How I use Claude Code subagents to make my Claude Pro limits last longer

Andrea Grandi is a developer on the €20 Claude Pro plan who kept hitting the 5-hour and weekly usage limits on his side projects. Every Claude Code session ran on Opus 5.5. His fix is to keep Opus for planning and review, and to hand the rest to cheaper models through subagents.

The setup is four Markdown files in ~/.claude/agents/. An implementer on Sonnet 5.5 writes the code and tests from a plan Opus has already settled. A scoper on Haiku 5.5 gathers facts before planning: it is read-only, stops after 12 turns, and has to cite a file for every claim. A shipper (also Haiku) commits, pushes and opens the pull request, and stops if anything looks off. A second Opus reviewer exists but stays out of the normal flow, since by his account a second review can cost almost as much as writing the code. Small changes stay in the main session. He publishes all four files, the delegation rules from his CLAUDE.md, and the status-line script that logs his usage.

Then the numbers. Ten delegated sessions produced 13 pull requests for 80% of one 5-hour window, or about 6% each. Against 24 earlier Opus-only sessions, his cost per pull request fell from $3.59 to $2.14 (API-equivalent dollars, which is not what he pays). From that he estimates about sixteen pull requests per window where he used to get ten. He also says the saving is mostly not Sonnet’s lower price: repricing the subagents’ tokens at Opus rates makes the ten sessions only 5% more expensive. What changed is the main session, which made fewer requests (54, down from 75) with less context in each.

His published figures add up, and Anthropic’s release notes confirm the Sonnet 5.5 cache-read price cut he mentions (from $0.20 to $0.10 per million tokens, the day the post went up). But the comparison is looser than the table suggests. The two groups of sessions are not matched: he gives no task types or dates, and the delegated pull requests are larger on average. The 10% figure for an Opus-only pull request is converted from cost, since he has no usage readings from before the change. So “sixteen where there were ten” is an estimate. That quality held after the switch is his word; the post counts no rework.