Using more than one model

One model that both hunts for cases and judges them has no independent check. A second model, given a narrow job, can catch what the first one missed. Splitting the work also keeps cost down, because most of the work is reading, and reading does not need the strongest model.

Model names and descriptions change often. Everything product-specific below is accurate as of October 2026, and you should check the current docs before relying on it.

Choosing a model

You can set the model during a session with /model. You can set it at launch with claude --model <alias>. You can also use the model setting. A subagent has its own model field in its header. That field takes sonnet, opus, haiku, fable, a full model ID, or inherit.

Effort level is a separate setting from the model. You set it with /effort.

As of October 2026, the docs describe the tiers this way.

• Haiku is "fast and efficient" for simple tasks.
• Sonnet is for daily coding tasks.
• Opus is for complex reasoning tasks.
• Fable is for the hardest and longest-running tasks.
• An opusplan alias uses Opus to plan and Sonnet to execute.

We describe the models by role in this chapter. If the names change, keep the roles and swap the names.

The three-stage pipeline

Stage 1. Smaller models search and read

Hand the first pass to smaller, cheaper models. This is the case hunting, the reading, the quote pulling, and the first verification. Set the model field in your research agent to a smaller model. Run several agents at once if you have several questions. Subagents can run in parallel, up to 20 at once by default.

Stage 2. The main session checks for gaps

When the agents report, your main session runs a short gap check. Ask whether anything obvious was missed. Did the search cover state courts? Did it try other framings? Is a key case missing from the forward-citation list?

This is a check and not a redo. Do not hand it to another cheap agent. It needs to see the whole picture of your case, and the main session has that.

Stage 3. The strongest model reviews

The strongest model reviews the combined list of already-verified results. It weighs strategy, spots tension between authorities, and tells you where an argument is thin. It should not do raw case hunting. That is the expensive way to do the cheap job.

One exception. A single quick lookup is faster to do yourself. Spinning up an agent for one search costs more than the search.

The blind second read

Some cases carry an argument. If your position stands or falls on one case, read it twice, with two readers who do not talk to each other.

1. Agent A reads the full opinion and writes down the holding.
2. Agent B gets only the opinion and your question. Do not show it Agent A's summary. Ask it to state the holding on its own.
3. In your main session, compare the two.

If they differ, report both. Do not average them. Then read the opinion yourself at the disputed passage.

This catches the most common way a real citation goes bad. The case exists, but it holds something other than what a summary says. A second reader who never saw the first summary is not pulled toward it.

Ambiguity resolution before you ask the human

Extraction work throws up unclear items. Scanned text has garbled names. Initials are unclear. A header line is uncertain. Do not send every one to yourself.

Use a second AI pass first.

1. Look for context clues in the surrounding text.
2. Search the whole record for other appearances of the same name or phrase.
3. Look at the actual page image, if the text came from a scan.

Only items that stay unresolved, and that matter to your case, come to you. Apply a materiality test. Ask whether the answer could change a fact in a filing.

Record each resolved reading with a note on how it was resolved. In practice, most provisional readings turn out right and few of them matter. Your time is limited, so use it on the few that do.

What multi-model review catches

• Cites that do not exist, because the second reader cannot find them.
• Holdings that a summary got wrong, through the blind read.
• Gaps in coverage, such as a missing court system or a missing framing.
• Weak points in your argument, when a strong model plays the other side. Chapter 8 has a prompt for that.
• Inconsistencies between a draft and your facts file.

What it misses

• Mistakes that all the models share. Models can have the same blind spot and repeat the same wrong summary. That is why you read the opinion yourself for any case that matters.
• Whether a case is still good law. A model can scan for negative words, but it can miss a change in the legal foundation under the case. Chapter 6 covers treatment checks, and they are not a paid citator.
• Recent decisions that have not reached any free database.
• Facts outside your record. A model does not know what happened unless you wrote it down.
• Your judgment about strategy and risk. The decision is yours, and you sign what you file.

Using more than one model reduces errors. It does not guarantee accuracy. Two models can agree and both be wrong. The only proof is the opinion text, read and saved, and you remain responsible for what you file.

Wait for the research

A related habit. While a research task is still running, give the record facts and what the research needs to show. Do not give a verdict on whether a legal move will work. Say you are holding the verdict until the memo lands, then revise and note what changed.

A confident prediction made before the authorities are in can steer you away from an option you wanted. If the cases then cut the other way, you have lost time.

Cost control with caps

Unbounded runs burn money. Set limits.

• Put caps in the agent file. For example, up to 10 searches and 10 reads for a quick verify, and up to 30 searches and 25 reads for a sweep.
• Tell the agent to stop and report when it hits a cap, and to say what it did not get to.
• Name the model when you dispatch work, and use the smaller one by default.
• Use lower caps when you only need a quick answer. The caller can always set them lower.
• Give each agent one question. Ten narrow tasks cost less than one sprawling one.
• Keep an eye on your plan's usage. Plans and limits change, so check your account.

If you are on a plan with usage limits, caps also protect your allowance for the work that needs it most.

Next: Chapter 8 gives you four copy-ready prompts you can use with this setup.

Sources

• Claude Code model configuration
• Claude Code subagents