GPT-6 Astra: Features, Benchmarks, and Practical Guidance

The world’s best computer-use model, with caveats

Astra marks a new frontier in how quickly and accurately a model can perform computer-use tasks. It can autonomously handle tedious duties such as filling online forms, updating CRM records, organizing calendars, conducting online research, and drafting summaries for emails or documents. In professional contexts, Astra’s computer-use capabilities extend to analyzing scientific data, generating plots, building websites, and running frontend QA checks to verify site functionality. OpenAI reports substantial performance gains, including faster task completion when Codex is updated alongside Astra, translating into meaningful time savings for knowledge workers.

However, it is crucial to understand how benchmark scores translate to real-world usage. The ARC-AGI-3 score, cited at 99.9%, is achieved when Astra runs with a provider adapter harness that maintains state across reasoning steps. In standard stateless API calls, independent runs place ARC-AGI-3 in a much lower range (roughly 17% to 63% depending on the reasoning tier). In other words, the headline 99.9% reflects a highly harnessed configuration that does not generalize to typical stateless API usage.

A step change in professional work

Alongside computer-use improvements, Astra is designed to tackle multistep professional workflows, producing polished documents, spreadsheets, and presentations that align with templates and business styles. It can create slide decks that adhere to a given format and tone, assemble coherent narratives, and manage context so that outputs stay on target as requirements evolve. Astra’s ability to preserve relevant context across tasks and outputs helps maintain continuity in longer projects, reducing the need to reassert background details with every new instruction.

Coding, tooling, and context preservation

In software engineering contexts, Astra represents a notable leap. It performs well on complex terminal-based tasks, software engineering, and system configuration benchmarks, and it benefits from a Codex feature that preserves and retrieves context as the context window fills. Specifically, Astra can keep notes across context windows, allowing it to locate prior requirements or test results even if those items weren’t captured in notes. This experimental capability can be enabled via the Codex config.toml and is expected to become the default for Astra in the coming weeks.

In practical terms, coders can rely on Astra to stay oriented as a task evolves, asking focused questions only when necessary and using stored context to guide decisions. The combination of improved context handling and faster task completion contributes to more efficient coding workflows and better adherence to project templates.

Benchmark performance and important caveats

Astra reports top-line scores across several major benchmarks, including FrontierMath Tier 4, ARC-AGI-3, ExploitBench, and OSWorld 2.0. The FrontierMath Tier 4 (v2) score is listed as 97.6%, and ARC-AGI-3 is listed as 99.9% when using the provider adapter harness. ExploitBench is reported at 100%, and ExploitGym at 42.4%. On OSWorld 2.0, Astra achieves 72.6% in a 40-minute-per-task regime, signaling faster real-world task completion relative to Sol in similar conditions. It’s important to interpret these numbers with the harness caveat in mind: the 99.9% ARC-AGI-3 score and many other “best-in-class” marks rely on a stateful setup that goes beyond typical stateless API usage and can be associated with higher cost and more specialized configurations.

OpenAI has described FrontierMath Tier 4 saturation and other benchmarks as reaching a ceiling, meaning that the single-metric dominance may not straightforwardly predict future progress. In particular, the FrontierMath result is characterized as saturation, which signals that the test may be too easy to track further progress with, and that a harder benchmark would be required for meaningful longitudinal comparisons. A similar caveat applies to ARC-AGI-3 in stateless deployments, where results can differ substantially from adapter-harness runs.

Below is a compact view of Astra’s standings against a leading contemporary competitor, Claude Fable 5.1, and a broader cross-check against several other models. It highlights where Astra excels and where competitors may outperform Astra on specific benchmarks:

<th Claude Fable 5.1

These results reveal a nuanced picture: Astra leads on several domains—computer use, math-heavy reasoning, and cybersecurity—but in some benchmarks simulating broad knowledge or general coding, Claude’s Fable 5.1 and related models match or outpace Astra. This is why readers should view Astra as a strong performer with standout strengths in particular tasks, rather than a universal best across every index.

Safety, alignment, and responsible deployment

Astra marks a notable leap in cybersecurity capabilities, designated as meeting the Critical threshold under OpenAI’s Preparedness Framework. This enables defensive workflows—such as secure code review and patching—while gating more advanced offensive cyber tasks behind Daybreak access. OpenAI emphasizes that these capabilities come with stronger safeguards to minimize misuse and misalignment, including misalignment monitoring and automated stopping heuristics for potentially unsafe activity.

Given the heightened cybersecurity capabilities, Astra’s safeguards can pause, slow, or stop legitimate work. In conversational surfaces like ChatGPT or Codex, you may be asked to review an action before continuing; in the API, the task can stop outright. OpenAI notes that these checks are essential for safety but are being refined to reduce unnecessary interruptions while expanding frontier capabilities through programs like Daybreak.

Beyond automation safeguards, Astra’s alignment program relies on layered protections, including system-level safety classifiers, robust jailbreak defenses, and monitoring that can stop unauthorized activity. The model’s ability to respect safety restrictions appears stronger than GPT-5.6 Sol in many scenarios, though monitorability of written reasoning remains an area of active research. A recent evaluation framework indicates Astra is three times less likely than GPT-5.6 Sol to make inaccurate representations about capabilities, even as some concerns about chain-of-thought visibility persist. These safety and alignment considerations are central to OpenAI’s deployment strategy and influence how broadly Astra is made accessible at launch.

Availability, access, and pricing

Astra rollout begins with a limited set of organizations, followed by broader availability to ChatGPT Plus, Pro, Business, and Enterprise users, as well as through the OpenAI API, Microsoft Azure, and AWS Bedrock. Access is off by default for Enterprise workspaces and can be enabled by administrators. Astra Pro variants are available on Pro, Business, and Enterprise plans.

For developers, Astra is accessible in the API as gpt-6-astra, and pricing is standard API pricing: $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. A fast mode exists, delivering up to approximately 2x the speed at up to 2x the standard price. Astra usage counts against existing subscription quotas, with credits available for higher usage. Zero Data Retention is available for eligible API customers, and Private Safety Processing tests are being explored to preserve privacy while maintaining safety monitoring.

Additionally, Astra supports access through Microsoft Azure and Amazon Bedrock, and free-tier access is not part of the current rollout. This aligns with practical considerations for developers evaluating deployment options and cost profiles.

Codex context preservation: how it works in practice

One notable capability is Astra’s approach to context preservation in Codex, which keeps notes across context windows instead of packing everything into a single summary. This allows earlier instructions, requirements, and test results to remain discoverable as a project grows. The feature is described as experimental and is configured via the Codex config.toml; OpenAI plans to make it the default for Astra in coming weeks. Practically, this means developers can work through longer debugging sessions or large refactors with more reliable access to prior decisions and verifications, reducing the loss of historical context.

Recurrent depth: a technical nuance with safety implications

Astra uses a novel reasoning technique described in some sources as “recurrent depth,” which can obscure portions of the model’s chain-of-thought. The efficiency gains are notable, but this approach has safety and monitorability trade-offs. Independent reviewers have highlighted concerns about how easily such techniques can affect the ability to monitor the model’s reasoning. OpenAI and safety researchers consider this a priority area for ongoing study and refinement as frontier capabilities expand.

Where Astra stands against Claude Fable 5.1

Astra demonstrates clear strengths in several domains—computer use, mathematics, and cybersecurity—where it often leads. However, in some benchmarks, Claude’s Fable 5.1 shows stronger performance. For example, on the Artificial Analysis Intelligence Index v4.1.1, Claude Fable 5.1 scores higher (65.7) than Astra (61.2). Humanity’s Last Exam with tools also favors Claude Fable 5.1 (65.0%) over Astra (57.2%). On the other hand, Astra excels on OSWorld 2.0 computer-use tasks and cybersecurity-focused benchmarks, and it achieves top marks on ExploitBench. Practitioners should weigh task-specific strengths when choosing between Astra and Claude Fable 5.1, recognizing that neither model dominates across all benchmarks.

To help readers compare directly, a concise snapshot shows Astra’s edge in computer use and cybersecurity, while Claude Fable 5.1 leads on certain general knowledge and coding indexes. Consider the implications for your own tasks: if your work prioritizes robust, edge-of-automation computer-use performance and defensive cybersecurity workflows, Astra may offer a stronger match; if your priority is broad general knowledge or coding across diverse tasks, Claude Fable 5.1 may outperform Astra in some areas.

Access notes for readers and practitioners

One practical detail: free ChatGPT users do not have a direct access path to Astra at launch. This aligns with prior rollout patterns where frontier capabilities are gated behind paid plans or partner programs. If you are evaluating Astra, plan on a paid path and consider Daybreak access for early defensive use. Astra’s telemetry and safeguards are designed to balance safety with broad utility, but expect interruptions or pauses in some workflows as safeguards are refined.

Frequently asked questions

How do you get access to ChatGPT Astra? GPT-6 Astra is available on the ChatGPT Plus, Pro, Business, and Enterprise plans, as well as through the OpenAI API under the name gpt-6-astra, on Microsoft Azure, and on AWS Bedrock. Enterprise workspaces may have access disabled by default and must be enabled by an administrator. Pro, Business, and Enterprise plans include the GPT-6 Astra Pro variant.

How much does GPT-6 Astra cost in the API? OpenAI’s launch pricing is $10 per million input tokens and $50 per million output tokens, with separate rates for cache reads and writes. A fast mode exists, delivering up to 2x the speed for up to 2x the price. Current figures are shown on the OpenAI API pricing page.

Does Astra beat Claude on every test? No. On the Artificial Analysis Intelligence Index v4.1.1, Claude Fable 5.1 scores 65.7 versus 61.2 for Astra. For Humanity’s Last Exam with tools, Claude Fable 5.1 scores 65.0% versus Astra’s 57.2%. Astra does lead on computer use, mathematics, and cybersecurity, but Claude outruns Astra on several general reasoning and coding indices.

Why does my task stop halfway through? OpenAI runs production monitors that halt tasks when they violate safety or alignment constraints. In ChatGPT and Codex, you are asked to review the action before continuing; in the API, the task stops outright. This safeguard exists to prevent dangerous behavior, but OpenAI continues to refine it to minimize unnecessary interruptions for legitimate work.

Can Astra help with defensive cybersecurity? Yes, Astra can assist with secure code review and patching, which OpenAI explicitly permits in public versions. It does not support offensive work like creating proofs of concept exploits beyond the Daybreak program, where more advanced capabilities are gated.

Looking forward

As AI models take on more consequential work, the ability to align and reliably control these systems remains critical. OpenAI emphasizes ongoing testing, transparent evaluation, and careful deployment to ensure frontier capabilities remain safe and beneficial. The community should expect continued iteration on safeguards, monitoring, and policy as Astra and successor models evolve.

Note: All benchmark and capability figures referenced reflect public disclosures from OpenAI and third-party evaluations and are subject to caveats about harnesses, data windows, and test conditions. Readers should interpret headlining scores in the context of practical deployment constraints and safety safeguards that can affect performance in real-world usage.