How ChatGPT + Codex Turns the Chat Window Into a Full‑Featured Workbench

Opening Update

On July 10 OpenAI launched GPT‑5.6 and combined ChatGPT with Codex into a single desktop application called ChatGPT Work . The update also removed the five‑hour usage limit for Plus, Business, and Pro plans, optimized Sol model efficiency, and pushed active users past six million.

From Chat to Execution

Previously ChatGPT functioned as a supercharged input box: users asked questions, received answers, pasted code for explanations, or received draft documents. Codex introduced a different logic where the user provides a goal and the agent reads files, modifies code, runs tests, and returns results, operating continuously over a task cycle.

The merger places this execution‑oriented workflow directly into ChatGPT’s main entry point. The new desktop app bundles Chat, Work, and Codex modes in one download for macOS and Windows, while the older desktop client is renamed ChatGPT Classic .

GPT‑5.6 Model Tiers and Pricing

GPT‑5.6 is split into three tiers:

Sol : flagship tier for complex reasoning and long‑running autonomous work, suited for heavy code migration or deep analysis.

Terra : balanced tier, performance close to the previous GPT‑5.5 flagship at half the price.

Luna : lightweight tier emphasizing speed and low cost.

API pricing per million tokens reflects this hierarchy: Sol $5 input / $30 output, Terra $2.5 / $15, Luna $1 / $6. Free and Plus users default to Terra; Plus‑plus and higher can select other models. Sol also offers “max” and “ultra” inference levels, with ultra enabling multiple sub‑agents to run in parallel for Pro and Enterprise customers.

The article stresses that model choice should match task demands: lightweight models for routine drafting, heavier models for large‑scale code refactoring or prolonged autonomous work, akin to allocating cloud resources.

ChatGPT Work’s Expanded Scope

ChatGPT Work aims to extend Codex’s execution capabilities beyond pure coding to everyday office and business processes. Users can give a goal, have the system break it into steps, operate across applications and files for hours, and finally produce spreadsheets, slides, documents, or shareable web apps.

Integration plugins include Slack, Teams, Google Drive, email, calendar, and CRM, invoked with the @ symbol to fetch information from specific apps. Combined with scheduled tasks, the agent can generate weekly meeting agendas from channel discussions or continuously update reports based on incoming emails. The built‑in browser and desktop automation let it handle tasks lacking public APIs, such as interacting with web‑only interfaces.

Relaxed Usage Limits Signal an Execution‑Efficiency Race

On July 13 OpenAI temporarily lifted the five‑hour cap and further optimized Sol’s token consumption, allowing users to accomplish more work with fewer credits. The article argues that as AI shifts from answering questions to executing tasks, time and quota limits become critical friction points; long‑running tasks would suffer disruptive interruptions under strict caps.

Competing products like Claude Code, Claude Cowork, and other AI workbench solutions vie for the same execution layer, where model capability, cost, stability, tooling ecosystem, and cross‑application ability all matter.

Practical Guidance for Developers

Treat ChatGPT as a workflow entry point rather than a simple search box; provide complete task descriptions, constraints, input files, and acceptance criteria in one prompt.

Select the appropriate model tier per task—use lighter models for document summarization or simple scripts, reserve Sol or ultra for heavy refactoring or multi‑agent parallelism.

Define deliverables precisely, e.g., instead of “generate a table,” specify “create a table from these 20 feedback items, grouped by module, severity, and owner, and include top‑three priority recommendations.”

Maintain audit trails and human confirmation for privileged actions such as deleting files, sending emails, or committing code.

Relationship Between ChatGPT Work and Codex

Codex can be seen as an execution agent focused on development scenarios—codebases, command lines, testing, and commits. ChatGPT Work is a more generalized work agent that emphasizes documents, spreadsheets, calendars, messages, web pages, and business systems. After merging into a single desktop app, the distinction blurs, allowing developers to generate release notes while operations teams can build data dashboards.

For Chinese users facing subscription, network, or stability hurdles, services like Code80 can proxy official API keys, swapping endpoints to enable integration with IDEs, CLIs, or automation pipelines.

FAQ Highlights

Did Codex disappear after the merger?

No; its capabilities are now embedded in the unified ChatGPT entry point, reducing the number of entry points while expanding task types.

How to choose among the three GPT‑5.6 tiers?

Use Luna or Terra for lightweight tasks, Sol for complex reasoning and long‑running autonomous work, and consider Sol’s ultra level only when multiple agents need to run in parallel.

Which is better for developers, ChatGPT Work or Claude Code?

Claude Code leans toward deep code‑base development, whereas ChatGPT Work emphasizes cross‑application and general office automation; the choice depends on the primary workflow of the team.

Will the temporary removal of the five‑hour limit stay?

It is currently a temporary relaxation, highlighting that execution‑oriented AI is highly sensitive to continuous usage time; future policy will depend on system load and strategic decisions.

How can domestic users integrate these capabilities into their tools?

They can use the official API or adopt endpoint‑replacement services like Code80 to connect models such as GPT, Claude, or Gemini to IDEs, CLIs, or automated pipelines.