One grep result cost us $1.14.

Not the reasoning. Not the fix. One search that dumped 86,000 tokens into the transcript. The model kept two lines — and then re-read the whole dump on every later step, because that's how agent transcripts work. Paid once to write. Paid again, and again, until the session ends.

The wire log doesn't lie. Tool output isn't output — it's a deposit, and it earns negative interest.

That receipt is why we rebuilt Empryo's engine. Then we put the rebuild in a cage.

Full disclosure as always: I build Empryo — alone, nights and weekends, no funding. Every number below comes from recorded runs (Round 1, Round 2, and the new Round 3 board).

The cage match

Four bugs that were really reported and really fixed in hono and opencode. The fix PRs' own tests, hidden from both agents. Git history scrubbed so the answer is unreachable. Six models, two providers, 96 metered runs.

pi at its floor, Empryo at its ceiling. Every real pi install grows past this shape — ours is the shape.

Mid-series we even found pi's cost meter was broken on the GPT side and upgraded our opponent so its numbers would be right. Making the other guy's bill more accurate is the cheapest credibility you'll ever buy.

What the bill taught us

Forge v1 — our old engine — gave every model the same everything. The whole map, all 37 tool contracts, on every step, for every model. The wire log said: that's not intelligence, that's postage.

v2 in one picture: strong models get the short tour, small models get the full manual and a scout, and every tool result gets a budget.

Same bugs, same models, same keys — the engine swap alone cut Opus −56%, terra −46%, and took luna from fixing half the bugs to fixing all of them.

One thing v2 deliberately does NOT contain: benchmark-shaped hints. We wrote some. They worked. We deleted them, because an engine tuned to a test is a worse engine wearing a better score.

The run that explains everything

Two files, one name, one bug. Strings guess. The shape knows.

That's the whole idea in one cell. Code isn't a wall of text that happens to compile. Functions call functions, files move together — and an agent that can see that doesn't guess.

The board

Round 3, the wins. Full board with every cell — losses included — on the results page.

The honest parts, because a benchmark without them is an ad: on cheap and mid-price models, pi's tiny prompt keeps the cost edge — when a million tokens costs a quarter, nothing beats twenty lines. And pi's totals reproduce almost exactly between runs while ours wander ($1.20, then $1.47 on Sonnet). Tightening that band is the next milestone.

But the pattern is the thesis: the smarter the model, the more the shape pays. Frontier models handed a map take fewer, better steps — and at frontier prices, fewer steps is real money. Small models want the map served in slices. Intelligence isn't a feature you include. It's a delivery you price.

Want to check us? The whole cage — tasks, hidden tests, metering — lives in the public bench repo. The engine is one click away: the V2 button beside Empryo's prompt. The bill is the judge either way.