Major Update for ChatGPT: Cross-Platform Functionality, One-Click Website Creation, and Lower Costs

On July 9, OpenAI released the GPT-5.6 series models and announced the official merger of the ChatGPT desktop application with Codex. In this update, the most noteworthy aspect is not merely the increase in parameters, but the reshaping of boundaries in two dimensions: first, in model pricing, Terra offers performance close to GPT-5.5 at half the price, while Luna reduces API costs to a very low level; second, in terms of tool functionality, the merged application introduces a Work mode that supports executing complex projects across platforms like Slack and Google Drive, along with scheduled task capabilities. This marks a transition of AI tools from "human-triggered - machine responds" synchronous dialogue to "machine listens - machine executes" asynchronous workflows.

ChatGPT Work supports cross-platform execution and multi-device synchronization, image source: OpenAI official blog

Pricing Strategy and Cost Restructuring of GPT-5.6

OpenAI has implemented an aggressive "market segmentation by execution cost" strategy for the GPT-5.6 series. The new series includes the flagship Sol, cost-effective Terra, and lightweight Luna models, each targeting different execution scenarios and cost tolerances.

In terms of API pricing, Sol charges $5 per million tokens for input and $30 per million tokens for output, consistent with the previous flagship GPT-5.5. Luna charges $1 per million tokens for input and $6 per million tokens for output, positioning itself as a lightweight model. The real game-changer for developers is Terra, which has an input price of $2.5 per million tokens and an output price of $15 per million tokens, offering performance close to the previous flagship GPT-5.5 but at half the price.

For enterprise procurement decision-makers and mid-tier developers, Terra provides a clear path to cost reduction. When handling large-scale text analysis, code refactoring, and other non-extreme reasoning-intensive tasks, using Terra instead of GPT-5.5 can directly halve API expenses. For instance, if a team consumes 1 million tokens for input and 500,000 tokens for output monthly, migrating from GPT-5.5 to Terra would reduce monthly API costs from about $20 to about $10. This pricing strategy indicates that OpenAI no longer solely relies on Benchmark scores to differentiate models but instead offers clear cost-performance gradients, allowing enterprises to choose computing power based on the actual ROI of tasks.

Meanwhile, Luna's extremely low pricing is not merely aimed at the lower market but provides economic feasibility for asynchronous workflows. In backend tasks requiring high-frequency calls and low-latency responses, such as data cleaning, log analysis, or scheduled inspections, the cost per invocation of the model becomes crucial for large-scale deployment. Luna's input and output prices are only one-fifth of Sol's, enabling developers to keep AI running in the background without worrying about runaway costs. This "cheap labor" positioning naturally complements the scheduled task functionality in the subsequent Work mode—using Luna for background scheduled summaries and Sol for handling complex reasoning in the foreground.

Sol, on the other hand, raises the ceiling with newly added max and ultra reasoning tiers. The max tier targets deep reasoning needs, while the ultra tier supports multi-agent parallel processing, designed for high-level and enterprise-grade workflows. In complex codebase refactoring or cross-system data integration scenarios, the ultra tier can dispatch multiple sub-agents to handle different sub-tasks simultaneously and then aggregate the results. This tiered approach not only covers all needs from low-end to high-end but also deeply binds model capabilities to specific execution scenarios, laying a computational foundation for future tool mergers.

From Synchronous Dialogue to Asynchronous Cross-Platform Execution

If model layering provides different cost execution engines, the merger of the ChatGPT desktop application with Codex offers the physical carrier for executing these engines. The merged application uniformly provides three work modes: Chat, Work, and Codex, creating a complete productivity spectrum from daily Q&A to cross-platform project execution and sandbox-level code engineering.

The Chat mode still handles daily synchronous dialogue functions, used for lightweight creation and quick Q&A, with interaction methods largely consistent with previous versions of the ChatGPT desktop application. The Codex mode focuses on heavy code engineering, adding file editing, Pull Request review, and Ultra mode, providing developers with a sandbox-level programming environment. The browser side of the Codex mode has also upgraded its support for CDP (Chrome DevTools Protocol), allowing AI to inspect network traffic, perform multi-tab analysis, and manage login states, further expanding the boundaries of code debugging and web application testing.

What truly breaks the boundaries of traditional AI tools is the Work mode. ChatGPT Work supports executing complex projects across platforms like Google Drive and Slack, and includes scheduled task functionality. This means AI is no longer just a passive response box but an automated terminal that can actively listen, schedule, and execute tasks.

Specifically, the Work mode connects external platforms through a Plugins mechanism. Users can describe task objectives in natural language, and AI will automatically break them down into multiple steps, retrieving discussion records from Slack and project documents from Google Drive, then synthesizing them into an analysis report. More critically, the scheduled task functionality allows users to set up a background asynchronous execution plan similar to Cron, such as "every Monday at 9 AM, summarize the key discussions in the Slack channel and generate a progress report based on the project documents in Google Drive." At the scheduled time, AI will automatically execute the entire process without manual triggering.

This capability signifies that AI tools are beginning to penetrate the territory of traditional RPA (Robotic Process Automation) and automation workflow platforms. Compared to traditional RPA, which requires writing complex rule scripts, the Work mode significantly lowers the barrier to building automated workflows through the advantages of natural language understanding. The productivity boundary of AI tools has expanded from "human-triggered - machine responds" synchronous interaction to "machine listens - machine executes - human approves" asynchronous workflows. ChatGPT is transforming into a combination of an AI-brained Zapier and an automation terminal.

One-Click Website Creation and Code Migration Behind Ecological Excavation

While reshaping workflows, OpenAI is also accelerating ecological binding through functional loops and lowering conversion costs.

The Sites feature allows users to publish visual content as a website with one click, and the browser side upgrade supports login state operations and multi-tab functionality. This means that analysis results or code applications generated in Work or Codex mode can be directly deployed as web apps with login states. After completing data analysis in Work mode, users can publish an interactive dashboard as an accessible webpage through Sites, allowing team members to log in and view it via Sign in with ChatGPT. For internal enterprises, the Sites feature locks in the distribution channel for lightweight applications. In the past, developers needed to find servers and configure deployment environments to allow team usage after generating code with AI. Now, this process has been compressed into a one-click operation, significantly shortening the delivery cycle from idea to application.

In the developer tools market, OpenAI has demonstrated a more direct competitive strategy. Codex has added functionality for migrating from Claude Code. The official GitHub repository provides a dedicated migration script that can accurately scan the configuration directory of Claude Code and convert agents, MCP servers, and hooks into Codex format with one click.

As a strong competitor in the AI programming assistant field, Claude Code's configuration system represents user habits and project accumulation. By actively providing official migration tools, OpenAI is lowering the conversion costs for competitors' users in the red ocean of AI programming assistants. This indicates that as the gap in fundamental model capabilities narrows, the focus of competition has shifted from pure model performance to ecological binding and user migration costs. Whoever can more smoothly take over the existing assets of competitors' users will occupy a larger share of the developer market.

The Bulky and Permission Risks Brought by the Merger

Despite the layered pricing of GPT-5.6 and the tool merger showcasing ambitions to reshape productivity boundaries, this strategy also comes with significant limitations and risks.

First is the issue of application bloat and cognitive burden. Merging a heavy code sandbox like Codex with the everyday office terminal of ChatGPT into a single desktop application has raised concerns among some developers. Whether the switching between different modes is smooth and whether the stacking of functions will lead to software bloat are practical issues that need to be addressed. For developers who only need simple code completion, facing a super application that includes various complex modes like Work and Codex may increase unnecessary cognitive load. How to achieve seamless yet non-interfering switching between different modes will test OpenAI's product design capabilities.

Second is the challenge of enterprise data privacy and permission control. The Work mode deeply integrates with core data sources like enterprise Slack and Google Drive, and Sites supports one-click publishing of public applications. This requires enterprise procurement decision-makers to establish strict permission management systems while enjoying the convenience of automation. Although OpenAI emphasizes default settings for enterprise-level data not being trained and access control, in practice, preventing AI from accidentally leaking sensitive information during cross-platform execution remains a key risk that enterprise compliance departments need to evaluate. Especially when AI needs to schedule data across platforms and generate publicly accessible Sites, auditing the flow of data becomes exceptionally complex.

More critically, there is the risk of open permissions at the browser level. After the upgrade of Codex on the browser side, CDP (Chrome DevTools Protocol) permissions were introduced, allowing AI to inspect network traffic, perform multi-tab analysis, and manage login states. While this deep control over the browser significantly expands AI's execution capabilities, it also opens potential security vulnerabilities. If the AI's prompt is maliciously injected, an AI with CDP permissions could be induced to steal user login credentials or intercept sensitive network traffic. Although the official requires explicit user approval and administrators can globally disable this permission, this leap from "read-only" to "low-level control" raises higher demands for terminal security protection.

Through the GPT-5.6 series and tool merger, OpenAI clearly outlines a path for reshaping productivity from dialogue to execution. However, behind the cost advantages brought by model layering and the expanded execution boundaries brought by tool mergers, how to balance functionality richness with software lightweighting, and how to maintain security while opening permissions, will determine how far this strategy can go.