From GPT-6 Astra to Work and Codex: A Complete Guide to ChatGPT

Author: AIB Team·Sunday, September 20, 2026
From GPT-6 Astra to Work and Codex: A Complete Guide to ChatGPT

“Everyone says GPT-6 Astra is incredible. Is that different from ChatGPT?”
“And what exactly is Codex? Is it another AI? I keep hearing it’s better for coding.”

Once you start using AI, the names alone can get confusing.

Even within ChatGPT—the AI product most people are familiar with—you’ll run into names like ChatGPT, Work, Codex, and GPT-6 Astra.

Then there’s Effort. What does that control? What do you actually get by paying for ChatGPT? And when people say a paid AI is “smarter,” is that really true?

Using OpenAI’s ChatGPT as our example, let’s break down the models, services, agents, and subscription plans—and how they fit together.

GPT is the “brain.” ChatGPT is the service.

  • GPT-6 Astra, the recently released model generating so much attention, is an AI model.
    The simplest way to think about a model is as the AI’s “brain.” Roughly speaking, the name can be broken down as GPT (model family) + 6 (generation) + Astra (model variant).

  • ChatGPT is the service that lets you interact with AI models through a conversational interface.
    The name itself is a useful clue: Chat + GPT. ChatGPT originally emerged as a way to have a back-and-forth conversation with GPT models. The ChatGPT app on your phone or computer is simply one way of accessing that service.

  • Work and Codex are both AI agents.
    An AI agent goes beyond answering a question. It can use tools, break a task into multiple steps, and carry out actual work on your behalf. The biggest difference between Work and Codex is not the underlying idea of an agent, but the kind of work each one is designed for.

  • Work is an AI agent designed for general knowledge work—research, analysis, documents, spreadsheets, presentations, and more.
    It can gather information, inspect files and connected services, work through multiple stages of a task, and produce a finished deliverable. Ask it to “research these materials and turn them into a report,” for example, and it can handle the research, analysis, and document creation as one larger job.

  • Codex is an AI agent built specifically for software development and technical work.
    Within the permissions you give it, Codex can work with project files, source code, terminals, and development tools. It can write and modify code, run programs and tests, investigate bugs, and make fixes.

A simple way to remember the distinction is:

Chat is for conversation, Work is for general work, and Codex is for software development. With ChatGPT Plus and Pro, you can use GPT-6 Astra and Codex, although access levels and usage limits vary by plan. We’ll come back to those differences later.


Why the same model can produce different results: the “harness”

A product like ChatGPT is not just the intelligence of the model itself.

Imagine hiring an extremely capable employee but giving them no documents, no computer, no software, and no access to the systems they need. Their ability hasn’t changed—but what they can actually accomplish certainly has.

AI works much the same way.

A service like ChatGPT combines the underlying model with tools such as web search, file analysis, and image generation, as well as memory, instructions, permissions, and other settings.

Suppose you ask:

“Fix the bug on our website.”

The model may first reason about what needs to be inspected. It can then read the relevant source files, use tools to modify the site, inspect the result, and make another change if the problem is still there.

The software structure that supplies a model with the right context and tools—and manages how those tools are used—is often called the harness.

Put simply, the harness is what turns a model from a “brain” into something that can actually interact with tools and get work done.

That’s why two products can both say they use “GPT-6” and still produce noticeably different results.

Even if the underlying model is identical, what the model can see, which tools it can use, what permissions it has, and how the surrounding system orchestrates its work can make a major difference.


What do you actually get with each subscription plan?

Prices below are U.S. list prices as of September 20, 2026. OpenAI supports localized billing in many markets, so the amount you see in another country or app store may be different. Always check the final checkout price before subscribing. OpenAI

PlanMonthly priceWhat you should know before subscribing
FreeFreeText chat with the standard model is available without a paid subscription. Images, files, advanced tools, and Codex have separate limits. Unlimited features remain subject to abuse safeguards.
Go$8/monthGives you more image generation, file uploads, analysis, and everyday tool usage than Free. It is mainly aimed at people who need more capacity than the free tier—not people looking specifically for GPT-6.
Plus$20/monthExpands access to features such as deep research and Codex. GPT-6 Astra can be used in Work and Codex, subject to usage limits.
Pro 5x$100/monthAdds access to GPT-6 Pro in regular Chat and provides roughly 5× the Work/Codex allowance of Plus.
Pro 20x$200/monthIncludes Pro features with roughly 20× the Work/Codex allowance of Plus.

The important point is that 5× and 20× refer to usage capacity—not to intelligence.

A Pro 20x model is not somehow “20 times smarter” than Plus.

That said, Free and Go do not give you access to GPT-6, and Plus has tighter GPT-6 usage limits. So depending on the plan, there can still be a real difference in the level of model intelligence you are able to use in practice.

There’s one more source of confusion: the same ChatGPT app presents models differently depending on whether you’re using Chat, Work, or Codex.

On Pro, regular Chat currently lets you select “6 Pro,” a Pro model based on GPT-6 Astra.

In Work and Codex, you can select GPT-6 Astra and then choose an Effort level—Light, Medium, High, Extra High, Max, and so on—depending on how difficult the task is. The exact options you see can vary by plan, settings, and rollout status.

Work and Codex can also call sub-agents that use different models and run parts of a job in parallel.

Ultra is not simply a “think longer” switch. It uses the highest reasoning setting and can, when appropriate, proactively delegate parts of the job to multiple sub-agents.

That additional compute can consume your included allowance very quickly, so it is worth using selectively.

Before you pay: check current student offers

TIP

The offers below were confirmed as of September 20, 2026. Eligibility and availability can change.

  • U.S. college and graduate students · 4 months of ChatGPT Plus free — Eligible full-time and part-time students at qualifying U.S. degree-granting institutions can verify enrollment through SheerID and claim four free monthly billing periods of Plus. The offer must be claimed by October 31, 2026. Direct ChatGPT billing is supported; subscriptions billed through the Apple App Store or Google Play are not eligible while they remain store-billed. OpenAI Help Center
  • U.S. and Canadian university students · $100 in Codex credits — Verified university students residing in the U.S. or Canada can receive $100 in Codex credits. A ChatGPT Free, Go, Plus, or Pro account is required. OpenAI Help Center

What is an API, and how is it different from a subscription?

NOTE

A ChatGPT subscription gives you access to a finished AI product.
You pay a monthly fee and use the models and features included in your plan, within the plan’s usage limits.

An API is a way for another piece of software to call an AI model.
You might use it to add an AI support assistant to an online store, automatically summarize documents inside a company system, or build your own AI-powered product. OpenAI API usage is billed separately from your ChatGPT subscription, generally based on the model and the amount of data processed.

A useful rule of thumb is this: if you are personally talking to an AI while writing, researching, or coding, a ChatGPT subscription is usually the natural starting point. If your software needs to call the AI automatically, you’re usually looking at the API.

The line is not absolute. Developers often use APIs while building and testing products as well. But if you want to build your own AI service or more advanced automations, learning how APIs work becomes increasingly useful.


Practical ways to stretch your subscription and usage limits

1. Choose the right mode for the job

Regular Chat and Work/Codex do not draw from the same usage pool.

Work and Codex share an allowance, while normal Chat usage is separate from that shared Work/Codex pool.

That means it often makes sense to do quick editing, brainstorming, or general discussion in Chat, and save Work or Codex for jobs that actually need file access, tool use, or multi-step execution.

Image generation is similar.

Generating an image in regular Chat is accounted for differently from generating an image as part of a Codex task. If all you need is one standalone image, generating it inside Codex can unnecessarily consume your shared Work/Codex allowance.

On the other hand, if image generation is part of a larger multi-step job, keeping everything inside Work or Codex may still be more convenient.

2. Don’t confuse reasoning effort with speed

The Effort setting in Work and Codex controls how deeply the model reasons about a task.

You do not need to start every task at the maximum level.

For clearly scoped editing or format conversion, Light may be enough. For typical planning and analysis, try Medium. For problems that require comparing multiple sources, debugging complex behavior, or validating a difficult decision, High or Extra High may be a better starting point.

Higher effort can take longer and may consume more of your allowance.

Fast mode is different. It does not make the model think more deeply—it makes the job run faster.

For GPT-6 Astra, Fast mode currently carries a 2.5× credit rate relative to standard speed, so it makes the most sense when turnaround time really matters.

3. Even “good prompts” have a shelf life

Take a look at the instructions, skills, and project rules you added months ago.

Rules such as:

“Always read every document before starting.”
“After every change, review the entire project again.”

may now be triggering unnecessary work even for simple tasks.

AI models improve quickly. Prompt patterns and skills that were genuinely useful a few months ago can become redundant—or even get in the way of a more capable model.

Instead of:

Always read everything first.

a better instruction may now be:

Read the material relevant to this task.

As models become more capable, the value of a universal “magic prompt” decreases. It becomes more useful to clearly explain what you want done, what the scope is, what must not change, and what counts as finished.

4. Don’t upgrade the moment you hit a limit

Before paying for a higher tier, first check which limit you actually reached and when it resets.

For Work and Codex, you can see your remaining allowance and reset timing under Settings → Usage.

Running out of image generation is not the same problem as exhausting your shared Work/Codex allowance, so the right solution depends on what you actually used up.

OpenAI also occasionally runs promotions or events that reset Work/Codex usage or provide reset benefits that can be used later.

Eligibility and expiration dates vary, so check both official announcements and your Usage screen to see whether any reset benefit is already available on your account.