CONCEPT Cheap sub-tasks want a cheaper model — but where?
Often a sub-task inside an agent loop is easy enough for a cheaper, faster model
(claude-haiku-4-5 or claude-sonnet-4-6) while the main reasoning wants
claude-opus-4-8. The instinct is to switch the main loop's model for that one call. That's
the trap. The right move is to spawn a sub-agent on the cheaper model and keep the main loop's
model unchanged.
CONCEPT Why not just switch the main loop's model?
Because the prompt cache is model-scoped. The main loop builds up a large cached prefix (system prompt, tool definitions, growing history). If you change the model on the next call, that cache no longer applies — the new model can't reuse the other model's cached prefix, so the entire prefix is re-processed at full price. You'd pay to re-ingest everything just to save on one sub-task: a net loss, plus added latency.
A sub-agent sidesteps this entirely. It's a separate context (4.4), so running it on a cheaper model doesn't touch the main loop's cached prefix at all. The main loop resumes on its original model with its cache intact.
| Approach | Effect on main loop's cache | Verdict |
|---|---|---|
| Switch main loop's model mid-conversation | Invalidated — cache is model-scoped; whole prefix re-processed at full price | ❌ Avoid |
| Spawn a sub-agent on the cheaper model | Untouched — separate context; main cache preserved | ✅ Do this |
DIAGRAM Two ways to run a cheap sub-task
cache warm"] -->|switch to Haiku| H1["Haiku call
cache INVALID → reprocess prefix"] H1 -->|switch back| O2["Opus loop
cache cold again"] end subgraph GOOD["✅ Spawn a sub-agent"] direction LR O3["Opus main loop
cache stays warm"] -->|delegate sub-task| SUB["🔵 Haiku sub-agent
separate context"] SUB -->|result| O3 end classDef bad fill:#f8d7da,stroke:#dc3545,color:#721c24; classDef good fill:#d4edda,stroke:#28a745,color:#155724; classDef b fill:#d1ecf1,stroke:#17a2b8,color:#0c5460; class O1,H1,O2 bad; class O3 good; class SUB b;
CODE Delegate the cheap sub-task to a sub-agent
The main loop stays on Opus (cache intact). A cheap, isolated sub-agent handles the easy sub-task.
# Cheap sub-task (e.g. classify/extract) → its OWN sub-agent on a cheaper model. def classify_cheap(text): r = client.messages.create( model="claude-haiku-4-5", max_tokens=64, # cheap + fast, separate context messages=[{"role": "user", "content": f"Classify as billing/tech/other:\n{text}"}], ) return "".join(b.text for b in r.content if b.type == "text") # Main loop stays on Opus — its model-scoped prompt cache is NEVER invalidated. label = classify_cheap(user_msg) # cheap sub-agent does the easy bit resp = client.messages.create( model="claude-opus-4-8", max_tokens=4096, # SAME model → prefix cache reused system=BIG_CACHED_SYSTEM_PROMPT, # large cached prefix, reused at cache price tools=TOOLS, messages=main_messages + [{"role": "user", "content": f"(category={label}) {user_msg}"}], ) # ANTI-PATTERN — do NOT do this inside the main loop: # client.messages.create(model="claude-haiku-4-5", messages=main_messages, ...) # → different model ⇒ Opus's cached prefix can't be reused ⇒ full re-processing cost.
EXAM TRAPS
PRACTICE
Q1 · Mid-loop you want a quick, cheap classification. Best way to keep costs down?
Spawn a sub-agent on claude-haiku-4-5 for the classification and keep the main loop on
its model. Switching the main loop's model would invalidate its cache and re-bill the whole prefix.
Q2 · Why exactly does switching the main loop's model cost so much?
The prompt cache is model-scoped. Change the model and the previously cached prefix (system, tools, history) can't be reused, so it's re-processed at full input price — often far more than the sub-task saves.
Q3 · Does running the sub-agent on Haiku cool down the main loop's Opus cache?
No. The sub-agent runs in its own separate context, so the main loop's cached prefix is untouched and stays warm for the next Opus call.