OpenAI Product and Platform Head Tibo: Developers Can Use Plugins to Build Complete Native Apps and Directly Deploy to ChatGPT
OpenAI's product and platform head Tibo Sottiaux stated that the company is opening its platform: developers can use plugins to build complete native applications and directly deploy them to ChatGPT; the official target is approximately 1.2 billion weekly active users, with relevant plugins recommended during conversations.
A new version of Codex Cloud is being launched, emphasizing a configurable cloud environment compared to last year's version. The statement is: after spending time configuring the cloud environment, it is difficult to revert to building solely on a notebook. This cloud environment is aimed at teams migrating building, running, and collaboration away from local machines, rather than just local completions.
The same technology driving the company's cloud agents (including Dots) is being previewed with the Agents API, which supports computer usage. Developers can build agents similar to products launched on the same day, with OpenAI managing orchestration, sessions, and context, eliminating the need to build a complete harness from scratch.
The Decisions API is based on gpt-6 luna, aimed at rapid decision-making with limited options, supporting visual input. The official claim is that end-to-end processing can be completed in hundreds of milliseconds; reports indicate about 150 milliseconds to return preset answers and confidence levels, approximately ten times faster than the standard luna interface, currently in limited preview with expansion in a few days. Input consists of context plus questions plus a closed set of answer options, with output restricted to the given options.
The plugin layer connects skills, applications, and mcp into reusable packages, with Chat and Codex sharing the same directory, usable in web, desktop, and mobile Chat and Work; Codex CLI has plugin browsing, but IDE extensions are not yet supported. In-conversation distribution means the distribution entry shifts from the app store to ongoing conversations.
In market mechanics, developers are buying to leverage the 1.2 billion weekly active users for a native experience and to separate the "next step" in the agent loop from large model calls; what is sold includes ChatGPT session slots, hosted agent runtimes, and luna decision files. Funding shifts from self-built orchestration to platform commissions; beneficiaries are plugins that can enter conversation recommendation slots and cloud agents that are isomorphic to Dots, while those under pressure are only selling local IDE plugins and solutions that use complete generative models for three-way routing. This is an event-driven platform opening, not just a single model release.
In public comparisons, Decisions is placed in the same category as independent decision models: closed options, visual input, sub-second returns. The Agents API extracts computer usage from Codex's in-app capabilities into an embeddable managed service, forming two sides of the same technology sold externally—one for user products and one for developer interfaces.
Source: Public Information
ABAB AI Insight
Tibo's product line history involves expanding Codex from a command-line coding assistant to a desktop-capable, plugin-installable, cloud-running workstation, and then renting out the same harness externally. The 2025 Codex Cloud still emphasizes "throwing tasks to the cloud," acknowledging that the default sandbox is insufficient, with real stickiness coming from teams integrating keys, images, and intranet tools into cloud hosts, creating migration costs. Plugins have evolved from earlier skill packages targeting Claude Code to native application distribution within ChatGPT, first serving coding users and then leveraging conversation recommendations to reach consumer-level weekly activity.
The capital path divides into three accounts. Plugins benefit from distribution: 1.2 billion weekly active users serve as the shelf, and recommendations in conversations act as shelf lights. The Agents API benefits from runtime: orchestration, sessions, and computer usage remain within OpenAI's cluster, with developers paying for inference and hosting, no longer needing to build long-term agent infrastructure. The Decisions API benefits from frequency: each routing, each button click, each email response is a small model with high QPS, and luna's low unit price can only support hundreds of micro-decisions in agents when made into dedicated interfaces. The motivation is to sell the "layer we use for Dots" as a platform tax, rather than just selling tokens.
A comparable example is Apple harvesting Cocoa capabilities into the App Store, and collecting push notifications and in-app purchases as system services; more recently, TypeSafe's Jev has turned "smart if-then" into an independent model, with OpenAI countering using image-enabled luna interfaces. The industry phase is transitioning from model APIs to operating system phases: whoever controls the conversational surface, the agent harness, and sub-second decisions defines what others can build on top of it. Claude and Cursor remain strong in coding loops, while OpenAI this time is breaking the loop into three external bricks.
Structural judgment indicates a reconstruction of the industry chain. The mechanism involves slicing an internally integrated agent stack into separately priced layers: distribution layer (ChatGPT plugins), execution layer (Agents API plus computer usage), and decision layer (Luna Decisions). Reconstruction occurs because complete large models performing three-way selections are slow and expensive, and local notebooks cannot adequately configure enterprise environments, while 1.2 billion weekly active users, if not opened up, would stagnate into only nurturing their own Dots. Once the platform simultaneously masters these three layers, the bargaining space for application developers shifts from "changing models" to "changing shelves."
ABAB News · Cognitive Laws
- First rent out the self-use engine, then the platform is considered open.
- Large models handle thinking, while small models handle decision-making.
- Weekly activity is the shelf, while recommendations in conversations are the rent.