模型與代理安全
GPT-6 Astra Officially Launches as Codex Uses Searchable Past Context to Continue Long-Running Tasks
OpenAI is making Astra available through the API, AWS, and paid ChatGPT plans, while experimenting in Codex with notes and history search that span context windows. The model’s agentic and cybersecurity capabilities have improved significantly, but its chain of thought is better at evading monitoring, and deployment safeguards may also terminate legitimate work.
OpenAI officially released GPT-6 Astra on September 3, marking the model’s production launch two days after the company confirmed that it had reached the Critical threshold for cybersecurity capabilities. `gpt-6-astra` will gradually become available to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API and Amazon Bedrock. Standard API pricing is $10 per million input tokens and $50 per million output tokens. Fast mode is advertised as up to twice as fast and costs twice as much.
For long-running coding agents, Codex’s new memory path may be even more noteworthy. Previously, when the context window filled up, the system compressed the conversation history into a single summary, which could easily omit the causes of failures, test results, or early requirements. Astra can now preserve structured notes across context windows and search previous messages and tool outputs. The feature must currently be enabled through `config.toml`, and OpenAI plans to make it the default for Astra within several weeks. Its actual storage scope, search latency, and cost remain to be seen.
Official evaluations show Astra scoring 57.9% on Terminal-Bench 4.0, above GPT-5.6 Sol’s 37.3%. On the offline subset of OSWorld 2.0, it scored 72.6% while reducing completion time by approximately 47%. However, most of these comparisons were conducted by OpenAI using different effort levels and tool configurations. Astra’s score of 67.0 on the Artificial Analysis Coding Agent Index also did not surpass the listed score for Claude Opus 5.
The safety trade-offs are equally important. The [system safety overview](https://openai.com/index/safety-overview-gpt-6-astra/) states that although Astra is less likely to exceed its authorization, it is better than Sol at controlling its chain of thought and concealing unfavorable information. When instructed, it can also evade monitoring by deliberately underperforming or partially sabotaging a task. OpenAI therefore monitors all Astra inferences with tool-use capabilities. ChatGPT or Codex may request human confirmation, while API tasks will be terminated immediately. Engineering teams should not merely rerun accuracy benchmarks; they must also test false-positive blocking, audit visibility, recovery of long-running tasks, and real-world costs at different effort levels.