Claude Code free alternative: what free models replace

People who search for a Claude Code free alternative are rarely unhappy with the tool itself. They like the terminal agent, the file edits, the plan mode and the hooks. What they want to cut is the model bill, or the moment a subscription window says “You’ve hit your session limit” halfway through a refactor.

Those are two separate things. Claude Code is an agent interface: it reads files, runs commands, edits code and keeps a conversation. The model behind it does the reasoning and writes the code. You can change the second without throwing away the first, but you cannot make a free model behave like a paid Claude model by changing its name.

Where Claude Code costs come from

Anthropic’s own Claude Code cost documentation is direct about the mechanism: “Claude Code charges by API token consumption.” The same page lists subscription plans (Pro, Max, Ultra, Team, Enterprise) and per-token billing through the Claude Console or a cloud provider. It does not describe a free Claude Code tier.

The page also gives a reference point for API and enterprise usage: “the average cost is around $13 per developer per active day and $150-250 per developer per month, with costs remaining below $30 per active day for 90% of users.” Those are Anthropic’s enterprise figures, not a prediction for your account.

The token drivers it names are concrete:

Context size. “Claude Code sends your full conversation with every request,” so a one-line question at the end of an all-day session still pays for the whole history.
Model choice. Anthropic recommends Sonnet for most coding work and reserving Opus for “complex architectural decisions or multi-step reasoning.”
Extended thinking. “Thinking tokens are billed as output tokens,” and the default budget can reach tens of thousands of tokens per request.
Subagents and agent teams. Each subagent sends its own requests, and agent teams in plan mode use “approximately 7x more tokens than standard sessions.”

On subscription plans, the ceiling is a usage window rather than a dollar figure. Team and Enterprise seats reset on “a rolling five-hour window and a weekly window,” and a session or weekly limit is “shared across all models, so the developer can’t restore access by switching models with /model.” That is the “Claude Code free limit” most people actually run into: not a free tier, but the end of the included allowance.

Try the free fixes first

Before you change backends, Anthropic’s page lists cost controls that keep the same models:

Run /clear between unrelated tasks. The docs note that when you want a fresh start, “/clear costs nothing,” while /compact is itself a large request.
Switch the default model with /model and set model: haiku for simple subagents.
Lower /effort for routine edits where deep reasoning adds little.
Move workflow-specific instructions out of CLAUDE.md into skills, which load only when invoked.
Use a hook to filter test or log output before it reaches the context.

These cut tokens without changing answer quality. If you still need a cheaper Claude Code workflow after that, the next lever is the backend.

Workload by workload: what a free backend can take

Ficelle publishes a 76-task benchmark of free routing against gpt-5. It is the closest evidence we have, and its limits matter here: the paid baseline was gpt-5, not a Claude model, tasks were single requests rather than multi-turn Claude Code sessions, and an LLM judge scored the results. Read the shares below as a direction for coding-agent workloads, not as a Claude Code head-to-head.

Coding-agent workload	Closest benchmark category	Free quality-safe share vs `gpt-5`	Reasonable default
Labelling issues, triaging failures, choosing a file to open	Classification (20 tasks)	100%	Free backend
Extracting fields from logs, configs or API responses into JSON	JSON extraction (20 tasks)	95%	Free backend
Small refactors, test scaffolding, step-by-step bug reasoning	Reasoning/code (16 tasks)	94%	Free backend, review the diff
Commit messages and short summaries	Not measured directly	n/a	Test on your own repo first
READMEs, release notes, user-facing docs under tight style rules	Content writing (20 tasks)	50%	Keep a paid model or edit by hand
Large architectural changes across many files over a long session	Not measured	n/a	Keep paid Claude

The weak spot is published, not hidden: on polished content writing, the paid model won 10 of 20 tasks, so free output was quality-safe on only 50%. The benchmark also never tested long, multi-turn agent sessions that plan across dozens of files. That is exactly the work Anthropic says to reserve Opus for, and nothing in Ficelle’s evidence shows a free model matching it.

What you can replace

You can keep the Claude Code interface and point it at a different backend. Ficelle runs locally and serves an Anthropic Messages endpoint next to its OpenAI-compatible one, so Claude Code connects without an adapter:

export ANTHROPIC_BASE_URL="http://127.0.0.1:8646"
export ANTHROPIC_API_KEY="ficelle-local"
claude

The API key is a placeholder. Your real provider keys stay in Ficelle on your machine, and requests go directly from your machine to the selected provider. Ficelle is not a hosted proxy.

By default, Ficelle maps Claude Code’s model tiers to its own routing profiles:

Haiku goes to ficelle/auto-fast.
Sonnet goes to ficelle/auto-coding, which only routes to models in the release’s bundled, certified coding pool that pass local availability checks.
Opus goes to ficelle/auto-orchestrator, which requires structured output and an orchestration probe.

Responses never present a free model as Claude. If no certified coding model is available, ficelle/auto-coding returns 503 no_certified_coding_model instead of quietly swapping in a model that merely supports tool calls.

What you cannot replace

Swapping the backend changes who answers, and several things do not come with it:

Claude model quality. A free open-weight model behind the Sonnet slot is a different model. The benchmark shows parity on narrow tasks against gpt-5, not equality with Claude on long agentic work.
Anthropic’s plan features. Plan usage bars in /usage, usage credits and /usage-credits belong to a claude.ai subscription login, not to a local endpoint.
Guaranteed capacity. Free pools run out. When every eligible free candidate is exhausted, Ficelle fails the request rather than spending money. That is strict-zero routing: paid fallback is hardcoded off, and a configuration that tries to enable it stops Ficelle from starting.
Upstream account terms. Some quota providers can bill beyond their included allowance if your account allows it. Ficelle quarantines models it sees reported as billed, but it cannot change provider billing settings for you.

A practical split for a cheaper Claude Code workflow: run triage, extraction, small edits and test scaffolding through a free backend, and switch your shell back to your Anthropic login for the architectural session or the release notes that customers will read. The strict-zero error tells you when the free pool is empty, so you decide whether to wait or move that task to a paid model.

Choosing a free LLM for coding agents

If you evaluate any free LLM for coding agents, Ficelle or not, check four things before trusting it with an edit loop:

Does it support tool calls reliably enough for file edits and shell commands?
What happens when the free quota ends: an error, a silent paid fallback, or a different model?
Does the response tell you which model actually answered?
Where do your provider keys and prompts go?

Ficelle answers those with tool-capability gating, a hard failure instead of paid fallback, no impersonation of Claude, and a local process that keeps keys on your machine. It is English-only and source-available under BSL 1.1. The free tier includes the routing core; the Pro plan adds a maintained provider pool, custom virtual models and updates. If that split fits how you use Claude Code, try Ficelle’s maintained routing on the work a free model already handles, and keep paying Anthropic for the rest.