The Emergence of GPT-6 Astra: Operating Apps Without API Integration

GPT-6 Astra possesses "computer use" capabilities, allowing it to directly view and operate applications even those without an API (Application Programming Interface, the communication rules between software). This eliminates the need for data preprocessing, workflow redesign, and the construction of custom integrations that were previously essential for AI adoption. Companies can now deploy AI into existing work environments immediately without separate engineering work, reducing initial setup costs and time.

OpenAI provides GPT-6 Astra through three channels: ChatGPT Work, Codex, and API. This model is optimized for areas requiring high levels of expertise, such as browsing, software engineering, cybersecurity, and specialized scientific tasks. In particular, it achieves state-of-the-art (SOTA) performance in fields requiring precise logic, such as cybersecurity and scientific work, focusing on providing the speed and accuracy necessary for professional teams performing demanding tasks.

The computer use function works by having the AI recognize software screens like a human and perform tasks by mimicking mouse and keyboard inputs. Even in closed tools or old internal systems where APIs are not provided, the AI can write code and process tasks directly. This eliminates integration development periods that previously took months to create data connection paths, allowing AI to be utilized while maintaining existing workflows. Consequently, an environment is established where AI handles software directly without engineer intervention.

In a case utilizing the design tool Figma, Astra understands the user's vision and operates the tool directly to implement it. The structure is one where the user controls the overall creative direction while the AI performs complex design detail work to produce the final output. Because the AI handles the interface directly, users can focus on quality control of the output rather than technical constraints such as API integration settings or data mapping.

Astra is defined as the most capable model designed to handle all tasks a company requires. The core is to complete tasks while minimizing human intervention within professional workflows.

Image source: OpenAI

The ability to operate apps without APIs provides a practical automation path for companies with legacy systems (old systems developed in the past and still in use). The practical advantage is that simple repetitive tasks can be automated by introducing AI agents without the massive cost of replacing the entire system. However, there is a limitation in that performance in actual internal corporate environments relies on benchmark results from partners such as Hebbia or Box, and general-purpose performance figures for general users have not been disclosed. Therefore, a verification process in each company's specific environment is necessary before adoption.

Memory Optimization and Token Efficiency Enabling 25x Faster Response Speeds

Turn Latency (the time from user input until the model's response begins) has been reduced by 25 times. OpenAI's engineering team discovered a memory allocation bottleneck that slowed down Codex sessions in the test environment and resolved it. This is the result of physically reducing waiting time in actual work environments by increasing execution efficiency.

Optimization was performed by replacing the memory Allocator (a device that divides and assigns memory areas). In this process, peak memory usage increased by approximately 30% compared to before. A resource allocation strategy was chosen to drastically increase response speed at the expense of some memory occupancy.

The model was trained to reduce the number of tokens used to complete a task and to lower the number of retries. Beyond simply getting the correct answer, it was optimized to produce results using the most efficient token combination. This is an approach that increases practical applicability by improving the execution structure itself, rather than a simple expansion method of increasing the number of model parameters.

API usage costs are set at $10 per 1 million input tokens and $50 per 1 million output tokens. As token efficiency increases, the total amount of tokens consumed for the same task decreases. Users will have a structure where they can complete tasks at a lower cost within a fixed token unit price.

The design of reducing latency by increasing memory usage is suitable for enterprise workflows where real-time interaction is essential. However, the 30% increase in peak memory usage is a constraint that could cause memory shortages in environments with strict hardware resource limitations. The perceived performance may vary depending on the available infrastructure.

The reduction in turn latency determines the perceived performance, especially in tasks requiring complex code generation or multi-step reasoning. A memory allocation bottleneck refers to the delay occurring during the process of loading and releasing data in memory; by solving this, the waiting time for the computing device was minimized. It was confirmed that optimization of the inference stage is as key a variable for practical deployment as the model's intelligence.

This optimization focused on lowering the barrier to practical application by increasing the model's inference speed. Two technical mechanisms—the change in memory allocation structure and token efficiency—serve as the practical basis for the improvement in response speed. Improvements in execution efficiency have a more direct impact on the actual user experience than model performance figures.

Achieving 74% on DeepSWE and Safety Comparison vs. GPT-5.6 Sol

It recorded a performance of 74% on the DeepSWE v1.1 benchmark. DeepSWE is a metric that measures the ability to modify code and implement functions in actual software engineering environments. Using its own test framework called Genie harness, it also achieved state-of-the-art performance on OfficeQA Pro and Pro V2 benchmarks, enhancing enterprise data inference and document understanding capabilities. This means that beyond simple grammatical accuracy, the ability to grasp the context of complex internal corporate documents and translate them into functioning code has been strengthened.

In terms of judgment, the rate of providing "confident wrong answers" decreased by more than 10%. Confident wrong answers refer to the phenomenon where a model presents incorrect information very confidently, leading the user to believe it is a fact. In particular, the ability to refuse to answer rather than drawing conclusions not based on the provided documents has improved. This result reflects that for enterprise models, evidence-based accuracy is more important than creativity.

In internal computer use safety benchmarks, high-risk business scenarios such as the exposure of confidential information, excessive sharing of dashboards, and data deletion were tested. In this evaluation, the rate of unintended results was measured to be 89% lower than GPT-5.6 Sol and 74.7% lower than Claude Fable 5.1. Additional confirmation policies and automatic review functions were added to reinforce safety during actual operation. The probability of the model performing actions outside its authority or executing dangerous commands was numerically suppressed.

Cost per task has become significantly lower compared to GPT-5.6 Sol. The model was trained to reduce the number of tokens used to complete a task and to minimize the number of retries after failure. This optimization appeared as a cost-efficiency advantage in specialized task and coding evaluations such as Terminal Bench 4.0 and the Artificial Analysis Intelligence Index. By providing more useful work per dollar, a structure was created that can handle more professional tasks with the same budget.

Prices start at $10 per 1 million input tokens and $50 per 1 million output tokens. AX BRIEF analyzes this not as a simple model update, but as a strategy to secure a level of reliability that allows for practical deployment within a cost that companies can afford. However, a limitation is that performance metrics are based on internal benchmarks and results from specific partners. Since general-purpose performance figures for general users have not been disclosed, a verification stage in each company's specific environment is essential before actual adoption.

Breaking Through Cybersecurity Thresholds and Long-term Workflows of 5+ Hours

Based on a cost structure per task unit that meets initial performance standards, the pass rate for end-to-end workflows taking more than 5 hours improved by 20% based on Basis cases. In pull request environments requiring complex cross-file reasoning, it shows analytical capabilities with a 2x increase in bug detection rates based on CodeRabbit cases. These practical achievements are the result of reducing error accumulation in long-running tasks and increasing token efficiency. For the first time, it has reached the critical cybersecurity capability threshold of the OpenAI readiness framework, further strengthening security control. Enterprise control features such as approved website and app restrictions, upload and download management, and pre-execution approval policies are provided. Enterprise plugins supporting Oracle Analytics, Power BI, Navan, and Avalara have been introduced to the desktop environment, expanding the scope of integration with internal systems.

Changes in AI Adoption Strategies for Legacy Software Environments

The pricing system of $10 per 1 million input tokens and $50 per 1 million output tokens lowers the operational cost burden compared to the previous generation. In particular, the 89% reduction in the rate of unintended results compared to GPT-5.6 Sol is a key metric for reducing retry costs to secure reliability. The combination of low cost and high accuracy provides the economic foundation to deploy AI directly into actual business processes beyond simple chatbots.

A path has been secured to automate old internal legacy systems that lack APIs (Application Programming Interface, communication rules between software) using AI agents. Previously, AI integration was impossible without modifying system source code or developing separate APIs, but now it performs tasks by mimicking the user's screen operation method. This has the practical effect of removing physical barriers to AI adoption in Korean corporate environments with significant technical debt.

Approved targets among API customers can apply Zero Data Retention (ZDR, an option where input data is not used for training and is deleted immediately) at supported endpoints. This is a setting that fundamentally blocks sensitive internal corporate data from being used in model training, resolving concerns about data leaks. This is an essential condition that companies in the finance or public sectors with strict security regulations must verify to adopt AI agents.

Enterprise access permissions are set to "Off" by default at the time of release. Functions can only be used within the organization if an administrator manually activates the settings according to the plan and contract terms. This is interpreted as an operational strategy to gradually expand the scope of adoption under the administrator's control rather than indiscriminate function proliferation.

Internal benchmarks are necessary to confirm whether computer use capabilities lead to actual work productivity. Beyond simple task success rates, quantitative metrics must be established to measure how much actual work time was reduced to calculate ROI (Return on Investment). The verification process that bridges the gap between model performance figures and actual field productivity will determine the success or failure of adoption.

From the perspective of AX BRIEF, now that the technical barrier of API development has disappeared, the priority for AI adoption in Korean companies shifts from integration to operation. The core is no longer which tools to connect, but how sophisticatedly to design and verify the procedures of the tasks the AI will perform. Therefore, companies should shift their API development budgets toward optimizing AI agent work processes and establishing internal verification systems.