Introducing Vetta: the most efficient harness for long-horizon tasks

One agent, one turn. The loop is an alarm rather than a call stack, so a task measured in hours is a sequence of short slices that each commit before they end — and a crash costs one turn, not the run.

dashed — shipped in the runtime, not exercised by the benchmark in this post

Drawn from the implementation, and what is missing is deliberate. No box here runs on spot compute and none queues a batch job: the completion window is a tariff on the request, and the live path issues an ordinary call on every window. Delegation is sequential — the board takes one run per turn, so two live children run one after the other. It buys context isolation, never wall-clock, and the instrument that would falsify a parallelism claim reads one in-flight sub-agent on every trial ever recorded. The measured arm in this post runs Vetta's prompt, tool surface and thinking budget on the pi agent core; the durable loop, the store and delegation are the shipped runtime, and the dashed box is the one the benchmark did not exercise.

Axis: measured dollars per COMPLETED task, indexed to Vetta = 1.00. Shorter is cheaper. Beside each bar: the share of tasks that harness finished, and the tokens it spent per finished task — a count that depends on no price at all.

One model, zai-org/GLM-5.2-FP8, held fixed on every row. Same benchmark, same wall-clock cap, same default completion window, one machine class per harness, all eight long-horizon tasks on every row. Dollars are ledger dollars: our rate card books every prompt token at the full input rate while the vendor bills cache-read input at a fifth of it, so each row's card figure is divided by an over-statement factor measured for that harness at that window. Those factors run from 1.017x to 4.176x, which is why an uncorrected card ranking is not the same ranking — and they are single-bracket measurements good to about ±10%, so the dollars are printed to three decimals and the ranking rather than the digit is the claim. Cost per completed task divides everything an arm consumed — including what it consumed on tasks it did not finish — by the tasks it did finish. The bracketed rate counts only attempts that never hit the wall-clock cap: claude-code never failed a task it was given time to finish, and hit the cap on more than a third of its attempts, so the clock is part of the raw ranking.

Axis: everything INDEXED to asap = 1.00. The measured price per token lands on the published input tier to within a thousandth. What you pay for it is latency — the median model call, printed in raw seconds beside its bar.

Vetta on all three rows — same harness, same model, same tasks, only the requested window differs. The measured price is cost ÷ tokens off the same meter, and it lands on the published input tier (1.000 / 0.625 / 0.500) because this workload is 97.6-98.3% prompt tokens; a workload with more output in it would sit higher, since the output tiers fall less steeply than the input ones. Cost per completed task is ledger dollars, and the over-statement factor was measured separately at each window rather than assumed equal across them, so the three rows are corrected independently and the ratio between them carries three separate measurement errors rather than none. Latency is the one series not drawn from the benchmark runs: it was measured directly against each window, and the same three readings set the per-window wall-clock cap in the suite — so the completion figures compare unequal time budgets by design, and the price and cost figures are unaffected by that scaling.

Axis: each arm's own bill at loose divided by its own bill at priority, so 1.00 means the cheaper tier bought nothing. Both arms were offered the same tariff on the same model and the same tasks. bare pi is Vetta's harness without Vetta's configuration.

The same tariff cannot be worth two different amounts, so the difference between these two rows is not the price — it is what each agent could do with the extra latency the cheaper tier costs. bare pi hits the wall-clock cap on roughly three attempts in four at both windows: it has no slack, so it spends the discount on more tokens and its bill comes back where it started. Vetta hits the cap on one attempt in six, converts the tariff, and its bill falls by more than the tariff alone. These three ratios are the one place a card figure appears in this post, and only because none of them is a dollar and none is ranked against another arm: each is one arm measured against itself off the same meter on both sides. bare pi has no measured over-statement factor at either window, so no dollar for it appears anywhere here. The number the post ranks on remains cost per FINISHED task, which cannot be improved by finishing fewer.

Each cell is the outcome for one task on one window. 6 of 8 tasks scored identically on all three — reading across a row asks whether the price tier moved the result, and on most rows it did not.

Vetta on all three windows, same model, same tasks, repeated attempts on each. The two rows that moved are named. A price cut paid for in unfinished tasks is not a price cut, which is why this figure prints every task rather than a summary rate — and why the one task loose lost is worth naming twice: it failed on wrong answers, not on truncation, which is the opposite of what a running-out-of-time story would predict. Two rows out of eight cannot be told apart from ordinary run-to-run variance here; what the figure can say is the shape.