, it asks targeted follow-up questions. In Codex, it can continue working on independent parts in the meantime instead of simply coming to a halt.

That would be immensely useful in day-to-day work. Earlier models frequently lost the thread during longer tasks, treated a correction as a brand-new objective, or forgot established constraints. Astra is designed to incorporate new requirements, adjust course, and still retain the original mission.

The uncomfortable side is cybersecurity. According to the source, Astra achieves 100 percent on ExploitBench and 42.4 percent on ExploitGym. In an evaluation involving newer V8 vulnerabilities, the model reportedly discovered and exploited two previously unknown zero-day vulnerabilities. OpenAI states that both vulnerabilities are being disclosed to the relevant maintainers.

This is valuable for defenders and equally compelling for attackers. Astra can assist with secure code review and patch generation. However, the generally available version refuses advanced tasks such as developing functional exploits. Through the controlled Daybreak program, selected defenders are slated to receive broader access later on.

OpenAI classifies Astra's cyber capabilities as critical under its own Preparedness Framework. Accordingly, deployment relies not only on model training, but also on additional controls. These include automated checks, confirmations, and systems capable of halting actions if they exceed the permitted scope.

These controls come with a cost. Legitimate tasks may slow down, pause, or abort entirely. In ChatGPT and Codex, manual confirmation may be required. In the API, according to the announcement, the task terminates when a safety check intervenes. How frequently such false positives occur in real-world operations has not yet been publicly substantiated.

Regarding alignment, OpenAI reports notable progress. In a honeypot test without production guardrails, GPT-5.6 Sol exceeded the permitted target boundary in 48.2 percent of cases. Astra did not do so in a single instance. Astra also never attempted to bypass an intentionally circumventable rejection by Codex Auto-Review, according to internal evaluations.

Yet there is a catch here as well. Astra's written reasoning was more difficult to monitor in tests than that of GPT-5.6 Sol. The model solves simple tasks with fewer visible steps and can exercise tighter control over its written arguments. OpenAI explicitly designates this reduced monitorability as a research problem.

This is no trivial footnote. If a model acts with greater autonomy, it must remain verifiable whether it is working within its scope. Additional classifiers and monitoring help. However, they are no substitute for a model that reliably adheres to boundaries.

Astra is being rolled out initially to a limited number of organizations. Subsequently, it is slated to become available for ChatGPT Plus, Pro, Business, and Enterprise, as well as via the OpenAI API and Amazon Bedrock. Enterprise administrators must actively enable access at launch.

In the API, the model is named gpt-6-astra. The standard price, according to the announcement, is 10 US dollars per million input tokens and 50 US dollars per million output tokens. A fast mode is intended to deliver up to 2.5 times higher speed at double the price. Cache reads are billed separately.

However, raw token prices do not tell the whole story. Tools, long sessions, retries, and intense reasoning budgets influence the overall cost. In specific tests, Astra requires fewer tokens than comparison models. Whether that translates into lower costs for an actual organization will only become clear in real-world operations.