ChatGPT’s New Work Features Explained: Skills, Tokens, Tools, and What You’re Actually Seeing

ChatGPT now exposes more of the machinery behind a task. The labels are useful, but only if skills, tools, plugins, context, tokens, credits and usage limits are not mistaken for one another.

Published: August 15, 2026
Estimated reading time: 10 minutes

ChatGPT used to make most interactions look the same: you typed a message, the model returned an answer, and almost everything between those two events remained out of sight.

That description is no longer enough. A substantial task can now show progress updates, tools being called, a skill being used, files being created, approvals being requested and a usage meter moving. The same app may offer Chat, ChatGPT Work and Codex. Projects can carry shared material into later chats, while personalization can carry selected preferences more broadly. Long conversations may also be compacted so the system can keep working without retaining every earlier detail in its original form.

These features make the product more capable, but the interface mixes terms from several different layers. A token is not a credit. A tool is not a skill. A project is not memory. A visible activity log is not the model’s private reasoning. ChatGPT Work is not simply a longer reply box.

The cleanest way to understand the new interface is to ask two questions: what kind of thing is each label, and what does it let you control?

Chat, ChatGPT Work and Codex are three working modes, not three intelligence levels

Regular Chat is still the direct conversation surface. It fits questions, explanations, brainstorming and shorter drafts. It can use capabilities such as web search or file analysis when they are available, but the normal interaction centers on the answer.

ChatGPT Work is built around a task with a reviewable outcome. You give it a result to produce, the source material and constraints, and it can work through multiple steps with files, plugins and approved tools. The interface is designed so you can follow progress, answer a question, redirect the job or approve an important action before it happens. A Work task may run locally when it needs files or applications on your computer, or in the cloud when that option is available and the task should continue without your computer remaining awake.

Codex is the software-and-project-oriented surface. In the command-line version, the directory where it starts becomes the project context. In the desktop app or an editor, it can work with files, run development tools and make changes inside the access boundaries it has been given. OpenAI’s current documentation says ChatGPT Work and Codex share the same core capabilities for non-coding work; the practical difference is the working environment and the way the job is organized, not a simple ranking of which one “thinks harder.”

Model choice, reasoning settings, available tools, plan and workspace policy can still change what any of these surfaces can do. Switching from Chat to Work does not automatically make a weak instruction precise, and switching to Codex does not make an unverified claim true.

A skill is a reusable method

A skill packages instructions and supporting material for a particular kind of work. It might define how to research a topic, format a report, review a spreadsheet or verify a website. A skill can include templates, examples, reference material and optional scripts or connected tools. Its purpose is consistency: instead of restating the same process in every chat, the process can be applied when the task calls for it.

Skills can activate in two ways. You can select or mention one explicitly—ChatGPT supports @ mentions and Codex supports $ mentions—or ChatGPT and Codex can choose one when your request matches the skill’s description. The second route is why a task may report “Skills used” even though you did not type a skill name.

When the interface shows that label, it is reporting that a packaged workflow was selected or invoked. Available, selected, invoked and successfully followed are different states. The label does not mean the base model was retrained for you, and it does not prove that every instruction inside the skill was followed perfectly. The useful evidence is still the result: whether the required sources were used, whether the expected checks ran and whether the output satisfies the stated constraints.

A skill’s availability can also vary by product surface, installation and workspace policy. Seeing a skill in one ChatGPT or Codex environment does not guarantee that the same skill is enabled everywhere else.

A tool performs an action; a plugin packages capabilities

A tool is something the system can call to do work beyond generating text. Web search can retrieve current pages. A browser can interact with a site. A file tool can read or write a document. A connected service can fetch repository, drive or messaging data. Tools have inputs, outputs and permission boundaries, and a task may call the same tool more than once.

A plugin is the installable package around capabilities. Under OpenAI’s current terminology, a plugin can contain skills, connectors or both. A connector is commonly backed by a Model Context Protocol server that exposes approved data or actions from another service. The distinction matters because installing a plugin does not mean every bundled action runs automatically. The task still has to select the relevant capability, receive permission where required and execute the call.

Installed, connected, authorized, invoked, completed and validated are also different states. A plugin can be present without being used. A connector can be signed in without giving every task access to every action. A tool call can complete without producing a correct conclusion.

The activity shown during a Work or Codex task is therefore a reported execution trace. It may tell you that a tool was called, a page was searched, a file changed or a skill was applied. It is valuable because it exposes concrete actions that can be checked. It is not a transcript of the model’s private chain of thought, a complete security audit or proof that an external source was interpreted correctly.

Treat activity reporting as evidence about what happened, not a guarantee about the quality of the conclusion. If correctness matters, inspect the cited source, open the created file, run the relevant test or ask for the result that proves the step succeeded.

Tokens measure model input and output

Models do not process language as whole words and paragraphs. They process tokens: small units of text and other encoded content. A task can consume tokens from your prompt, earlier conversation, project instructions, attached files, retrieved pages, tool results and the model’s output.

The context window is the amount of material available to the model during a particular run. Context is the content; tokens are the units used to measure it. A long conversation can occupy substantial context even when the latest prompt is only one sentence. A short prompt can also start an expensive job if it asks the system to inspect many files, search multiple sources or iterate through a difficult task. A published model context-window specification does not reveal how much of a live ChatGPT or Codex conversation is currently occupied or what remains active inside it.

That is why prompt length alone does not predict usage. OpenAI’s current usage documentation says model choice, context, reasoning, tool use, retrieval and caching all affect consumption. Two requests that look similar in the chat window can use different amounts because the work behind them is different.

Tokens should not be read as a measure of intelligence, effort or “thoughts.” A higher token count can reflect a larger source file, repeated tool output or a verbose response just as easily as deeper analysis.

Credits and usage limits are accounting layers

Credits translate eligible product usage into an account balance or billing mechanism. They are related to token consumption, but they are not interchangeable with tokens. The credit cost of a task can vary with the model, context, reasoning and tools involved. One credit is therefore not a fixed number of words, messages or minutes across every kind of work.

Usage limits are the boundaries applied to a plan, workspace or time window. Some plans include a shared allowance for local messages and cloud work; eligible accounts may be able to buy credits after included usage is exhausted. Exact allowances, supported models and time windows can change, so a static article should not freeze today’s pricing table into a promise about tomorrow.

The practical reading of a usage meter is modest: it shows how much of an applicable allowance remains under the current product rules. It does not tell you precisely how many future tasks you can complete. A simple request may use a small fraction of the allowance, while a long job with extensive context, tools and reasoning may use much more. The ChatGPT usage dashboard is the current account-level reference; in Codex CLI, the documented /status command can report context usage and rate-limit information for the active session.

Projects, personalization and context solve different problems

A ChatGPT project groups related chats, files, instructions and connected sources. Chats started with regular Chat and ChatGPT Work can live in the same project and use its shared material. That makes a project the right place for context that belongs to a continuing body of work rather than to every conversation you will ever have. Project membership makes that material available as shared context; it does not prove that every file or prior chat was opened, retrieved or reviewed for a particular answer.

A local project is different. In Codex, a working directory or editor workspace supplies the files that the task can inspect. The ChatGPT Projects view and a folder on your computer are not automatically the same thing. The command-line interface does not expose the ChatGPT Projects view, and durable repository guidance belongs in project documentation such as AGENTS.md rather than in an assumption that every new chat will remember it.

What users sometimes call “Personal Context” is not one single, precisely bounded feature in the current documentation. The official controls are described under Personalize ChatGPT. Personality changes communication style, not capability. Custom instructions carry preferences across chats. Memories can carry useful facts, recurring workflows and conventions. Project instructions apply within the project. ChatGPT memories and a separate local Codex memory store should not be treated as one universal memory database. These sources can overlap, but they have different scope and different controls.

That separation is important. A preference remembered from an earlier chat is not a substitute for a requirement that must govern a project. If an instruction is essential, put it in the current request or in the project’s durable guidance and verify the output against it.

Compaction keeps long work possible by reducing prior context

Context cannot grow without limit. In long-running work, earlier material may need to be reduced so later turns still fit. OpenAI’s public technical documentation calls one version of this process compaction: prior state is carried forward in a smaller representation so the interaction can continue with fewer tokens. Codex CLI also documents an explicit /compact command.

The public compaction guide describes an API mechanism, not a promise that every ChatGPT interface uses an identical implementation. The safe general conclusion is that long tasks may preserve the information needed to continue without retaining every earlier message verbatim in the active context.

Compaction is not the same thing as memory. Memory is selected context intended to carry across chats. Compaction manages the working state of a long interaction. It is also not perfect recall. A compressed representation can preserve the main objective and still lose a small exception, an exact quotation or a constraint that seemed unimportant when it was condensed.

For long or high-stakes work, restate the non-negotiable requirements, keep authoritative facts in files and check that the current output still obeys them. That is more reliable than assuming an old instruction remains fully represented simply because the conversation continues.

What you can control and what you can only observe

You can control the requested outcome, the sources and files supplied, explicit skill selection, connected plugins, model and mode where those choices are offered, and whether a task runs locally or in the cloud where that choice exists. You can set project instructions and personalization controls. Permissions can constrain which files or network destinations Codex may access. Approval prompts can stop sensitive actions until you authorize them. During a task, you can answer questions, redirect the work or tell it to stop.

You can observe progress messages, named skills, tool calls, requested approvals, citations, created files and some usage information. These are useful windows into execution. You generally cannot observe every internal model operation, recover a complete private reasoning trace or infer the exact correctness of an answer from how busy the activity panel looked.

The distinction is simple: controls change the task’s permitted inputs, behavior or boundaries; indicators report some part of what the system did. Neither replaces verification.

If the failure you are trying to prevent is skipped steps, unsupported claims of completion or confident arithmetic that does not survive a check, this article’s companion is Why Does ChatGPT Sometimes Skip Steps and Think 1+1=3?. ART-003 focuses on evidence and verification. ART-047 explains the interface machinery around that work.

The bottom line

The new labels are not decorative. They expose real differences in how a task is organized, what it can access and how usage is counted. But the labels only help when they are read at the correct layer.

A skill supplies a reusable method. A tool takes an action. A plugin packages skills and connections. A project supplies shared work context. Personalization carries selected preferences. Tokens measure processed material. Credits and limits account for product usage. Compaction reduces the active history of a long interaction. ChatGPT Work organizes a delegated task, while Codex supplies a project-centered environment for work with files and software.

What you see is more transparent than the old single-response interface, but it is still a partial view. The durable rule is the same one that applies to any capable technical system: give it clear boundaries, inspect the actions that matter and verify the finished result.

Product definitions and availability were verified against official OpenAI documentation on August 15, 2026. Features, plan rules, limits and terminology can change.