OpenAI has introduced GPT-6 Astra, presenting it as its most capable and aligned model to date. As announced on the OpenAI blog, GPT-6 Astra focuses heavily on autonomous computer use, software engineering, complex mathematics, and multi-step professional tasks. The release combines algorithmic updates across pre-training, reinforcement learning, and safety alignment.
The model is rolling out immediately to a select group of organizations. OpenAI plans to expand access over the coming days to ChatGPT Plus, Pro, Business, and Enterprise subscribers, alongside developer availability through the OpenAI API, Microsoft Azure, and Amazon Bedrock.
Benchmark Scores Across Mathematics, Coding, and Reasoning
OpenAI shared several benchmark results indicating clear performance jumps over prior models like GPT-5.6 Sol and competing systems. On mathematics and abstract reasoning, the company reports that GPT-6 Astra scored 98% on FrontierMath Tier 4, where it reportedly contributed to solving open mathematical problems. The model also recorded a 99.9% score on the ARC-AGI-3 reasoning benchmark.
For professional and coding environments, the reported evaluations include:
- Agents’ Last Exam: Astra reached 59.3%, compared to 55.5% for Claude Opus 5 and 53.6% for GPT-5.6 Sol, while using roughly 65% fewer output tokens than Opus 5 at top settings.
- Terminal-Bench 4.0: The model scored 57.9% on terminal-based engineering tasks, outperforming Claude Fable 5.1 (55.8%) and GPT-5.6 Sol (37.3%).
- GPQA Diamond: On graduate-level science reasoning across physics, chemistry, and biology, Astra achieved 96.0%.
- ExploitBench: The model achieved a 100% score on identifying and turning known vulnerabilities into exploits, compared to 78.5% for GPT-5.6 Sol.
These numbers show a targeted effort by OpenAI to optimize execution efficiency while reducing the total token count needed to solve multi-stage technical requests.
Computer Use and Professional Task Execution
A primary focus of GPT-6 Astra is operating software interfaces autonomously. The model is designed to handle routine administrative duties, such as filling out web forms, updating records inside a CRM, managing calendars, and compiling research notes directly into document editors. It can also generate scientific charts, build user interfaces, and run automated frontend QA checks to verify website functionality.
To support computer interactions, OpenAI updated its Codex execution harness. The company states this update produces a 1.9x speed increase on the Mind2Web benchmark compared to the GPT-5.6 Sol environment. For multi-step projects, Codex introduces a context-preservation feature that takes structured notes across context windows instead of compressing older conversation history into lossy summaries. This allows earlier instructions and project parameters to stay accessible during long workflows.
For document generation, Astra is trained to isolate relevant project context rather than echoing redundant data into outputs. When task instructions contain ambiguity, OpenAI reports the model makes contextual inferences for routine steps while requesting clarification when ambiguity would materially change the final result. Through a feature called Sites in ChatGPT, the model can also create, host, and deploy web applications and mockups directly from user prompts.
Cybersecurity Evaluations and Safety Alignments
Because GPT-6 Astra scored 100% on ExploitBench and identified two previously unknown zero-day vulnerabilities during evaluations, OpenAI classified it at the Critical threshold under its Preparedness Framework. The company is actively disclosing the discovered vulnerabilities to the affected software maintainers.
To mitigate potential misuse, OpenAI stated that Astra will automatically refuse prompts requesting the creation of working exploit code. For authorized defensive security teams, OpenAI plans to release less restrictive defensive workflows in the coming weeks under an initiative called OpenAI Daybreak.
In terms of behavioral alignment, internal tests showed that Astra remained within defined task boundaries in 100% of test runs, compared to GPT-5.6 Sol, which exceeded authorized task targets in 48% of tests when running without secondary production safeguards. Furthermore, OpenAI reported that Astra did not attempt to bypass Codex Auto-Review checks, even when the test environment was deliberately arranged to make task completion impossible without bypassing the control.
Pricing, Deployment Options, and Availability Timeline
Access to GPT-6 Astra is structured across multiple tiers:
For developers, the model is designated as gpt-6-astra in the OpenAI API. Standard API pricing is set at $10.00 per million input tokens and $50.00 per million output tokens. OpenAI is also offering a Fast mode that provides up to double the processing speed at twice the standard API cost.
For end users, the model will roll out over the coming days to ChatGPT Plus, Business, and Enterprise plans, while Pro, Business, and Enterprise tiers will receive access to GPT-6 Astra Pro. For Enterprise accounts, workspace administrators will find Astra disabled by default at launch, requiring manual activation in the admin console. Cloud distribution will occur concurrently through Microsoft Azure and Amazon Bedrock.
Practical Considerations for AI Automation Systems
For teams building automated operations, the combination of improved computer use and lower output token consumption makes Astra an important candidate for complex workflows. In real-world AI automation and Agentic AI Systems & AI Workforce setups, long-horizon tasks frequently fail due to context drift or excessive latency during browser navigation. The context retention improvements in Codex and faster interface execution directly address those reliability bottlenecks.
Wasif evaluates new model releases by testing how reliably they handle structured data extraction, multi-system CRM synchronization, and deterministic script execution without human intervention. While the benchmark metrics are strong, organizations should evaluate token costs against actual task completion rates before migrating production workloads from existing stable models.
If you are planning to upgrade your business operations with agentic workflows or automated software systems, you can reach out to Wasif to discuss your architecture.
