Goose is free, Claude Code is two hundred a month, and the comparison everyone is making is wrong

There is a line going round at the moment that goes something like this: Claude Code costs you up to two hundred a month, Goose does the same job for free, so why are you paying. I pay that bill, so I went and actually worked out whether it is true. It is not, and the reason it is not tells you something useful about where the money in this whole field actually goes.

The short version: "free" software that drives a frontier model is not free, because you still pay for every token the model burns, and an agent burns a frightening number of them. What you are really buying when you pay a subscription is not a chat box. It is autonomous compute and deep tool integration. Once you see that, the free-versus-paid argument turns into a much more interesting question about how you work, not how stingy you are.

What Goose actually is

First, credit where it is due, because Goose is genuinely good and I am not here to talk you out of it. Goose is an open-source, local AI agent originally built by Block, the company behind Square and Cash App [1]. It runs on your own machine as a desktop app or a command-line tool, it is licensed under Apache 2.0, and it is properly free as in software: you can use it commercially, modify it, ship it [1]. In December 2025 Block handed it to the Linux Foundation's new Agentic AI Foundation, alongside Anthropic's Model Context Protocol and OpenAI's agent standards, so it is now a neutral, foundation-governed project backed by most of the industry rather than a single vendor's pet [2][3].

The clever bit is the architecture. Goose does not ship a model. You bring your own, and it treats the model as a configuration choice [1]. On first run it walks you through picking a provider, and it supports a long and growing list of them: Anthropic, OpenAI, Google, local models through Ollama, and many more [1]. It speaks MCP natively, which is how it plugs into tools and systems, and its capabilities extend through MCP-based extensions [1]. It is a real agent, not an autocomplete: it reads and writes files, runs commands, executes tests, installs dependencies, and loops on its own output until the job is done [4].

If you run it against a local model through Ollama, nothing leaves your machine, which for sensitive or air-gapped work is a real advantage no hosted product can match [1]. The vendor neutrality is the other genuine win. You can swap Claude for a cheaper model, or a local one, without changing the agent you have learned. That is not nothing.

So why am I still paying Anthropic.

The bit the "free" framing skips

Here is the sleight of hand. Goose is free. The model it drives is not. The moment you point Goose at Claude or GPT through an API key, you are paying that provider by the token, and agentic work is the most token-hungry way to use a model that exists.

The reason is structural. An agent does not send one prompt and get one answer. It runs a loop, and on every turn of that loop it resends the entire accumulated context: the files it has read, the commands it has run, the history of what it has tried [5]. A single involved coding session can run to a million input tokens against forty thousand output [5]. Input dwarfs output by a factor of twenty or more, and you pay for all of it [5].

Put real numbers on it. Claude Opus, the model you actually want for hard agentic work, runs at roughly five dollars per million input tokens and twenty-five per million output on the API [6]. An hour of continuous agent runtime lands somewhere around five to fifteen dollars at those rates, and people running these tools hard all day routinely report monthly API bills of four hundred to fifteen hundred dollars, with extreme cases well past that [5]. None of that is a Goose problem. It is the cost of the compute, and it does not care which agent wrapped it.

And here is the detail that gives the whole game away. Anthropic literally meters autonomous compute as a separate line item. On its managed agent pricing you pay for tokens, and then you pay an additional eight cents per session-hour just for the agent to be running [6]. They are charging for the run, not the chat. The chat box is the cheap part. The expensive part is the thing thinking and iterating on your behalf, and that is true no matter whose software is in front of it.

The real cost comparison

So let me lay it out honestly, because the honest version helps both arguments.

First, the subscription side, with the numbers corrected. Claude Code is bundled into Anthropic's plans: Pro at twenty dollars a month, Max at one hundred, and the big Max tier at two hundred [7][8]. That top tier is where "two hundred a month" comes from, though if you are in the UK the real figure is more like a hundred and fifty to a hundred and ninety once you account for the exchange rate and VAT, not a flat two hundred pounds. Worth getting right if you are going to quote it. The plans work on usage limits shared across Claude chat and Claude Code, not unlimited use [9].

Now the comparison that matters, by how hard you actually push it.

Read that table both ways, because both ways are true.

If you are a light or occasional user, the free route genuinely is cheaper. Bring your own key, run a mid-tier model like Sonnet, and you will spend less than any subscription, because you are simply not consuming the expensive autonomous compute the subscription is priced around [10]. If your work fits a free tier like Gemini's command-line tool, or a local model, your marginal cost really is close to zero [13]. Anyone telling you the paid plan is always better is selling something.

But if you push these tools hard, the flat subscription is the better deal by a distance, and not by a little. In heavy agentic coding, more than ninety per cent of your token spend is cache reads from long autonomous context, not the prompts you type [12]. On the API you pay for those. On a flat plan they are included. People running Claude Code all day have measured well over a thousand dollars of API-equivalent usage absorbed by a two-hundred-dollar plan in a single month [11]. At that volume the subscription is not the expensive option. It is the cap that stops the expensive option running away from you.

So the crossover is roughly this: under heavy, all-day, Opus-grade work, the subscription wins on both cost and predictability. Everywhere below that, bring your own key is cheaper, and Goose is an excellent way to do it [14].

One trap worth knowing

There is an obvious move people try here: keep my cheap Claude subscription, but drive it from free Goose instead of paying the API. As of the last few months that does not work cleanly. Using a Claude Pro or Max subscription inside third-party open-source agents is not sanctioned, and API-key billing is the only supported path for Claude models in those tools [10]. So "free tool plus my existing sub" is not the loophole it looks like. If you want Claude inside Goose, you are paying the API, and we are back to the token maths above.

What this actually tells you

Strip it all back and the lesson is not about Goose or Claude Code at all. It is that in agentic software, the interface is the cheapest thing in the room. The model is expensive, the autonomous compute is expensive, and the deep, first-party tool integration that makes a hosted agent feel effortless is expensive to build and maintain. A free wrapper removes the one part that barely cost anything. It does not remove the part that actually costs money, because that part is the model doing the work, and you rent that by the token whoever you are.

Which is why the question "free or paid" is the wrong one. The right question is how you work. If you dip in and out, or you can live on smaller and local models, Goose plus your own key is the smart, cheaper, more private choice, and you should take it. If you live in an agent all day on frontier-grade work, a flat subscription is buying you a predictable cap on a genuinely expensive resource, and "but it is free" misunderstands what you would be giving up.

I keep paying the bill because of how I work, not because I have not noticed Goose. And if you are about to switch to save two hundred a month, just check first whether you were ever going to spend two hundred a month in the first place. Most people quoting that number never were.

Sources

1. Goose on GitHub, Agentic AI Foundation - https://github.com/aaif-goose/goose
2. Linux Foundation, formation of the Agentic AI Foundation - https://www.linuxfoundation.org/press/linux-foundation-announces-the-formation-of-the-agentic-ai-foundation
3. Block, Block, Anthropic and OpenAI launch the Agentic AI Foundation - https://block.xyz/inside/block-anthropic-and-openai-launch-the-agentic-ai-foundation
4. Block, Block open source introduces codename goose - https://block.xyz/inside/block-open-source-introduces-codename-goose
5. Vantage, the cost of agentic coding - https://www.vantage.sh/blog/agentic-coding-costs
6. Anthropic, API and managed agent pricing - https://platform.claude.com/docs/en/about-claude/pricing
7. Anthropic, plans and pricing - https://claude.com/pricing
8. Anthropic, what is the Max plan - https://support.claude.com/en/articles/11049741-what-is-the-max-plan
9. Anthropic, using Claude Code with your Pro or Max plan - https://support.claude.com/en/articles/11145838-use-claude-code-with-your-pro-or-max-plan
10. Shareuhack, the real cost of running Claude through open-source agents - https://www.shareuhack.com/en/posts/openclaw-claude-code-oauth-cost
11. Product Compass, Claude Code pricing in practice - https://www.productcompass.pm/p/claude-code-pricing
12. CloudZero, Claude Code pricing and where the tokens go - https://www.cloudzero.com/blog/claude-code-pricing/
13. Open Source Alternatives, the best open-source AI coding assistants - https://www.opensourcealternatives.to/blog/best-open-source-ai-coding-assistants
14. Verdent, Claude Code pricing guide 2026 - https://www.verdent.ai/guides/claude-code-pricing-2026