GPT-6 Astra Launch: OpenAI Says AI Has Entered the “AGI Era”

OpenAI has launched GPT-6 Astra, its newest flagship AI model, and says it may represent the beginning of the “AGI era.” OpenAI President Greg Brockman said he personally believes the company has reached a point that could reasonably be described as artificial general intelligence, although he said people can decide for themselves whether Astra meets that definition.

Brockman further stated that AGI is not being considered as a mere technical or contractual milestone. He described it as a bigger idea related to OpenAI’s objective, and stated that the launch is merely the beginning of a much longer journey.

GPT-6 Astra Uses OpenAI’s Largest Training Run

Astra is the result of OpenAI’s most extensive training program to yet. Aidan Clark stated that the model was trained using over 100,000 GPUs at the company’s Stargate facility in Texas.

The new system also has a unique link with previous AI models. According to OpenAI, Astra is the company’s first model, with earlier models playing an important part in training supervision.

Access is being introduced gradually. Enterprise clients that are already using the Daybreak program will have priority, with Plus, Pro, Business, and Enterprise users receiving Astra in the coming days. The model will also be made available through the API and AWS.

Customers with Pro, Business, and Enterprise plans will have access to Astra Pro. Select API customers will be allowed to use the model with zero data retention.

GPT-6 Astra Shows Strong Benchmark Results

Astra is charged at $10 per million input tokens and $50 per million output tokens when utilized via the API. This is far greater than some competing models, but OpenAI claims that the cost of completing the complete assignment is more relevant than the price of each token.

According to the firm, Astra utilizes less tokens in various evaluations and partner tests. However, there is insufficient launch data to assess whether these savings will offset the higher API charge.

The model performed well on a variety of demanding tests. Astra received 74.1% on DeepSWE v1.1, an agentic coding benchmark with 113 jobs. GPT-5.6 Sol achieved a score of 70.8%.

Its most significant benefits appear in areas other than coding. OpenAI achieved a 98.6% score on ARC-AGI-3 and 97.6% on FrontierMath Tier 4. Astra scored 95.9% on BenchCAD’s Vision2Code test, compared to 84.3% for Fable 5.1 and 83.3% for Sol.

Astra also scored 64.6% on Terminal-Bench Science, which evaluates AI agents on 70 command-line research activities across five scientific fields.

Codex Gets New Capabilities With Astra

Codex may contain some of the most practical updates for developers. Astra may maintain notes across context windows and search older messages and tool results, rather than relying solely on compaction when a coding job grows too huge.

This could assist an agent remember facts like why an earlier repair failed, which tests were passed, or what requirements were supplied at the beginning of a task.

Astra can also operate on independent pieces of a job while waiting for a response from the user. According to OpenAI, this prevents one unanswered question from bringing the entire job to a halt.

The model has been proved to function with software such as Excel, Blender, KiCad, and Power BI. Astra got 72.6% in the OSWorld V2-Offline benchmark, compared to GPT-5.6 Sol’s 65.7% score. According to OpenAI, the average time it takes to complete a task has decreased from 75 minutes to 40.

Safety Remains a Challenge for Astra

The launch also includes warnings about monitoring and safety procedures. OpenAI discovered that Astra’s written reasoning was more difficult to monitor in experiments intended to detect attempts to dodge scrutiny.

Jakub Pachocki, Chief Scientist at OpenAI, stated that increasing intelligence does not always result in better alignment. The corporation believes it may postpone further growth until it is confident in its capacity to supervise increasingly sophisticated systems.

Another area where tougher regulations are being implemented is cybersecurity. According to OpenAI, Astra crossed the Critical cybersecurity barrier during testing and was able to create attacks for hardened browsers and operating systems.

During testing against recent V8 issues, the model discovered two previously undiscovered vulnerabilities. OpenAI stated that it intends to notify the appropriate maintainers about the issues.

Some advanced cybersecurity features will not be available to all. Daybreak programs let vetted defenders to gain broader access, whereas normal users may have specific cybersecurity tasks refused, stopped, or blocked.

According to OpenAI, these constraints will be included in the launch as it continues to study how to safely allow access to a more powerful AI system.