AI for programming: what a developer should buy

Three paths: where exactly the AI writes code

The question "which AI is best for programming" has no answer until it is clear where it will work. For a developer, AI lives in three places, and for each place we have a service of our own.

An editor with AI. This is Cursor: a regular editor based on VS Code in which the AI sees the whole project, completes the line while you type, and edits several files on one command. You stay at the wheel — you look at every diff and accept or revert it. This mode suits daily work on a live project where understanding every line matters. Details on the plans are on the Cursor page.

An agent in the terminal. This is Claude Code and Codex. You hand the agent the whole task: "add caching to this service and update the tests", and it reads the repository itself, edits files, runs checks and brings back the result. You step in at the end, not at every step. Claude Code is included in the Claude Pro subscription, Codex — in ChatGPT Plus and Pro. No separate API key is needed for this.

A chat for analysing code and architecture. This is the same ChatGPT, Claude or Gemini in a browser window: paste a function and ask why it crashes; discuss a database schema; get a draft module from a description. The chat is included in all subscriptions from the first two paths, so only those who don't need an agent or an editor yet buy it on its own: students and people who are learning. Gemini Pro stands apart in this row because of its context: a medium-sized repository fits into a single request, which is more convenient for "how is this project structured" questions than feeding the chat files one at a time.

Which path for whom

• You write code every day in one or two projects — the editor. Cursor Pro covers a working day; Pro+ is needed when the agent in the editor works in the background for hours and the monthly request pool runs out by mid-month.
• You do a lot of similar tasks: migrations, tests, refactoring, moving between frameworks — the agent. Claude Code behaves more carefully in someone else's code and rewrites less than needed less often; Codex is stronger where parallel work on several tasks in the cloud is needed. Developers who have tried both usually keep the one whose editing style is closer to their own.
• You are learning or write code once a week — the chat. ChatGPT GO or Gemini Plus: explanations, error analysis, small functions. An agent is harmful at this stage: it will write faster than you can understand.
• You lead a team — one set for everyone. Different tools for different people mean different prompts, different rules in the repository and arguments in reviews. Setting up for several people is covered on the AI for business page.

How Cursor differs from GitHub Copilot

We don't sell Copilot, but the question comes up in every other consultation, so here is a straight answer.

Copilot is an add-on to your editor: it lives inside VS Code, JetBrains or Neovim and suggests lines as you type. It has gained an agent mode, but it remains an overlay on someone else's editor, and the depth of access to the project is limited by what the extension allows.

Cursor is the editor itself. It is a fork of VS Code with AI built in at the core level: it indexes the entire codebase, sees the open tabs and the terminal, edits several files on one command and shows the changes as a diff. VS Code extensions work in it, so moving over takes an evening: settings, themes and key bindings are imported.

The practical difference is that Cursor gives you a choice of model. Inside one editor you can switch between Anthropic, OpenAI and Google models depending on the task: one writes new code better, another edits old code more carefully. Copilot runs on the models GitHub has chosen. If you are used to one specific model and want it in your editor, Cursor is the only one of the two where that can be configured.

On price, Copilot's entry plan is cheaper than Cursor Pro, and for those who only need line suggestions it is enough. But as soon as the task becomes "do it in the project" rather than "finish the line", the price difference stops being decisive: an hour of a developer's work costs more than the difference between the plans for a month.

Why developers buy two services at once

This is not marketing but an observation from orders: developers are the only audience in the AI section that regularly sets up two subscriptions at the same time. The reason is simple: the editor and the agent solve different tasks, and no single service does both well.

In the editor, response speed and precision matter: you edit a function and want to see the result in a second. For the agent in the terminal, stamina matters: it can work on a task for ten minutes, run the tests and fix what it broke. If you try to do the second in the editor, you sit watching a progress bar; if you do the first in the terminal, you wait a full cycle for every line.

The most common combination is Cursor Pro + Claude Pro. Cursor for hands-on work, Claude Code for "do the whole thing" tasks, and the Claude chat on top for analysis. The second most common is Cursor Pro + ChatGPT Plus: the same logic, but the agent is Codex and the chat is ChatGPT, which is more convenient for those who also need images, voice and search. Today's total for each set is in the "Typical sets" block above: it is recalculated from the exchange rate together with the prices in the table and matches what the manager will quote.

A set is ordered as one order. Message @timurjob1 which two services you need, pay one amount, and hand over access one at a time — first to one account, then to the other. This is more convenient than two separate conversations, and the manager sees right away that the order is a combined one.

Limits and what to do when you hit them

No developer subscription is unlimited. This is important to understand before buying, not at the moment the agent stops in the middle of a task.

Cursor Pro and Pro+. The limit is a monthly pool of requests to the models. Autocomplete is not counted and works without restrictions, but every request to the agent or the chat uses it up. The Pro pool is designed for a normal working day; if the agent works in the background for hours, the pool may run out before the end of the month. The solution is Pro+, where the pool is several times larger. You can switch at renewal: tell the manager which plan you need next month.

Claude Pro. The limit is rolling: usage is counted in a five-hour window, the window shifts, and after a pause the limit recovers. On top there is a weekly ceiling. In practice this means that with heavy Claude Code use, after two or three hours the agent will ask you to wait. We don't sell the Max plan with limits several times higher — we say so up front so you don't expect it from us. If Pro falls short every day, the working way out is to add a second agent: ChatGPT Plus with Codex. Two agents on separate limits add up to more working hours than one Pro and cost about what Max would.

ChatGPT Plus, Pro ×5 and Pro ×20. Codex also has a limit in five-hour windows plus a weekly one. Plus is enough for several tasks a day; if you hit the limit regularly, the next step is Pro ×5, an intermediate option with noticeably larger limits, and Pro ×20 — full Pro, where the limits are many times higher. How to tell that Plus is not enough is in the article ChatGPT Plus or Pro.

Gemini Pro. The limits are designed for a full working day; people rarely hit them. A separate feature is the context size: it is large, but long requests use up the limit faster than short ones.

The general rule: if the limit ran out once in a month, do nothing and wait. If it runs out every week — the next plan tier. If every day — a second service, not a higher tier of the same one.

How the setup works for each service

The scheme depends on the service, and that is normal: it is how the platforms themselves are built, not our rules.

1. Cursor — on your account. You hand over the login and password for your Cursor account, the manager sets up the subscription, and in 15–30 minutes it is active. Projects, editor settings and codebase indexes stay in place. After activation, change your password — the subscription won't drop.
2. Claude — on your account. The same scheme: access for the duration of the setup, then a password change. If you don't have an account yet, the manager will explain how to register — Claude has some quirks with phone numbers, which are covered in the article Claude where access is restricted. Claude Code switches on by itself: install it and sign in with the same account.
3. ChatGPT — your choice: invite link or account login. The invite link doesn't require handing over a password: you receive an invitation and accept it in your own account. Account login is the classic scheme with access handed over for the duration of the setup. The prices of the two schemes differ slightly, both are on the ChatGPT page; if you have two-factor protection enabled, the invite link is simpler.
4. Gemini — to the client's email. The subscription is tied to a Google account, so either the manager registers a new account and hands it over to you in full, or sets it up on yours — then access to it will be needed. If you don't want to give access to your personal mailbox, choose a new account.

Payment is the same in every case: cryptocurrency (USDT). You don't need a foreign card of your own, and you don't need to create a foreign account for payment. The manager confirms receipt in the chat and takes the order into work right away.

Working where access is restricted: where you need a VPN

The setup goes through without a VPN everywhere — that is our side. Usage depends on the service, and here we are straightforward.

• ChatGPT and Claude — in countries where OpenAI and Anthropic restrict access, the sites and apps don't open without a VPN. This applies to the chat, to Codex and to Claude Code: the terminal agent talks to the same servers as the website. You need a stable VPN for the whole working session; in the editor and terminal a system-level VPN works better than a browser extension. Details in the article do you need a VPN for ChatGPT.
• Cursor — the editor itself and autocomplete work from restricted regions, but requests to the models go through Cursor's servers, and for some models a VPN may be needed. It takes a minute to check: if the chat in the editor answers, everything is fine.
• Gemini — in countries where Google restricts access, the site is unavailable without a VPN. This applies to the chat and to Gemini in Docs.

If a VPN is a problem, the most tolerant path is Cursor: the main work happens locally, and it goes online only for the model's reply.

Renewal and guarantee

All developer subscriptions are set up for a month; these services have no annual options in our catalogue. Renewal works the same way: message the manager a couple of days before the end of the term and name the service and plan. At renewal you can change the plan: Pro to Pro+, Plus to Pro ×5, or the other way round if the limits are more than enough. There are no automatic charges — the subscription won't renew itself or take money without your decision.

The guarantee is valid for the entire paid period: if the subscription drops, we restore it or refund the money. It doesn't apply in one case — the account itself being blocked by the platform's decision. That is not in our hands; we'll help you write to support, but we can't promise the outcome. Reducing the risk is simple: don't hand the account to third parties, don't change the country in the profile after the setup, and use one VPN region rather than jumping between countries. More on the terms on the guarantees page.

What we don't sell

• API keys and API payment — no. The subscription covers Claude Code and Codex without an API key: both agents work on the subscription with the limits from the table above. The API is only needed by those who embed the model into their own product, and that is a different task and different money.
• Claude Max — no, only Pro. What to do when the limits run short is above.
• GitHub Copilot — no, only the comparison.
• ChatGPT Team and Claude Team team plans as a product — no; on request through the manager, see the for business page.

If you are torn between two paths, describe to the manager what exactly you do in a day: the language, the size of the project, how many hours in code. Picking a plan from such a description is easier than from a table.