7/7 shipped at 12% of raw Opus cost

Opus baseline: 7/7 shipped. $1.64 per session.
Hybrid cell: 7/7 shipped. $0.20 per session.

The hybrid: one Sonnet session to seed the store. Haiku for every session after.

Same results. 12% of the cost.

Tokenmaxxing is not the high-performance option. It is just the expensive one.

The common assumption: if you want Opus-level results, pay Opus prices.

The hybrid cell breaks that. Sonnet on session one to establish architecture and seed the memory store. Haiku on every subsequent session, with that context injected.

Subsequent sessions are implementation, not discovery. Implementation does not need frontier reasoning. It needs frontier context. Haiku with the right context outperforms Opus starting cold.

The tokenmaxxing argument goes like this: larger models produce better output, so higher token spend is a quality investment. Some teams are running Claude Opus on everything, justifying the cost as a performance decision.

I benchmarked five cells on the same project and tasks: Haiku baseline, Haiku with Hydrate, Sonnet baseline, Sonnet with Hydrate, and a hybrid routing cell.

The Opus baseline cost $1.64 per session. 7/7 tasks shipped. The hybrid cell cost $0.20 per session. 7/7 tasks shipped. Identical results at 12% of the cost.

The hybrid routing pattern: one Sonnet session to scaffold the project and seed the memory store with architectural decisions. Haiku for every session after that, with the store injected before each session starts.

The reason this works is that subsequent sessions are implementation, not discovery. The architecture is established. The conventions are set. Haiku is not being asked to make frontier reasoning decisions. It is being asked to implement against a specification it can hold in context. With the right context, it does that reliably.

The Opus premium pays for reasoning capacity. Most implementation sessions do not require that capacity. They require the context that tells the model what was decided.

Full benchmark data: gethydrate.dev/blog/tokenmaxxing-and-the-empirical-answer

Opus: 7/7 shipped, $1.64 per session.
Hybrid (Sonnet seed + Haiku+Hydrate): 7/7 shipped, $0.20 per session.

12% of the cost. Identical results.

Tokenmaxxing is not the high-performance option. Just the expensive one.