How to Run Claude Code Cheaper by Cutting Token Spend
Swapping Claude Code onto a cheaper model like MiniMax M3 lowers your per-token rate but not your token volume. The larger saving comes from retrieval: search an indexed vault before you call any model. RDK answers most queries from stored chunks, cutting token spend 80 to 90 percent.
Why a cheaper model only solves half the cost problem
Your Claude Code bill is rate multiplied by volume. A model swap to something like MiniMax M3 attacks the rate. It does nothing about volume, and volume is where the waste lives. Every session re-sends the same context, re-derives the same conventions, and re-answers questions the codebase already answered last week. You pay full price to regenerate an answer you have produced before.
Rate cuts are capped. A model is only so cheap before quality degrades on the hard 5 to 10 percent of work that actually needs a frontier model. Volume cuts are not capped the same way. If you stop sending a query to any model, its cost goes to zero. That is the lever this guide is about.
Retrieve the answer instead of regenerating it
RDK (Retrieval Development Kit) indexes files from your local vaults, Obsidian notes, docs, and code, as encrypted private chunks on a distributed network. When an agent has a question, it searches those chunks first. If the answer is already written down, the agent retrieves it and never spends a token generating it.
This is why retrieval beats a rate cut. A cheaper model still charges you for the query. Retrieval removes the query from the bill entirely. Across a real workload RDK cuts token spend 80 to 90 percent, because most of what an agent asks has been answered before and stored.
The savings scale with repetition, which is exactly the shape of coding work. A team asks the same setup, convention, and architecture questions dozens of times a week. With a plain model, every repeat is a fresh full-price generation. With retrieval, the first answer is written to a chunk and every repeat after that is a lookup. The more your team works, the wider the gap between a rate cut and a volume cut grows.
Stacked retrieval, tier by tier
Stacked retrieval routes each query through cheaper layers before the expensive one. Your private vault RAG answers 40 to 65 percent of queries directly. The public RDK network, chunks other builders have published, adds another 15 to 20 percent. Only the remaining 5 to 10 percent falls through to the LLM as a genuine fallback. You pay model rates on that thin slice, not the whole stream.
Combine both levers so they compound
These levers are not either or. Point Claude Code at a cheaper model for the fallback tier, and put RDK in front of it for everything else. Retrieval shrinks the volume that reaches the model. The cheaper model shrinks the rate on what is left. Multiply a 5 to 10 percent residual volume by a lower per-token rate and the frontier-model bill you started with is barely recognizable.
The order matters. Retrieval goes first because it is the larger multiplier and because it protects quality: the hard queries that still need a strong model are exactly the ones you do not want to starve by over-optimizing the rate.
There is a practical setup implication. Configure Claude Code to route through RDK for retrieval, and reserve your model choice for the fallback path only. That way a decision like moving to MiniMax M3 affects only the 5 to 10 percent residual, so a rate experiment can never quietly degrade the answers your team relies on most. You tune the cheap tier freely while the retrieved answers stay stable.
How billing actually works
RDK bills cost per connection, not commission and not a metered inference charge that climbs every time your team asks the same thing twice. Payments run on Base using USDC through the CryptoCadet rail. There are three user classes: Node Operator, Enterprise, and Builder. A Builder indexing a single vault and a team running Enterprise nodes both get the same core behavior: search before you generate, pay for connections rather than for regenerating known answers.
As a side effect of publishing chunks public, the same work you indexed to save tokens can earn USDC per retrieval when other agents pull it. Cost reduction and revenue come from the same index.
| Lever | What it reduces | Ceiling | Quality risk |
|---|---|---|---|
| Swap to a cheaper model (e.g. MiniMax M3) | Rate per token | Capped by how cheap a model can go before quality drops | Higher on hard queries |
| Retrieve with RDK before calling the model | Volume of tokens billed | 80 to 90 percent of token spend removed | Lower, model still handles the hard 5 to 10 percent |
Frequently asked questions
- Is MiniMax M3 or a cheaper model enough to run Claude Code cheaply?
- It helps, but it only lowers your rate per token. Your session still sends and regenerates the same volume of context and answers. To cut cost meaningfully you also need to reduce volume. Put RDK in front so most queries resolve from an indexed vault, then let the cheaper model handle only the fallback.
- How much does retrieval actually save on token spend?
- Across a real workload RDK cuts token spend 80 to 90 percent. That range comes from stacked retrieval: your private vault answers 40 to 65 percent of queries, the public network adds 15 to 20 percent, and only 5 to 10 percent reaches the LLM. You pay model rates on that residual instead of on everything.
- Do I have to choose between a cheap model and retrieval?
- No, and you should not. They target different parts of the bill and compound. Retrieval removes most queries from the model entirely, and the cheaper model lowers the rate on the small fraction that still needs generation. Add retrieval first because it is the larger multiplier and it protects answer quality.
- Is my code exposed if I index it for retrieval?
- No. RDK indexes your vault as encrypted private chunks by default. Agents search them before calling an LLM, and nothing is published unless you deliberately mark chunks public. If you do publish, other agents can retrieve them and you earn USDC per retrieval on the Base rail.