AI coding tools for founders (2026): Claude Code vs Cursor vs Codex, app builders and what they really cost

The best AI coding tool for a founder depends on one question: can you read code? If you can, use a coding agent such as Claude Code, OpenAI's Codex or Cursor, which edit your files and run commands while you review the changes. If you can't, use a browser app builder such as Lovable, Bolt, Replit or v0, which turn plain-language prompts into a hosted app. Code-completion assistants such as GitHub Copilot sit in between: cheap, fast and best for people already writing code by hand.

Every price below comes from the vendor's own pricing page, checked on 25 October 2026. This site isn't affiliated with any of these companies, and prices and limits in this category change often, so click through before you buy.

The three kinds of AI coding tool

There are three kinds of tool, and they suit different founders.

• Coding agents (Claude Code, Codex, Cursor, Windsurf, Jules). You give a task, the agent reads your codebase, writes changes and runs commands or tests. Claude Code, for example, runs in a terminal, IDE extensions, a desktop app and the web, according to Anthropic's docs. These fit founders who can read a diff, even if they're rusty.
• AI app builders (Lovable, Bolt, Replit, v0). You chat in a browser and get a working app with hosting. Lovable describes itself on its pricing page as a way to build websites and web apps "with no technical knowledge needed". These fit non-technical founders testing an idea.
• Code-completion assistants (GitHub Copilot's inline suggestions, plus the tab completions in Windsurf and Cursor). They predict the next lines as you type. These fit people writing code themselves who want speed, not delegation.

The lines blur. Copilot now has agents and a CLI, and Replit and Lovable both run agents under the hood. Pick by how you'll work, not by the label.

AI coding tools compared: prices and limits

The table shows the cheapest paid plan that includes the coding product, in US dollars, as each vendor's page showed it on 25 October 2026.

A few notes on reading this table:

• Claude: Pro includes Claude Code, Free doesn't, and Max starts "from US$100" a month, per Claude's pricing page. Every plan has limits that reset on a rolling five-hour window.
• Codex: The ChatGPT pricing page lists "Limited Codex access" on Free and on Go (US$8 a month), "Expanded Codex usage" on Plus at US$20, and Pro "from US$100". OpenAI's Codex pricing page gives rough local-message ranges per five hours by model, and says Pro plans currently have no five-hour limit.
• Cursor: The pricing page looked unusual on 25 October 2026. It listed Free, Plus (US$8), Premium (US$30), Super (US$100) and Ultra (US$200) next to X features such as a "Premium checkmark" and ad-free scrolling. Cursor's own docs explain that Cursor "is sold as part of The X Subscription, one subscription for Grok, Cursor, and X". Our reading: Plus at US$8 covers Cursor's own models only (Grok and Composer), and you need Premium at US$30 or above to use Claude, GPT or Gemini, which are then billed at API list price from prepaid credit on top of the subscription. Annual billing was US$80 a year for Plus and US$300 for Premium.
• Windsurf: windsurf.com/pricing now carries Devin branding. It showed Pro at US$20 a month with "increased quotas" and extra usage at API pricing, and a free plan with unlimited tab completions.
• Jules: Google's Jules limits page gives the free tier 15 tasks a day and 3 concurrent tasks, and says paid tiers come through Google AI plans, currently only for personal @gmail.com accounts. Google One listed Google AI Pro at US$19.99 a month when we checked.
• Lovable: The plan cards didn't load as text on the pricing page, so the price comes from Lovable's subscription docs: Pro is US$25 a month for 100 credits, or US$250 a year, and Business starts at US$50 a month.
• Replit: The Replit pricing page showed Core at US$20 a month, or US$18 a month billed yearly, including US$20 of usage towards more powerful models. Pro was US$100 a month.
• v0: The v0 page showed Plus at US$30 per user a month with US$90 struck through. Free is capped at 7 messages a day.

Most of these pages say their prices exclude local taxes such as GST or VAT.

Claude Code vs Cursor vs Codex. All three edit files and run commands, so the choice comes down to where you like to work and which models you want. Claude Code and Codex both start at US$20 a month and work from the terminal, an IDE, a desktop app or the web. Cursor is an editor first, and on the plan structure above you pay for Claude or GPT models on top of the subscription. Give each shortlisted tool the same small task on your own project and keep the one whose changes you find easiest to review.

How usage limits work and how to control the cost

Usage limits, not the sticker price, decide what you'll actually pay. Every tool meters the model's work somehow, and long sessions on a large codebase burn through any allowance fastest.

The four common models:

1. Rolling time windows. Claude plans reset usage on a rolling five-hour window, and Claude Code also shows "You've hit your weekly limit" messages, per its cost docs. Codex uses a five-hour estimate on Plus, with weekly limits that "may also apply".
2. Credits in dollars. GitHub Copilot meters in GitHub AI Credits worth US$0.01 each. Pro includes 1,000 base credits plus a 500-credit flex allotment, Pro+ (US$39) includes 7,000 in total, and unused credits are forfeited when the allowance resets on the first of each month. GitHub's docs now list request-based "premium requests" billing as legacy.
3. Build credits per message. Lovable charges 1 credit per Plan mode message and a variable amount in Build mode. Its pricing FAQ gives examples: 0.50 credits to make a button grey, 1.20 to add sign-up and login.
4. Tokens. Bolt counts tokens, and its pricing page warns that most usage comes from syncing your project's files to the AI, so "the larger the project, the more tokens used per message".

For a budget anchor, Anthropic says Claude Code averages around US$13 per developer per active day across enterprise deployments, or US$150 to US$250 a month, with 90% of users staying below US$30 per active day (Claude Code costs). That's for professional developers using it all day. A founder doing a few sessions a week will usually sit inside a US$20 plan's limits.

How to keep costs down, using the vendors' own advice:

• Clear between tasks. Claude Code's docs say stale context "wastes tokens on every subsequent message", so use /clear when you switch to unrelated work, and /compact to summarise a long session.
• Match the model to the job. Anthropic recommends Sonnet for most coding tasks and Opus for complex architecture. OpenAI suggests a smaller model or tighter task scope if usage runs high.
• Trim context. OpenAI's Codex pricing page suggests smaller prompts, fewer source files, a smaller AGENTS.md and fewer MCP servers, because each server adds context to every message.
• Watch the meter. Run /usage in Claude Code or /status in the Codex CLI. On Claude Pro and Max you can set a monthly spend limit for usage credits that let you keep working past your plan limit.
• Cap overage. GitHub lets you set a dollar budget for extra AI credits. Cursor's docs mention spending caps for usage beyond the pool. Lovable lets workspace owners set per-member monthly credit limits.

Pay monthly for the first two months, see where you land, then consider annual billing.

Security and code ownership basics

Treat an AI coding tool like a fast contractor with access to your laptop: useful, but you still check its work and keep your keys away from it.

Secrets. Keep API keys, database passwords and payment credentials in environment files that are excluded from version control, and never paste them into a prompt. If an agent needs to call a service, give it a test key with limited permissions.

Permissions. Coding agents can run commands on your machine. Claude Code's security docs describe a permission system where, in Manual mode, it starts read-only and asks before editing files or running commands, and they state plainly: "You're responsible for reviewing proposed code and commands for safety before approval." Commands that fetch from the web, such as curl, aren't auto-approved by default, which helps against prompt injection.

Review every diff. Read what changed before you accept it, and run the tests. If you can't follow a change, ask the agent to explain it line by line, or don't merge it.

Training on your code. Only what each vendor's page said on 25 October 2026:

• Anthropic: For Free, Pro and Max accounts, including Claude Code use, Anthropic trains new models when the user's setting is on. For Team, Enterprise and API it "does not train generative models using code or prompts sent to Claude Code under commercial terms" unless the customer opts in (Claude Code data usage). Consumer data is kept for 5 years if you allow training, or 30 days if you don't.
• OpenAI: The ChatGPT pricing page marks Free, Go, Plus and Pro as "Content is used to train our models" with an opt-out available. The Codex pricing page ticks "No training on API or business data by default" for Business, Enterprise and API key use, but not for Plus or Pro.
• GitHub: The Copilot Trust Center says GitHub doesn't use Business or Enterprise data to train models, but may use Individual subscribers' prompts, suggestions and code snippets for training unless they opt out in settings.
• Cursor: Its security page says that with Privacy Mode on, "we will not train on your data", and Privacy Mode is available to anyone.
• v0: The v0 pricing page lists "Training opt-out by default" on Business and "Your data is never used for training" on Enterprise, which implies the cheaper plans aren't opted out by default.

Ownership. Lovable's pricing FAQ says you own your code, apps and AI output, "subject to any third-party rights in the underlying AI models". GitHub's FAQ notes that Copilot output can, in rare cases, resemble code its model was trained on, and that admins can enable a filter that suppresses long matches with public code. For anything you plan to sell or raise money on, read the terms of the tool you actually use.

A starter workflow for shipping your first product

The workflow that works is the same one good developers use: write it down, cut it small, test it, review it, ship it.

1. Write a one-page spec. Who the user is, the three things they must be able to do, and what's out of scope. Paste it into the project as a file the agent can read.
2. Prototype in an app builder or plan mode. Use Lovable, Bolt, Replit or v0 to get a clickable version in front of five potential customers before you write serious code. Our guide to getting your first customers without ads covers finding them.
3. Break the build into small tasks. One feature or fix per prompt. Small diffs are cheaper on usage limits and far easier to review.
4. Ask for tests with every task. Tell the agent to write a test first, then make it pass. Run the tests yourself.
5. Review before you commit. Read the diff, check no secrets crept in, and commit each working step to Git so you can roll back.
6. Deploy early and small. Use the builder's hosting or a standard host, with a staging copy for testing. Add payments only after a review (see merchant of record vs Stripe).
7. Measure and repeat. Watch what users do, then go back to step 3.

When to hire a developer anyway. Bring one in before you take payments, store personal or health data, or sign a customer who'll ask about security. Also hire when you've spent more than a few sessions stuck on the same bug, or when your app builder's credit spend is climbing faster than your revenue. A part-time contractor reviewing your agent's work is often enough at first; our guide to hiring remote employees internationally covers contractors versus employer of record. For the wider picture of running a company with agents, see how solo founders use AI agents.

To compare notes on which tools are working with other remote founders, request an invite to Remote Builders.

Common mistakes

These are the mistakes that cost founders the most time and money.

• Choosing on sticker price. A US$8 plan that excludes the model you want, or a US$20 plan you exhaust by Tuesday, costs more than it looks. Read what limits usage.
• One giant prompt. "Build me Airbnb for dogs" produces a large, untestable codebase. Small tasks give better code and use fewer credits.
• Never clearing context. Long sessions re-send history with every message, which is why vendors tell you to clear or compact.
• Accepting diffs unread. Agents make confident mistakes. If you can't explain a change, don't ship it.
• Pasting secrets into prompts or committing .env files.
• Assuming your plan doesn't train on your code. On most individual plans the default allows it unless you opt out.
• Staying in the app builder too long. Builders are great for validation. Once the product is the business, make sure you can export or own the code and have someone who can maintain it.
• No Git history. Without commits you can't undo a bad agent run.

Checklist before you pay for an AI coding tool

1. Decide which kind you need: coding agent, app builder or completion assistant.
2. Check today's price on the vendor's own pricing page, including tax and annual versus monthly billing.
3. Find out what limits usage (time windows, credits, tokens or tasks) and what happens when you hit the limit.
4. Confirm the models you want are included and not billed separately.
5. Check the training setting for your plan and opt out if you need to.
6. Set a spend limit or budget for overage before your first long session.
7. Put your project in Git and keep secrets out of the repository.
8. Use a permission mode that asks before running commands until you trust the setup.
9. Write a one-page spec and break it into small, testable tasks.
10. Review every diff and run the tests before you commit or deploy.
11. Book a developer review before you take payments or store customer data.
12. Recheck prices and limits every few months, because they change often.