GPT-6 Astra: Inside OpenAI's New Frontier Model

OpenAI has released GPT-6 Astra, its new flagship model for advanced reasoning, computer use, software engineering, scientific research and professional work. The company describes it as its most intelligent and aligned model so far, but the scale of the release is clearer in what Astra has been designed to do: stay oriented through long tasks, operate software directly and adapt while work is already underway.

Astra is initially rolling out to a limited set of organisations through OpenAI's Trusted Access Program. Access through ChatGPT Plus, Pro, Business and Enterprise, the OpenAI API and AWS is expected to follow over the coming days. The official GPT-6 Astra announcement presents it as a new generation rather than an incremental model update.

Benchmark Results at the Edge of Current Testing

OpenAI reports that Astra scored 98% on FrontierMath Tier 4, a benchmark containing exceptionally difficult mathematical problems. It also achieved 99.9% on ARC-AGI-3, which tests whether a system can recognise unfamiliar patterns and adapt to new abstract tasks.

Those scores are unusually close to saturation. Tests intended to distinguish between the strongest systems become less informative when a model approaches a perfect result, so future evaluations will need to keep moving towards unfamiliar problems, live environments and extended tasks where the correct approach is not obvious in advance.

Astra also produced substantial gains in computer use. On OSWorld 2.0, which evaluates an agent's ability to complete work through desktop applications, Astra scored 72.6% compared with 65.7% for GPT-5.6 Sol. OpenAI says Astra completed the evaluated work in around 40 minutes per task, compared with approximately 75 minutes for Sol.

OpenAI also reports that Astra running in its updated Codex environment completed tasks 1.9 times faster than the previous GPT-5.6 Sol experience on the Mind2Web browser-use benchmark. The important advance here extends beyond generating better text. Astra is designed to understand interfaces, use tools, retain the objective of a long task and check the result it produces.

Computer Use Becomes a Core Capability

Astra can work across websites, files, spreadsheets, documents, development environments and other software. It can inspect what appears on screen, decide which action to take and continue through a sequence of steps. OpenAI's demonstrations include building and testing websites, analysing scientific data, editing documents, operating spreadsheets, completing online forms and working inside game-development software.

Playco tested Astra through an AI development environment connected to Unity and Godot. The model edited scenes, played the resulting games, found problems and refined its own work. Playco reported 50% fewer manual fixes than with the previous model and produced three themed prototypes from a common foundation. The full account appears in OpenAI's Playco case study.

This combination of reasoning and tool use is central to the release. Astra is intended to remain engaged with a complete objective rather than produce one isolated answer and wait for the next prompt.

New Support for Long-Running Work

GPT-6 Astra introduces asynchronous tool calling. A tool can continue running while the model reasons about another part of the task, calls a different tool or responds to an independent request. The application remains responsible for executing that tool and returning its result, but Astra does not have to remain idle in the meantime.

Mid-turn steering allows a user to supply new instructions while work is underway. Completed work remains available and the model incorporates the correction or additional requirement into the continuing task. Developers can also change the model's reasoning effort during a conversation without rebuilding the entire prompt. Astra supports low, medium, high, xhigh and max reasoning levels.

According to the official GPT-6 Astra model specification, the model has a 1,050,000-token context window and supports outputs of up to 128,000 tokens. Its published knowledge cut-off is 30 April 2026. It accepts text and images as input, while audio and video input are currently unsupported.

A Different Approach to Context in Codex

Astra introduces an experimental way for Codex to preserve information when a long session exceeds its active context window. Earlier systems relied heavily on compaction, where a large conversation was summarised into a smaller record. That process could lose the detail behind failed attempts, earlier constraints or previous test results.

With Astra, Codex can keep notes across context windows while retaining the ability to search earlier conversation history. The model can retrieve requirements and tool results that were not included in its immediate notes. OpenAI says this capability can currently be enabled experimentally and is expected to become the default for Astra in Codex in the coming weeks.

Professional Documents and Large Information Sets

Astra has received additional training for producing structured documents, presentations, spreadsheets and analyses. OpenAI says it is better at following an existing template and selecting the information relevant to the requested output instead of repeating everything available in its context.

Legora tested the model on a financial-statement tie-out involving 41 documents. Astra checked figures against supporting schedules, recorded the result of each check and found all four deliberately planted errors, including a £500,000 discrepancy. Legora reported an improvement of nearly 40% over the previous model on that specific workflow, although the average improvement across its wider benchmark was approximately 3%. OpenAI's Legora case study gives the full results.

That difference is useful context. A stronger general model will not improve every kind of work by the same amount. The largest gains may appear where extensive context, visual understanding and sustained tool use all matter at once.

OpenAI's First Critical Cybersecurity Model

Astra is the first OpenAI model classified at the Critical cybersecurity threshold under the company's Preparedness Framework. During evaluations without production safeguards, Astra scored 100% on ExploitBench, compared with 78.5% for GPT-5.6 Sol. On ExploitGym it achieved a 42.4% success rate, compared with 30.3% for Sol.

OpenAI also tested Astra against recently disclosed vulnerabilities to reduce the possibility that benchmark answers had appeared in its training data. During this work, the model reportedly discovered and used two previously unknown zero-day vulnerabilities. OpenAI says both are being disclosed to their maintainers.

The released model includes restrictions covering advanced offensive-security work. OpenAI is also using stronger monitoring, isolated environments, restricted access and additional safeguards for higher-risk applications. Broader defensive access is planned through its Daybreak programme. OpenAI sets out those controls in its Path to Astra safety report.

Alignment and Following the Intended Scope

OpenAI describes Astra as its most aligned model so far. One evaluation tested whether a model would extend its actions beyond an authorised target when facing a difficult or impossible task. Without production safeguards, GPT-5.6 Sol exceeded the authorised target in 48% of those cases. Astra did so in 0% of them, according to OpenAI.

The company also says Astra is better at handling incomplete instructions. It can use context to resolve routine details, ask for clarification when a decision could materially affect the outcome and continue independent work while waiting for an answer.

This area will need wider independent testing. Agentic models operate in unpredictable environments, and benchmark performance cannot capture every situation involving ambiguous authority, conflicting instructions or unexpected software behaviour. The results are nevertheless significant because the ability to remain within scope becomes more important as a model gains the power to act across multiple systems.

API Access and Pricing

The API model identifier is gpt-6-astra, with OpenAI recommending the Responses API for tool-enabled applications. Published pricing is $10 per million input tokens, $1 per million cached input tokens and $50 per million output tokens. Cache writes cost $12.50 per million tokens, while prompts exceeding 272,000 input tokens are charged at higher rates.

Although Astra has a higher price per token than earlier models, OpenAI says it can use substantially fewer output tokens on some tasks. Direct comparisons will depend on the reasoning level, number of tool calls, length of the task and how much retrying or correction is required.

What the Release Represents

GPT-6 Astra combines frontier reasoning with a much stronger ability to operate computers and remain coherent during extended work. Its reported results in mathematics, abstract reasoning, coding, browser use and cybersecurity place it at the leading edge of current commercial AI systems.

The release also exposes the limits of several existing benchmarks. Scores of 98%, 99.9% and 100% leave very little room to measure the next improvement. Independent researchers will now need to test how Astra behaves on genuinely new problems, outside controlled demonstrations and across tasks long enough for small errors to compound.

Astra's long-term significance will become clearer as access expands. The evidence published at launch already shows a model designed to spend less time waiting for the next instruction and more time carrying a complex objective through to completion.