Start with the distinction most of this results page blurs: Claude Code is free software — what costs money is the model inference behind it (its docs). So "without a subscription" does not mean "without paying"; it means paying per token instead of per month, which Claude Code supports natively through two environment variables. There is no configuration in which frontier-model inference is free and unlimited, whatever the video titles on this search say — and the routes that promise it each carry a cost that is real but not denominated in dollars.
Bias declared: Kunavo is one of the metered options. It is also not the recommendation for everybody — if you code daily and saturate a plan, a subscription is cheaper and this page says so at the end.
The routes, and what each actually costs
| Route | Money cost | The cost that is not money |
|---|---|---|
| Metered API key (Anthropic or a gateway) | Per token, scales to zero | No ceiling — spend tracks use |
| Point it at a cheaper or open model | Low to none | Quality drop larger than benchmarks imply |
| Third-party "free" proxy | None advertised | An unidentified operator sees all your code |
| Shared or purchased credentials | Low | Terms violation; revocable without recourse |
| Claude Pro / Max plan | Flat monthly | Usage windows; no unattended runs |
The legitimate route: a metered key
This is the answer for most people who do not want a subscription, and it is two lines. The base URL is an origin — no /v1 suffix, because Claude Code appends the path itself:
# Claude Code itself is free software. These two variables point it
# at a metered API key instead of a plan. The base URL is an ORIGIN —
# no /v1 suffix.
export ANTHROPIC_BASE_URL=https://api.kunavo.com
export ANTHROPIC_AUTH_TOKEN=sk-kn-...
claude
# Anything you unset goes back to the plan:
# unset ANTHROPIC_BASE_URL ANTHROPIC_AUTH_TOKENThe key can be an Anthropic API key or one from any gateway implementing the Messages API. Kunavo lists Claude Sonnet 5 at $2.00 / $10.00 per 1M against Anthropic's $2.00 / $10.00 — at list on this model today, and Claude Haiku 4.5 at $0.40 / $2.00 for the routine calls. Setup detail is on the Claude Code integration page; which of the three sources to pick is a separate decision.
What actually determines the bill is not the vendor, it is caching and model tier. Claude Code re-sends a long system prompt and file context on every turn, so uncached input dominates the invoice — cached input is $0.20 per 1M against $2.00 fresh. A source without prompt caching can cost several times more for identical work. The full set of levers on a Claude bill puts them in order of size.
The "free" routes, priced honestly
These are described as categories, because the specific projects circulating on this query change monthly and we have not audited any of them. Judge a given one by which category it falls into and what it asks you for.
- Substituting a cheaper or open model. Legitimate, and the only free-ish route with no ethical or security problem. The real cost is quality: Claude Code is not model-agnostic in practice — its prompts, tool patterns and editing behaviour are tuned against Claude, so a substitute that benchmarks close can still be noticeably worse at editing your actual repository. Evaluate it on real work, not a demo. Routers exist for this; Claude Code Router is the common one.
- A third-party proxy someone else operates. The operator sees every request Claude Code makes — your source code, your configuration, whatever else is in the context window. That is the price, and it is only acceptable if you know who they are and would have sent them the code anyway.
- Shared or purchased account credentials. Using someone else's account violates the provider's terms, the access can be cut at any time with no recourse and no refund, and you cannot know how the account was funded. Marketplace listings selling keys cheaply are cheap for a reason.
When the subscription is still the right answer
If you use Claude Code interactively most working days, one session at a time, and regularly reach the usage window, a plan is cheaper per unit of work than metered billing — often substantially, which is what the "API costs many times more" posts are really measuring. The metered route wins on bursty months, unattended and CI runs, concurrent sessions, work that must continue after a window is spent, and per-project spend records.
These are not exclusive, and that is the practical ending: Claude Code switches on two environment variables, so keeping a plan for daily interactive work and a key for everything that happens when the window is spent is a normal arrangement rather than a compromise. cc-switch manages the toggle, and what the plans actually include covers the subscription side.
FAQ
Can you use Claude Code without a subscription?
Yes. Claude Code is free software — what costs money is the model inference behind it. Set ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN and it bills per token against an API key instead of drawing on a plan. That key can come from Anthropic directly or from a gateway that implements the Messages API.
Is Claude Code free?
The program is free to install and run; the model calls it makes are not. This is the distinction most "free Claude Code" content blurs. Every route that keeps the tool genuinely useful pays for inference somewhere — through a Claude plan, through a metered API key, or by pointing it at a cheaper or free model and accepting the quality difference. There is no configuration in which frontier-model inference is free and unlimited.
How much does Claude Code cost per month on an API key?
It scales with use rather than being a fixed number, which is the point of the metered route. Two things dominate the bill: the model tier, and whether prompt caching is on — Claude Code re-sends a long system prompt and file context every turn, and cached input is billed far below fresh input ($0.20 against $2.00 per 1M on Claude Sonnet 5 on Kunavo). Light or bursty use typically lands well under a monthly plan; sustained daily use with a frontier model does not.
Are the free Claude Code proxies safe?
Judge them by what they require rather than by the claim. A third-party proxy sees every request Claude Code sends, which includes your source code and anything else in the context window — so an operator you cannot identify is the actual cost of that route. A setup that asks you to use shared or purchased account credentials violates the provider's terms and the access can be revoked at any time with no recourse. A setup that swaps in a different model is legitimate but is not the same product: Claude Code's prompts are tuned for Claude, and substituting a weaker model degrades results more than the model's own benchmarks suggest.
Can Claude Code use a free or open-source model?
Technically yes, through anything that presents an Anthropic-Messages-compatible endpoint, and there are routers built for exactly this. The honest caveat is that Claude Code is not model-agnostic in practice — its system prompts, tool-use patterns and editing behaviour are tuned against Claude models, so a substitute that scores close on public benchmarks can still perform noticeably worse at the actual task of editing your repository. Try it on real work before committing, not on a demo.
When is a subscription still the better choice?
When you use Claude Code interactively most working days, one session at a time, and regularly reach the usage window. In that pattern a flat monthly fee is cheaper per unit of work than metered billing, often by a lot. The metered route wins for bursty use, unattended and CI runs, more than one concurrent session, work that must continue after a window is spent, and per-project spend records — and the two can coexist, because switching is two environment variables.