A copper terminal console on a workbench, a robot arm slotting a small engine cartridge into it, a tray of other cartridges with blank price tags beside it

Claude Code free: we ran it on nine free models, and here is where each one stops

Posted by:

|

On:

|

, ,
A copper terminal console on a workbench with its hatch open, a robot arm slotting a small glowing engine cartridge into it, a tray of assorted engine cartridges with blank price tags beside it, and a larger sealed engine on a shelf behind a velvet rope

Can you use Claude Code for free? You can run the program for free, just without Claude’s models behind it. Anthropic’s free plan leaves Claude Code out, yet the tool is a free download and works with any server that speaks Anthropic’s Messages API. So we pointed Claude Code 2.1.330 at nine free models on OpenRouter and gave each one three coding tasks, graded by tests the model never saw. All nine fixed a bug and built a small feature, seven of nine got every edge case right on the third task, and the bill came to $0.00.

  • Free and real Claude: not on offer. Claude Code comes with the paid plans; Claude Code pricing shows how much work each plan holds.
  • Free through OpenRouter: works today for small, well-scoped tasks. What stops you is supply, not skill: rate limits, and a daily cap on free requests.
  • Free and local, with Ollama: documented by Ollama, not timed in this post. Claude Code’s first request alone was about 19,000 tokens, so the model needs a large context.

Is there a free version of Claude Code?

No. Anthropic’s pricing page marks Claude Code “No” on the Free plan and says it “is included in all paid plans”, which share one usage pool between chat and the terminal. Pro is the cheapest way in.

What is free is the program. Claude Code reads its endpoint from ANTHROPIC_BASE_URL, so any server that accepts Anthropic’s request format can sit behind it. OpenRouter runs one for hundreds of models, and a few of those cost nothing; their ids end in :free. On 22 November 2026, its model list held 476 models, 11 of them free, and all 11 support tool calls, which Claude Code needs to read and edit files. You keep Claude Code’s agent loop, its tools and its permission prompts. The thinking is done by someone else’s model.

How we tested Claude Code with free models

Each model got three tasks in a small invented Python repo, run with claude -p in a fresh config folder and the same tools allowed for everyone:

  • Bug fix: an invoice helper applies a 10% discount as 1000%. Three unit tests fail; fix the code, not the tests.
  • Feature: add parse_invoice(), which reads a multi-line order, skips comments and blank lines, and takes an optional VAT line. Four hidden tests grade it afterwards.
  • Edge cases: a free_slots() function that finds free time between meetings breaks on overlapping, nested and out-of-order meetings. The model gets one failing test and the docstring; seven hidden tests cover the cases a careful engineer would think of.

We skipped the two free Gemma models, which are Google’s; the other nine ran. Every request went through a small local filter that refused any model id not ending in :free, so a slip could never reach a paid model. Claude Code never asked for one: the only refusal in the log is the test we sent on purpose. OpenRouter’s usage counter for the account read the same before and after, to the ninth decimal.

Results: nine free models inside Claude Code

Scorecard of nine free OpenRouter models running Claude Code on three tasks. Bug fix: all nine pass, from 9 to 176 seconds. Feature: all nine pass hidden tests, from 13 to 636 seconds. Edge cases: seven score 7 of 7, Nemotron 3.5 Lightning scores 6 of 7, Ling 3.2 Flash gave up after 11 rate limits. Cost $0.00; 105 of 366 requests hit a rate limit

The easy tasks did not separate anyone. Every model found the discount bug and built the parser, including LFM 3 1.7B, a model with 1.7 billion parameters. The median bug fix took 19 seconds and the median feature 46.

The edge-case task separated them: Nemotron 3.5 Lightning sorted and merged the meetings correctly, then kept the old > min_len comparison, so a free slot of exactly 15 minutes vanished although the docstring says “at least”, and its own new tests passed anyway. Ling 3.2 Flash never wrote a line: every request came back rate-limited, Claude Code retried ten times and stopped with “API Error: Request rejected (429)”. The fastest model overall was ERNIE 5.5 Flash Preview, at 72 seconds for all three tasks.

At this size, the free models are good enough. The trouble is getting them to answer.

Where the free route breaks

  • Rate limits. 105 of the 366 requests Claude Code sent came back with a 429, and two models took 89 of them. Ling’s feature task took 636 seconds, almost all of it waiting. The four NVIDIA Nemotron models hit one rate limit between them in all sixteen of their runs. Pick a model with spare capacity, and move to another when one starts refusing.
  • The daily cap. OpenRouter counts free requests per day per account; ours reports a ceiling of 1,000, and accounts that never bought credits get a lower one (limits page). Our edge-case task needed a median of 11 successful requests, so 1,000 covers roughly 90 tasks of that size a day.
  • Claude Code’s own prompt. The first request of every run carried 17,600 to 19,400 input tokens, with 14 tool definitions, before the model had read a single file of yours. It is the price of the agent loop, and it is why a model with a short context fails here even when it can code.
  • Green tests are the model’s tests. Lightning’s miss passed every test it wrote. On a free model, read the diff the way you would read a stranger’s pull request.
  • It is not Claude, and Claude Code says so. Every run printed an “unrecognized model” warning. Features tied to Claude models still point at them: OpenRouter’s Claude Code guide notes that /fast on an unsupported model moves the session to your configured Opus model. Map every model slot to a :free id, or one stray command lands on a paid one.

How to use Claude Code with OpenRouter for free

Create an OpenRouter key, pick a :free model that lists tool support, and put this in ~/.claude/settings.json (or .claude/settings.local.json for one project). Leave ANTHROPIC_API_KEY empty, as OpenRouter’s guide says, so Claude Code does not fall back to an Anthropic login.

~/.claude/settings.json

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://openrouter.ai/api",
    "ANTHROPIC_AUTH_TOKEN": "sk-or-v1-your-openrouter-key",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_MODEL": "nvidia/nemotron-3.5-super-vl:free",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "nvidia/nemotron-3.5-super-vl:free",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "nvidia/nemotron-3.5-super-vl:free",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "nvidia/nemotron-3.5-super-vl:free",
    "CLAUDE_CODE_SUBAGENT_MODEL": "nvidia/nemotron-3.5-super-vl:free"
  }
}

The five model lines are the safety catch from the last section: the main model, the background model and the subagents all stay on the free id. If you were logged in to Claude Code before, run /logout first. The free list changes; when a model starts returning errors, OpenRouter’s model list shows what replaced it.

Claude Code with a local model (Ollama)

The other free route runs the model on your own machine. Ollama speaks Anthropic’s API and documents the setup: ollama launch claude, or the same three variables pointed at http://localhost:11434. For local models its Claude Code page asks for a context window of 64k or more.

We did not time this route here: the machine we tested on holds no local chat model. Expect the same 19,000-token opening prompt on every run, and work out first what your hardware can hold. The local LLM cost post compares buying a machine with renting one, and VRAM swap covers fitting a model larger than your card.

FAQ

Is Claude Code free?

The program is free to install, but Anthropic’s free plan does not include it. Claude Code with Claude models needs a paid plan or API credit. With OpenRouter’s :free models it costs nothing, and it is not running Claude.

Can I use Claude Code without a subscription?

Yes. Point ANTHROPIC_BASE_URL at OpenRouter or at a local Ollama server and set a non-Claude model. Our nine free models fixed bugs and built a small feature at $0.00; you trade Claude’s quality for rate limits and a daily request cap.

What is the best free model for Claude Code?

On our three tasks, ERNIE 5.5 Flash Preview was fastest, and the NVIDIA Nemotron models almost never hit a rate limit. The free list changes, so favour a model that lists tool support and answers without 429 errors on the day you use it.

Are OpenRouter’s free models really free?

On our account, yes: 27 tasks and nine warm-up checks, 366 requests in all, left the usage counter where it was. The limits are rate limits per minute and a daily request cap that depends on whether you have ever bought credits.

Does Claude Code work with Ollama?

Yes, according to Ollama’s own documentation, which recommends a 64k or larger context for local models. We have not timed it here.

What we changed in TinkerClaw, and why

The free route’s real failure is a model that stops answering, and plain Claude Code gives up after ten tries. TinkerClaw, our open-source build of OpenClaw, treats a rate limit as a cue to move: its router marks that supply as cooling and runs the next turn on the best model still open, and each chat’s model, from a paid plan to an OpenRouter or local one, sits behind one dial. The model router skill is the lighter version for a stock OpenClaw.

We tinker with ideas like this all the time and bake them in the moment we need one in our own day; from block to blob says why that pace is possible now. The skills worth loading into Claude Code have their own post, Claude Code skills, and choosing between paid models is in best AI for coding. The live demo lets you click through it all.

For builders

Leave a Reply

Your email address will not be published. Required fields are marked *