OpenAI’s GPT‑6 Astra is positioned as a computer‑use agent that can drive browsers, spreadsheets, terminals and desktop apps to finish multi‑step jobs without asking for step‑by‑step instructions. For developers and IT teams the biggest practical question is whether Astra can be dropped into existing workflows today.
Astra is not self‑hostable. It is a closed, hosted model available only to organisations in OpenAI’s Trusted Access and Daybreak programs. If you are outside those groups you cannot run it on your own hardware or via a private API key. Access is granted through OpenAI’s managed endpoints, which means you must rely on their service level agreements and data handling policies.
The model introduces a new context‑management approach: instead of compacting and discarding earlier turns, Astra keeps searchable notes across context windows. This reduces the loss of critical details such as why a fix failed or which tests ran, a common failure point for earlier agent systems. The feature is experimental now but will become the default in Codex soon. Developers can enable it via a simple config.toml flag.
Astra can also ask clarifying questions while continuing unrelated work, preventing a single unresolved decision from stalling an entire job. This addresses a frequent agent bottleneck where waiting for user input halts progress.
Tool coverage is broad: computer use, hosted shell, apply patch, skills, MCP and tool search. However, fine‑tuning is not supported, so you cannot specialise the model on proprietary codebases or internal APIs without using prompting or retrieval‑augmented techniques.
Performance numbers show gains in OSWorld (72.6% vs 65.7% for GPT‑5.6 Sol) and strong results on ARC‑AGI‑3 (98.6%) and FrontierMath Tier 4 (97.6%). Coding remains a relative weakness: DeepSWE v1.1 scores 74.1% for Astra versus 75.4% for Muse Spark 1.3, a gap of only one or two tasks on a 113‑task benchmark.
Cost is $10 per million input tokens and $50 per million output tokens, with cached input at $1.00. Requests above 272 K input tokens are billed at 2× input and 1.5× output for the whole request. Batch and Flex modes run at 50% of the base rate, while Fast mode doubles the price.
Security‑wise, Astra is the first OpenAI model to reach the Critical cybersecurity threshold in the Preparedness Framework. It can discover exploits and has found previously unknown V8 vulnerabilities. For most users this triggers automatic safety checks that block or pause tasks deemed risky, which may cause slowdowns even during unrelated work.
In short, if you belong to a Trusted Access or Daybreak organisation and need an agent that can autonomously operate software across contexts, Astra offers a tangible step forward—provided you accept the hosted‑only model, the current coding limits, the cost structure, and the strict cybersecurity safeguards. For teams outside those programmes, the practical path is to wait for broader access or to use alternative models that can be self‑hosted or fine‑tuned.
#AI #GPT6Astra #OpenAI #ComputerUse #DevTools #Cybersecurity