Publication

GPT-6 and the New ChatGPT Architecture New capabilities, user tools, the agentic model of work, and OpenAI's direction

Sep 26, 2026 · 2 authors

Abstract

The central shift: from answers to completed workOpenAI is moving ChatGPT away from the traditional question -> model -> answer paradigm toward a broaderoperating model: goal -> agent -> context -> tools -> actions -> verification -> finished result.GPT-6 matters not only as a more capable model. It increasingly acts as the intelligence layer behind a largerexecution environment that includes ChatGPT Work, Codex, Skills, Plugins, Workspace Agents, Cloud Browser,Computer Use, Desktop, long-running tasks, and persistent context mechanisms.KEY THESISThe most important transformation is larger than a model upgrade. ChatGPT is gradually becoming ageneral-purpose AI work environment in which the model is only one layer of the system.Verification notes and corrections The core GPT-6 family is best described as Astra, Sol, and Luna. GPT-6 Pro is a product-level configuration oraccess mode powered by Astra, not a fourth foundational base model. Workspace Agents should not be described as universally available. Access depends on plan, workspacesettings, administrator policy, region, and rollout status. Memory is not synonymous with every persistent context source. Projects, Library, connected apps, files, andComputer History are separate context mechanisms. OpenAI has documented a migration path from Custom GPTs toward plugins. The standard retirement date forCustom GPTs is December 11, 2026, with specific Enterprise exceptions and extensions. WebMCP should be described as a proposed web standard. Site Tools already exist, but WebMCP is not yet auniversally adopted industry standard. Sites and the Agents API remain beta/public-beta products; their strategic importance should not be confusedwith full maturity or universal availability.Important GPT-6 agentic capabilitiesSeveral GPT-6 features are especially relevant to agent architectures: async tool calling, mid-turn steering,reasoning-effort changes during continuation, subagent delegation in supported environments, and persistent notesfor long-running work in Codex. These capabilities push the interaction model beyond a simple synchronouschatbot

Showing the abstract — retrieve the full paper via the Exa API.

Authors

A.A. MalachevskyAlex Malachevsky

About

PublishedSep 26, 2026
Citations0

Powered by the Exa API