OpenAI Updates Model Spec and Prepares for Multi-Agent "Astra" Architecture

OpenAI published an update to its Model Spec on August 18, 2026, aiming to deepen public conversation about the intended behavior for models powering its API and consumer products. [1] In this specific update, the company added clarity around principles on appropriate relational interactions for teens and clarified how AI assistants should handle false or unsupported premises. [3] The revision also removed outdated guidance for pre-reasoning models and introduced a new section titled "Be clear about capabilities and limits." [3] To encourage wide use and collaboration, the Model Spec is dedicated to the public domain, though OpenAI explicitly notes that current production models do not yet fully reflect the complete guidelines. [1]

These governance and behavior updates follow the July 9, 2026, rollout of GPT-5.6 Sol, which OpenAI describes as its flagship reasoning model for complex tasks spanning coding, scientific research, cybersecurity, computer use, and design. [3] Alongside this deployment, the artificial intelligence community is highly anticipating an upcoming OpenAI model codenamed "Astra," which reportedly emphasizes multi-agent capabilities. [4] According to a World of AI report, the Astra architecture allows the system to efficiently delegate specific functions to specialized sub-models, such as utilizing Codex for programming tasks while managing natural language queries with a completely different agent. [4]

To free up capacity and better serve its most capable reasoning models, OpenAI is continuing to retire older models that have limited usage in ChatGPT, including the OpenAI o3 model which will be officially sunset on August 26, 2026. [3] This transition follows an earlier interface update on May 28, 2026, where canvas functionality was removed from GPT-5.5 Instant and replaced entirely with direct writing and coding blocks in chat responses. [3] On the operational infrastructure side, OpenAI experienced a degraded performance event on August 13, 2026, resulting in elevated errors specifically for free ChatGPT users over a span of eight hours and 38 minutes. [2]

The dual focus on defining model behavior through the public Model Spec and deploying advanced reasoning engines like GPT-5.6 Sol highlights OpenAI's transition toward more autonomous systems. By retiring legacy models such as o3, the company is reallocating critical compute resources to support the complex routing required by upcoming multi-agent architectures like Astra. This shift mirrors broader industry trends toward specialized agent delegation, positioning Astra to compete directly with highly anticipated rival releases like DeepSeek V5. What the sources don't address: How the multi-agent delegation in the forthcoming Astra model will impact pricing and token consumption for developers relying on the API.