# GPT-6 Astra: OpenAI’s New Frontier in Agentic AI
OpenAI has introduced GPT-6 Astra, its latest large language model, with a distinct positioning that sets it apart from previous releases. Rather than framing Astra as a conversational assistant, the company has designed it to function as an autonomous computer-use system capable of navigating software environments the way a human user would.
## A System Built for Action, Not Just Conversation
Astra’s core philosophy centers on practical execution. The model is engineered to operate across browsers, spreadsheets, desktop applications, and terminal environments, completing multi-step workflows end-to-end instead of merely outlining the steps a user should take. This represents a significant shift in how AI assistants are conceptualized — from advisors that suggest to agents that act.
The model is a closed, hosted system with no released weights, meaning it cannot be self-hosted on personal infrastructure. As of launch, Astra is available exclusively to organizations participating in OpenAI’s Trusted Access and Daybreak programs.
## Context Handling: The Biggest Developer Change
One of the most impactful technical improvements in Astra is how it manages conversation context. Previous iterations of Codex relied on a compaction process, where earlier exchanges were summarized once the context window filled up. This approach had a critical flaw: it discarded the granular details that agents often need later, such as why a particular fix failed or which specific test was run at an earlier stage.
Astra addresses this through a notes-based system that preserves information across context windows and allows the model to search back into earlier messages and tool outputs. This feature is currently experimental, controlled by a setting in the configuration file, though it is expected to become the default behavior in the near future.
Additionally, Astra can pose questions to the user while simultaneously continuing to work on parts of a task that don’t require an answer. This feature eliminates a common bottleneck in agent systems where a single unresolved decision halts the entire workflow.
## Model Specifications at a Glance
Astra comes with a range of capabilities that distinguish it technically from its predecessors:
– **Context Window**: 1,050,000 tokens
– **Maximum Output**: 128,000 tokens
– **Knowledge Cutoff**: April 30, 2026
– **Input Modalities**: Text and images
– **Output Modalities**: Text only
– **Reasoning Effort Levels**: Standard, low, medium, high, xhigh, and max
– **Tool Support**: Computer use, hosted shell, patch application, skills, MCP, and tool search
– **Fine-Tuning**: Not currently supported
## How Does It Perform?
### Operating System and Browser Tasks
On the OSWorld V2-Offline benchmark, Astra achieved a score of 72.6%, compared to 65.7% for GPT-5.6 Sol. Beyond accuracy, the model also dramatically reduced the average time required to complete tasks, cutting it from roughly 75 minutes to approximately 40 minutes.
### Reasoning and Specialized Benchmarks
Astra’s performance on the ARC-AGI-3 benchmark is notable, scoring 99.9%. It is important to note that this result was achieved using a specific Responses API configuration that retains reasoning across turns and utilizes compaction for managing long contexts. OpenAI has previously demonstrated that these settings alone can substantially boost scores on this benchmark without any change to the underlying model. This means the score reflects the combination of the model and the agent system working together.
Other impressive results include 97.6% on FrontierMath Tier 4, 95.9% on BenchCAD Vision2Code, and 64.6% on Terminal-Bench Science.
### Cybersecurity Capabilities
Astra represents a milestone for OpenAI’s Preparedness Framework, being the first model to reach the Critical cybersecurity threshold. During testing, it successfully developed exploits for hardened browsers and operating systems, and identified two previously unknown vulnerabilities in the V8 JavaScript engine, which OpenAI is proactively disclosing to relevant maintainers.
Due to these capabilities, access to advanced cybersecurity features is restricted. Standard accounts will be blocked or slowed when attempting such tasks, and API developers will encounter a hard stop rather than a review pause for cybersecurity-related safety checks.
### Pricing Structure
Astra is priced at $10 per million input tokens and $50 per million output tokens, with cached input available at $1.00 per million. For requests exceeding 272,000 input tokens, billing applies at double the standard input rate and 1.5x the standard output rate for the entire request. Batch and Flex compute tiers run at half the standard rate, while Fast mode doubles the cost. Pro, Business, and Enterprise tier users also gain access to Astra Pro.
## Frequently Asked Questions
**Q: Can I run GPT-6 Astra on my own hardware?**
A: No. Astra is a closed, hosted model. Self-hosting is not possible, and the weights have not been released publicly.
**Q: Who can access Astra right now?**
A: Access is currently limited to organizations enrolled in OpenAI’s Trusted Access and Daybreak programs.
**Q: What makes Astra different from GPT-5.6 Sol?**
A: Astra is built specifically as a computer-use agent rather than a chat model. It features improved context handling with a notes system, can ask questions while continuing unrelated work, and has significantly higher reasoning effort tiers.
**Q: Why does Astra score so high on ARC-AGI-3?**
A: The 99.9% score was achieved using a Responses API harness that retains reasoning between turns and compacts long contexts. The score reflects the combined performance of the model and the agent system, not the model alone.
**Q: Is fine-tuning available for Astra?**
A: No. Fine-tuning is not supported for this model.
**Q: What are the cybersecurity restrictions?**
A: Standard access explicitly refuses advanced cybersecurity tasks, including exploit discovery. API users will encounter an immediate stop on cybersecurity safety checks rather than a pause for approval.
**Q: How much does Astra cost compared to other models?**
A: Astra lists at $10 per million input tokens and $50 per million output tokens, which is higher than previous models. However, cached input is available at a reduced rate of $1.00 per million tokens.
## Conclusion
GPT-6 Astra marks a pivotal moment in the evolution of AI agents. By shifting the focus from conversation to action, OpenAI has created a system that can genuinely interact with software environments in a way that mirrors human workflows. The improvements in context handling, the ability to parallelize question-asking with ongoing work, and the impressive benchmark results across diverse domains signal a model built for real-world utility rather than just impressive demos.
However, the cybersecurity threshold achievement raises important questions about access control and responsible deployment. The fact that standard users may encounter slowdowns, pauses, or outright blocks during unrelated work highlights the delicate balance between capability and safety that OpenAI is navigating.
As Astra rolls out to more organizations in the coming weeks, the true test will be how effectively it performs in real-world scenarios beyond the controlled benchmark environments where it has excelled so far.
Thank you for reading
