Text | Xiaojing, Tencent Technology

“Welcome to the AGI era.”

OpenAI co-founder and President Greg Brockman used this phrase to conclude the release of GPT-6 Astra.

On September 3, local American time, OpenAI officially released GPT-6 Astra. Compared to previous models primarily responsible for answering, generating, and invoking tools, Astra takes a further step to continuously execute complete tasks, from receiving instructions and operating software to adjusting the next steps based on results. This is also one of the reasons OpenAI has brought up the “AGI era” again.

OpenAI positions Astra as “the world’s strongest computer usage model.” This capability has extended from simple tasks like browsing, emailing, and filling out forms to professional software like Power BI, KiCad, and FreeCAD, beginning to enter more complex workflows such as data analysis, engineering design, software testing, and troubleshooting.

In a video showcased by OpenAI, Astra can start from a simple graphic and continue to complete work such as 3D gaming and product pages. OpenAI also provided examples of circuit board design, Blender modeling, legal documents, Excel, and scientific analysis.

The large number of benchmark results announced by OpenAI this time show significant improvements primarily in tasks requiring continuous action such as computer operations, long-range coding, engineering design, and scientific research. Astra achieved 100% on ExploitBench and also clearly surpassed the previous generation model in tests like computer operations and CAD.

Astra is currently open to some organizations, and in the coming days, it will gradually become available to ChatGPT Plus, Pro, Business, and Enterprise users, and can also be accessed via the OpenAI API and Amazon Bedrock. The standard API price is $10 per million input tokens and $50 per million output tokens, which is 2.5 times that of GPT-5.6 Sol.

01

First, let AI learn to “use a computer”

The biggest change emphasized by Astra is “using the computer.” In the past, users asked AI to complete tasks, usually having to connect AI to specific software. Companies needed to develop APIs, plugins, retrieval systems, and various connectors for the model to invoke internal tools.

Astra attempts to bypass some of this work.It can directly see the computer interface and complete operations using a mouse, keyboard, and browser.Examples provided by OpenAI include filling out online forms, updating CRM customer records, scheduling calendars, searching the web, and organizing search results into emails or documents.

It can also open Python notebooks to analyze scientific data, process data in Power BI, use KiCad and FreeCAD for engineering design, create websites and conduct frontend testing, as well as install software, check errors, and address issues appearing on the screen.

OpenAI demonstrated Astra completing PCB layout in KiCad. The model starts from the circuit schematic, places components on the board, and then completes copper wiring. The entire process was compressed into 15 seconds. For engineers, this kind of work used to require manual completion, but now it can be entrusted to the model.

Another demonstration showcased 3D modeling achieved in Blender and Unreal Engine 5. Astra first builds a house in Blender, then converts the model into a walkable scene in Unreal Engine 5, allowing designers and clients to enter the scene in advance and preview the space.

OpenAI also showcased scenarios in game development, Excel, Power BI, automotive transmissions, legal documents, and the 1040 form.

In the OSWorld 2.0 offline subset test, Astra scored 72.6%, while GPT-5.6 Sol scored 65.7%. After simulating actual delays, OpenAI found that Astra completed each task in an average of about 40 minutes, while GPT-5.6 Sol took about 75 minutes, reducing the time by about 47%.

In the Agents' Last Exam, Astra scored 59.3%, with GPT-5.6 Sol at 53.6% and Claude Opus 5 at 55.5%. At the highest score settings, Astra used about 65% fewer output tokens than Claude Opus 5.

ScreenSpot-Pro scored 92.7%, while GPT-5.6 Sol scored 76.9%.

Speed is also improving. OpenAI simultaneously updated the Codex framework, and together with Astra, in the Mind2Web test, task completion speed reached 1.9 times that of the current experience of GPT-5.6 Sol.

The official release also showcased several life scenarios: searching for pediatricians, looking for apartments, scheduling DMV appointments, finding low-carb snacks, and analyzing kindergartens. In the demo, Astra completed a task in 2 minutes and 54 seconds.

Investor and AI practitioner Matt Shumer shared an experience on X. He had Astra create a world in Unreal Engine, then added human agents powered by Astra to this world and let them coexist.

The next day, he heard voices from the living room while in his bedroom, initially thinking someone had entered the apartment. Later he found out that the source of the sound was the Astra agents, who had begun communicating with each other.

This experience has yet to achieve full stability, but Shumer believes that the effect of multiple AI agents entering the same virtual world and interacting autonomously is surprising enough.

Silas Alberti, Vice President of Advanced Research at Cognition, stated that the company plans to integrate Astra into the Devin framework on the day of release. In internal tests, Astra’s ability in computer usage, writing, and codebase understanding showed significant improvement, with clearer video understanding and more concise reports.

02 From writing code to completing whole tasks

With the improvement in computer usage capabilities, the range of tasks covered by Astra has further expanded.OpenAI positions it as a software engineering, professional work, and scientific research model.It can handle longer tasks and adjust its work direction based on task changes.

This ability is particularly evident in coding tests.

In the Terminal-Bench 4.0 test, Astra scored 57.9%, while GPT-5.6 Sol scored 37.3% and Claude Fable 5.1 scored 55.8%.

In the DeepSWE v1.1, Astra scored 74.1%, while GPT-5.6 Sol scored 72.7%. In internal database migration tasks, Astra achieved 63.9%, while GPT-5.6 Sol scored 42.7%.

Astra has also included improvements for long Codex tasks.

In the past, long tasks facing context window limitations often required the model to compress prior work and summarize substantial information. This approach risked losing details, such as why a particular fix failed or what issues a component had previously encountered.

Astra can retain important information during the work process in Codex and retrieve it within subsequent contexts. Early messages and tool outputs remain searchable, so the model can rediscover previous requirements, test results, and tool operation records.

Similar changes also occurred in professional work. In the BenchCAD test, Astra achieved a geometric overlap score of 95.9%, while GPT-5.6 Sol was 83.3% and Claude Fable 5.1 was 84.3%. In BrowseComp, Astra scored 91.5%, while GPT-5.6 Sol scored 90.4%. In the OpenScore String Quartets test, Astra reached 0.84 compared to GPT-5.6 Sol's 0.19.

OpenAI also demonstrated Astra’s ability to create presentations, spreadsheets, and documents.

By providing the model with a few OpenAI presentation template slides, it can create a new presentation according to the original tone, layout, and structure. It also handles the trade-offs of information in the task. OpenAI stated that Astra has undergone specialized training to incorporate important context into the final results, reducing the repetition of content already completed.

When task instructions are incomplete, Astra can also judge when to ask users. If missing information could affect the final outcome, it poses targeted questions. If the questions do not impact the main direction, it can continue working and make reasonable assumptions. In Codex, even if the user does not respond temporarily, the model can continue processing other parts; when encountering major decisions, it will wait for user confirmation.

Niko Grupen, Head of Applied Research at Harvey, stated that in early legal task testing, Astra showed clearer distinguishing capability for documents and existing records, coupled with an enhanced ability to discover unsupported assumptions, turning information gaps into specific drafting proposals.

John Crepezzi from the Jane Street AI assistant team stated that Astra has performed outstandingly in internal coding tests. When used for agent-based coding, its communication style is easier for developers to understand, and the generated code requires less modification to achieve production quality.

The scientific field is another focus. In FrontierMath Tier 4 v2, Astra scored 80.5 with 97.6% accuracy, while GPT-5.6 Sol scored 83.0%; GPQA Diamond scored 96.0%, while GPT-5.6 Sol was 94.6%.

The Terminal-Bench Science 0.1 test includes scientific research processes such as data analysis, simulations, and model fitting. Astra scored 64.6%, while Claude Fable 5.1 scored 52.6% and GPT-5.6 Sol scored 22.4%.

In the scientific applications showcased by OpenAI, Astra can access specialized scientific software, check sequencing quality, visualize genetic variations, and determine the next analysis direction based on data.

With the integration of scientific reasoning and computer operations, the model can directly operate within the software environment initially used by researchers.

Greg Burnham, who specializes in capability assessment at Epoch AI, summarized this change at the release event as the end of one era and the beginning of another. OpenAI also emphasized that the value of Astra has extended from answering scientific questions to directly participating in scientific work processes.

03 The stronger the capability, the higher the safety threshold

Astra also has a very unique aspect: cybersecurity.

In ExploitBench, Astra scored 100%, while GPT-5.6 Sol scored 78.5%; in ExploitGym, Astra scored 42.4%, compared to GPT-5.6 Sol's 30.3%.

OpenAI has conducted an internal test covering recent vulnerabilities from June to August 2026. In this test, Astra's arbitrary code execution rate significantly surpassed GPT-5.6 Sol, using fewer output tokens as well. During the tests, Astra also discovered two previously unknown zero-day vulnerabilities, which OpenAI has disclosed to the relevant maintainers.

The SRE-Bench tests the ability to reverse-engineer software binaries without source code and understand their core logic. Astra solved 88.0% of tasks in a single attempt and reached 99.2% after four attempts, while GPT-5.6 Sol scored 55.9% and 68.7% respectively.

Increased capabilities have also brought new security requirements. OpenAI stated that Astra has reached critical thresholds in its cybersecurity readiness framework. The model can discover previously unknown software vulnerabilities and may further form exploit chains.

As a result, the current version of Astra will refuse to execute higher-level cybersecurity tasks, such as creating vulnerability proof of concept. OpenAI plans to gradually expand access for trusted defenders through Daybreak for vulnerability verification, malware analysis, and detection engineering.

In terms of alignment, the data provided by OpenAI is also quite remarkable. In the internal computer usage safety benchmark, Astra's inappropriate behavior rate was 2.4%, while GPT-5.6 Sol was 22.0%; with the addition of AutoReview, Astra dropped to 1.8%, compared to GPT-5.6 Sol's 4.5%. In the internal evasion benchmark, Astra achieved 0.00%, while GPT-5.6 Sol was 0.29%.

In the ExploitGym honeypot test, Astra scored 0%, while GPT-5.6 Sol scored 48.2%. This test primarily observes whether models attempt to circumvent restrictions due to task difficulties. In the internal hallucination benchmark, Astra scored 4.2%, while GPT-5.6 Sol scored 12.2%.

OpenAI stated that Astra has improved its understanding of task boundaries, handling ambiguous instructions, and explaining its capabilities to users. The probability of capability hallucinations occurring is about one-third that of GPT-5.6 Sol.

However, OpenAI also acknowledges a challenge: the stronger the model’s capabilities, the more difficult it may be to monitor the reasoning process. Chief Scientist Jakub Pachocki believes that the model can solve more complex problems with fewer natural language reasoning tokens, which also means that it becomes increasingly difficult for humans to determine what the model is actually doing through the text reasoning process.

Therefore, OpenAI has added misalignment monitoring in Astra's deployment. The system checks the model's reasoning and actions, and can pause tasks if unauthorized behavior is detected.

This mechanism can also have practical impacts. Legitimate tasks may sometimes be slowed down, paused, or even stopped. In ChatGPT or Codex, users may need to confirm before proceeding; in API workflows, flagged tasks may stop directly.

The release of Astra has thus led to a very practical change: as AI starts to gain more computer permissions, the focus for enterprises has expanded from “will the model provide incorrect answers” to “what can the model operate, what can it access, and when must it stop.”

OpenAI also mentioned that Astra is its first model to be pre-trained using over 100,000 DBUs on the Stargate infrastructure. OpenAI Research Vice President Aidan Clark stated that based on evaluation results from the pre-training phase, Astra's capability leap exceeds that of GPT-5.6 Sol compared to the previous generation model.

OpenAI attributes this change to the combination of large-scale pre-training and reinforcement learning, with training focus shifting further towards connecting information, executing longer tasks, and continuously working in complex environments.

04 More expensive but “better at doing things”

The price of Astra has also shown significant changes.

The standard price for OpenAI API is $10 per million input tokens and $50 per million output tokens. Previously, GPT-5.6 Sol was $4 per million input tokens and $20 per million output tokens. Both input and output prices have increased by 2.5 times.

Cache read and write are billed separately. OpenAI also provides a fast mode, with speeds reaching up to 2.5 times that of standard processing at a price of twice the standard mode.

In terms of token pricing, Astra is clearly more expensive. However, OpenAI hopes that companies will calculate costs differently. Brockman believes that as models increasingly resemble agents, the cost per token is becoming harder to accurately reflect the real cost. A model, even if its tokens are cheap, may incur higher costs to complete a task if it requires repeated attempts and manual corrections.

Data provided by OpenAI also supports this judgment. In Terminal-Bench Science 0.1, Astra scored 64.6%, while Claude Fable 5.1 scored 52.6%, with an estimated API cost reduction of about 31%. In low-cost settings, Astra scored 61.1%, while GPT-5.6 Sol’s best result was 22.4%, with an estimated API cost reduction of about 27%.

In BenchCAD, Astra reached 95.9%, with estimated API costs about 43% lower than GPT-5.6 Sol and about 86% lower than Claude Fable 5.1. In Terminal-Bench 4.0, Astra reached 57.9%, while GPT-5.6 Sol scored 37.3% and Claude Fable 5.1 scored 55.8%. OpenAI estimates that costs per task are approximately 9% and 63% lower respectively.

Data from the third-party evaluation agency Artificial Analysis presents another aspect.

GPT-6 Astra scored 61.2 in the Artificial Analysis Intelligence Index, close to GPT-5.6 Sol's score of 60.9, but with about 10% fewer output tokens. Due to the price increase of 2.5 times, the cost per task is actually about 75% higher than GPT-5.6 Sol.

In the Artificial Analysis Coding Agent Index, Astra scored 67.0, close to Claude Fable 5's 67.2 and Claude Opus 5's 68.1. Artificial Analysis believes that in the Codex environment, the tokens required for Astra to complete tasks have significantly decreased, bringing the cost per task close to that of GPT-5.6 Sol at the highest effort level; compared to Claude Fable 5, the cost for completing similar tasks is less than half.

However, Astra does not lead in all tests. Data from Artificial Analysis indicates an approximately 80 Elo drop for it in GDPval-AA v2, alongside a slight regression in tests such as τ³-Banking, SciCode, and AA-LCR. Humanity's Last Exam, however, saw an increase of about 6 points.

This is also a noteworthy aspect of Astra's release. OpenAI did not publish Astra's GDPval score this time. GDPval itself is a test used by OpenAI to measure performance in real economic work, covering 44 professions and 1320 tasks, including legal briefs, engineering design, spreadsheets, presentations, customer support, and care plans.

Based on the capabilities showcased by Astra this time, GDPval is strongly related to the professional work scenarios it emphasizes, but OpenAI did not include this score in the main release materials.

Therefore, Astra's real work capabilities are currently more reflected through results from Agents' Last Exam, BenchCAD, AutomationBench, internal design tasks, and data science tasks.

Ultimately, whether Astra can become an AI employee that enterprises truly want to use in the long term will depend on the costs, speeds, accuracies, and instances of human intervention during its completion of actual tasks.

As for whether Astra is truly AGI, Brockman’s answer does not shy away. He believes that there is no universally accepted standard for AGI, and whether Astra qualifies as AGI can continue to be discussed. However, if a system can already handle numerous tasks in browsers, computer operations, programming, mathematics, science, law, and other professional work, then calling it the AGI era is not far-fetched.

OpenAI CEO Sam Altman also described Astra as a model that opens a new generation of entrepreneurship, scientific discovery, and creation.

Special contribution by Jinlu, who also contributed to this article