OpenAI launches GPT-6 Astra with 10 trillion parameters

OpenAI announced GPT‑6 Astra on Tuesday, unveiling a new flagship model that can write, debug, and run code while interacting with external software in real time. The company said the system is now available to a limited set of enterprise partners and will roll out to a broader audience later in the year. Astra’s launch marks a shift from the conversational focus of GPT‑4 to a more tool‑centric AI that can manipulate digital environments.

The move follows years of incremental upgrades, from GPT‑3’s 175 billion‑parameter model to GPT‑4’s multimodal 100‑billion‑parameter version and the rumored GPT‑5, which never saw a public release. Engineers at OpenAI claim that Astra’s architecture uses a new transformer design that scales to 10 trillion parameters—about 25 % more than GPT‑5’s theoretical peak. The company argues that this jump, coupled with tighter safety controls, brings the system closer to artificial general intelligence, the ability to perform any intellectual task a human can. Industry observers note that the claim hinges on how well the model can adapt to novel problems without human guidance.

In early tests, Astra solved 90 % of the coding challenges that stumped GPT‑4, including complex algorithms and real‑world debugging scenarios. It can execute code in a sandboxed environment, call APIs, and revise its own output based on new data. The model also passes several standard benchmarks for logical reasoning and natural language understanding, scoring 15 points higher than GPT‑4 on the LAMBADA test. While developers praise the leap, safety researchers warn that the increased autonomy could amplify hallucination risks and make alignment more difficult.

OpenAI plans a phased rollout, starting with a closed beta for large enterprises that need reliable code generation. The company will publish safety documentation and open a public forum for independent researchers to audit the model. Regulators are already discussing oversight frameworks for advanced AI systems that can modify software. If Astra proves stable, OpenAI intends to release a public beta in Q4 2026, potentially setting a new standard for AI tools and accelerating the broader move toward AGI.