Cheap ways to code with AI (right now).

One of the questions we get asked most is how to try AI coding agents without spending much money. The answer changes month to month, because prices move, free tiers shrink, and tools shut down with little warning.

This section is a snapshot of the current situation, but as you will read it's changed a lot recently.

The brain and the body

A coding agent is really two things. The harness is the software that edits your files and runs commands (the body), and the model is the brain it phones up for every decision.

The body is essentially free, and several good ones are open source. The brain is where the money goes, because at raw API rates a full day of agentic coding on Claude Sonnet typically costs $5 to $15, an Opus day $15 to $40, and heavy users often report $200 to $500 a month.

A subscription brings that cost down a lot. Claude Code has no free tier, Pro is $20 a month, and the Max plans at $100 or $200 include far more usage than the same money buys in raw tokens.

Importantly, since April Anthropic no longer allows Pro and Max subscriptions inside third-party tools like OpenCode, so if you already pay for a Claude subscription, Claude Code is your cheapest option almost by definition.

Why so much advice is already out of date

Between 15 May and 18 June, three well-known budget options disappeared.

• Roo Code, a VS Code extension with over 3 million downloads, shut down on 15 May. Its own docs now point users to Cline or the community fork ZooCode.
• Continue.dev was absorbed into Cursor in mid-June. The repository is read-only and user cloud data was deleted after 15 July.
• Gemini CLI's free tier ended on 18 June. Users were moved to Google's Antigravity CLI, where community testing puts the free quota in the low tens of requests per day, down from roughly 1,000.

Another couple of budget options seem to be dwindling. Aider has not shipped a release since February, and Alibaba closed Qwen's free coding tier in April, replacing it with a $50-a-month plan.

What is genuinely free today

• Codex on a free ChatGPT account. This is the best free option we're aware of right now. OpenAI includes Codex on every plan, even Free. The allowance is unpublished and the smallest of any plan, enough for small jobs and learning the workflow rather than a day of building. Cloud features like automatic code review start on the $20 Plus plan.
• GitHub Copilot Free. 2,000 completions and 50 chat requests a month, roughly two or three questions a day. A decent companion to the Codex free tier.
• Kiro's free tier. Amazon's agentic IDE gives free users 50 credits a month with Claude Sonnet 4.5. Startups can still apply for up to a year of free Pro+.
• MonkeyCode. Open source and self-hostable, with a hosted free plan reportedly around 30 million tokens a day (we could not confirm that on an official page). Unless you self-host, your prompts go through a Chinese-hosted service.
• OpenCode. A free, open-source harness you plug a brain into, with a rotating set of free hosted models built in (no card needed). The free models change without notice and your prompts may be used for training, so treat it as a playground rather than something to rely on.
• Fully local with Ollama. Point Claude Code's base URL at Ollama and run open models on your own machine, free forever. Agents need at least a 32K context window though, which most laptops struggle with.

The cheap tier

• GitHub Copilot Pro, $10 a month. Unlimited completions plus roughly 1,000 AI credits. The cheapest paid agent, from a company as stable as this space gets.
• GLM Coding Plan from Z.ai, Lite at $18 a month. Plugs into Claude Code, Cline and 20 or so other tools as a cheaper brain. Zhipu's own benchmark puts GLM-5.1 at 94.6% of Claude Opus 4.6, a vendor number that broadly matches real-world experience.
• Kimi Code, $19 a month. Moonshot's rival to the GLM plan. The subscription plugs into Claude Code, OpenCode and Codex with one settings change and runs the Kimi models, though the plan lineup is currently in flux, so check the price on the day.
• ChatGPT Plus, $20 a month. Codex everywhere, plus cloud integrations. Enough for a few focused sessions a week, though OpenAI's rate card says heavy users average $100 to $200 a month.
• Kimchi, free to $20 a month. An open-source agent that routes each subtask to a different open model. The free plan includes a 250K context window, and the $20 Coder plan includes $25 of credits.

The reality check

None of this stays still for long. Free tiers exist so companies can collect feedback and training data, and they get cut when that purpose is served.

OpenCode's free models rotate without announcements, and Google cut its CLI free tier by roughly 50 times. These offers are all genuine, they just change often, so check the price on the vendor's own page the day you sign up.

What this means for you

If you already pay for Claude or ChatGPT, use the agent that comes with it and ignore the noise. If you pay for nothing, start with the Codex free tier plus Copilot Free, and if you are paying API rates, Copilot Pro or a GLM plan is the cheapest option that handles daily work.

Whichever you pick, keep your instructions in a CLAUDE.md or AGENTS.md file in the project. That way moving between tools takes an afternoon rather than a rebuild, and given the past three months, a little portability is worth having.

And if you want to see what building with these agents looks like in practice, James just wrote up the exact stack he uses to build iOS apps with Claude and Codex, which he's used to ship two production apps without writing the code himself.

One more thing that made us laugh this week. Builders have become so fed up with token bills that a Claude Code skill called Caveman has picked up over 95,000 stars on GitHub. It cuts token use by around 65% by making the AI talk like a caveman ("why use many token when few token do trick"), and the answers are still pretty good.