XNews

OpenAI GPT-5.6 and ChatGPT/Codex/Work Launch

OpenAI GPT-5.6 and ChatGPT/Codex/Work Launch

OpenAI released GPT-5.6 (Sol, Terra, Luna) on July 9, 2026, and unified ChatGPT and Codex into a single desktop application with three modes: Chat, Work, and Codex. ChatGPT Work is a new agent that connects to Gmail, Slack, Drive, calendars, and CRMs to autonomously execute multi-step tasks and produce finished documents, spreadsheets, and web apps. - GPT-5.6 pricing: Sol $5/$30, Terra $2.50/$15, Luna $1/$6 per million tokens input/output. - Free/Go users default to Terra; Plus+ can choose models and effort levels; Ultra effort for Pro/Enterprise. - Work and Codex share an agentic usage pool separate from Chat message limits; quotas based on tokens. - Work supports Scheduled Tasks, browser login-state operations, and desktop Computer Use. - Codex can import settings from Claude Code; original Codex app merged into ChatGPT desktop. - Microsoft/Google competition: OpenAI's super-app strategy targets workspace productivity, coinciding with IPO filing. - User insights from a DevEx interview: heartbeat routines for triage, file-based memory (Obsidian vault for agents), write-me skill to mimic writing style, goal.md/plan.md/worklog.md for long tasks, using session history to prune skill libraries.

Key insights

GPT-5.6 Model Family Covers Full Price-Performance Spectrum with Effort Controls

OpenAI introduced three GPT-5.6 tiers: Sol for complex reasoning and high-difficulty coding ($5/$30 per million tokens), Terra as the balanced default ($2.50/$15), and Luna as the affordable speed option ($1/$6). Users can adjust reasoning effort, with Ultra mode in Work restricted to Pro/Enterprise and in Codex available from Plus. This structure allows users to match model cost to task difficulty, avoiding overpaying for simple tasks while retaining frontier power for hard problems.

ChatGPT Work Transforms the Assistant into a Cross-App Task Execution Agent

ChatGPT Work moves beyond Q&A: it connects to plugins (Gmail, Google Drive, Slack, SharePoint, Teams, CRM) and autonomously gathers data, integrates information, and produces finished deliverables such as reports, spreadsheets, and web apps. Unlike Chat, it stays engaged for hours, breaking down complex projects, asking for input only when needed, and supports scheduled recurring tasks. Enterprise examples include lead-tracking review systems and competitor analysis pipelines compressing weeks into hours.

ChatGPT and Codex Unification Creates a Single Super-App Platform

OpenAI merged the standalone Codex app into ChatGPT desktop, offering Chat, Work, and Codex modes sharing the same agent engine. Codex focuses on codebases, diffs, and PRs; Work hides code and targets business documents. This unification aims to capture both developers and office workers, with over 1 million existing Codex users already doing non-programming tasks. A direct Claude Code import feature targets Anthropic's developer base.

Agentic Usage Quota Is Separate from Chat, Token-Based, and Shared Across Work and Codex

ChatGPT Chat, image generation, and voice have independent quotas. Work and Codex draw from a single agentic usage pool, meaning heavy Work tasks reduce available Codex capacity. Quotas are measured by token consumption, with plans from Free (trial) to Enterprise (150+ users). Plus ($20/month) is the first ad-free full-feature tier; Pro ($100/$200) offers 5–20x the Plus quota. Credits can be purchased after exhaustion.

Advanced User Workflows Center on File-Based Context and Minimized Automation Scope

An OpenAI DevEx team member demonstrated turning ChatGPT Work + Codex into a personal OS. Key practices: important context is stored in files (markdown notes, Obsidian vault) rather than conversation history; a chief-of-staff agent reads emails/Slack but never auto-sends because of information breadth; output is directed to actionable interfaces (browser tabs for shopping, pre-filled draft links); skill libraries are continuously pruned by analyzing session history to remove unused or feedback-heavy items.

Long-Running Tasks Benefit from Goal-File Architecture and Write-Me Personalization

For complex tasks, the user recommends a three-file method: goal.md with verifiable success criteria (enforced by the model), plan.md for implementation details, worklog.md for human-readable logs. The 'write-me' skill reads a user's past written communication to generate style-matching rules per audience. These patterns reduce context drift and improve output relevance in prolonged agent sessions.

OpenAI Positions the Unified Platform as a Workspace Competitor Ahead of IPO

By merging chat, coding, browsing, file operations, and cross-app automation into one desktop and mobile app, OpenAI directly challenges Google Workspace and Microsoft 365. The inclusion of Sites (shareable dashboards and tools from conversations), upgraded browser with login-state and multi-tab support, and computer-use capabilities signal an enterprise lock-in strategy. This release coincides with OpenAI's SEC IPO filing, underpinned by enterprise clients contributing 40% of revenue.

Voice and Cloud Automation Are Next Frontiers for ChatGPT Work

The DevEx interview revealed that Codex capabilities are migrating to the cloud, enabling automation to run server-side rather than only on a user's desktop. Voice dictation is already heavily used, and a future feature will allow making phone calls to instruct Codex. This indicates OpenAI's roadmap for always-on, headless agent execution and ambient computing interfaces.

Source posts