← vibemind.club

VibeMind · The receipts

Every number VibeMind publishes traces back to a run we actually did. This page is the full record: what we measured, exactly what the numbers were, and the one caveat that keeps each one honest. We publish only what we ran, so where you see a gap, it means we have not tested that yet and we say nothing about it.

The dataset

Every number below comes from one pile of real, paid-for runs.

Nothing here is a marketing estimate. We ran the same coding work through different setups, paid the real bill, and audited the logs. Where a result is an average, it is the middle of five runs on real open-source code, not a best-case cherry-pick, and the few single-run anchors are labeled as such.

69
audited runs, logged and cross-checked
$103.97
of real spend behind these numbers
5 runs
per measured result, the middle one reported
∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿
What we mean by a task

A task was real work, not a toy.

This matters for reading the calculator on the front page: when it asks how many tasks you do a day, here is the size of task we measured, so you can size its slider against your own work.

Most of our runs used real small features, the kind of change you actually ship: each one took roughly three minutes of agent working time on a 24,000-line open-source codebase. On top of that, we ran one larger feature that took about sixteen minutes in a single long session, to see whether the story holds when the work gets heavier.

So when you set the front-page calculator to, say, six tasks a day, picture six real features of that size. If your own tasks run longer than three minutes, the time you get back grows, not shrinks.

∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿
Three ways to do the same work

The same three tasks, run three ways, side by side.

We gave the same batch of three real features to three setups and measured what each one cost. The reader takeaway is simple: isolated lanes finish in under half the wall time of one long chat, at about the same token count, and they make far fewer tool errors than throwing several sessions at one shared copy of the code.

How the work was run Wall time Output tokens Dollars Tool errors
One long chatall three tasks in one session, in a row 547 s 38,185 $1.50 9
Naive shared checkoutseveral sessions on one copy of the code 302 s 44,996 $2.26 28
Isolated lanesVibeMindeach session in its own copy of the code 259 s 39,078 $1.89 17

Medium repo, 24,000 lines, sonnet coder. Each number is the middle result of five runs. The machine was shared, so the exact wall-clock seconds are directional; the ratios between the setups are what hold.

The row that costs you is the middle one. Opening several sessions on a single shared copy of the code was the most expensive setup we measured, and it made the most tool errors (28, against 17 for isolated lanes), because the sessions kept stepping on each other's changes to the same files. Isolated lanes avoid that by construction: each lane works in its own copy and owns a separate set of files, so that particular collision cannot happen in the first place. On dollars, isolated lanes came in about 16 percent under that naive shared-checkout run, the one comparison where running organized is measurably cheaper on the bill.

The technical detail
Batch of 3 tasks, medium repo (~24k LOC), sonnet coder, medians of 5 trials. Serial (one session, three tasks in sequence): 547 s, 38,185 output tokens, $1.50, 9 tool errors. Parallel-shared (naive): 302 s, 44,996 tokens, $2.26, 28 tool errors. Parallel-isolated (VibeMind): 259 s, 39,078 tokens, $1.89, 17 tool errors. Derived deltas: isolated vs serial = 2.11x wall faster, +2% tokens (parity), +26% dollars. Isolated vs naive-shared = -16% dollars, -39% tool errors (17 vs 28). Naive-shared vs serial = +18% tokens, +51% dollars. Shared machine, so absolute wall is directional; token and dollar ratios are not affected. We observed no hard merge failures in these trials, so we make no hard-failure-rate claim; collision safety is by construction, not a measured zero.
∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿
When the work gets heavier

On a bigger feature, the gap gets wider, not narrower.

The story above holds on small tasks. To check the heavier end, we ran that one sixteen-minute feature two ways: in a single long session, and in an isolated lane. Here is what changed.

about 3x
faster wall time than the one long session (3.03x)
+8%
dollars, a smaller premium than on small tasks
-36%
fewer tokens than the one-long-session run
One run each way, no error bars

On the small tasks, the dollar premium for running organized was larger and the token counts were about even. On this heavier feature, the isolated lane pulled ahead on every axis: it finished about three times faster, used about 36 percent fewer tokens than the one-long-session run, and the dollar premium shrank to 8 percent. This is a single measured run each way, stated as such, so treat it as one solid data point rather than a five-run average.

The technical detail
Large-task edge, N=1 per mode (measured anchor points, no error bars). Serial: 944.6 s (15.7 min), 69,339 tokens, $1.83. Isolated: 311.5 s (5.2 min), 44,431 tokens, $1.98. Derived: 3.03x wall speedup (up from 2.11x on the small-task batch), isolated -36% tokens vs serial (the token story flips to isolation at this size), dollar premium narrows from +26% to +8%. Per-lane batch wall stays flat even at the larger task.
∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿
What the organization costs you

The organizing itself is nearly free to run.

A fair question: all these lanes and merges and memory, do they quietly eat the savings? We timed the machinery's own footprint. It is small enough to round off.

1.2%
of the run's wall clock is the organizing itself: carving a workspace per lane and merging the results back, timed at 408 milliseconds
92.6%
of the tokens you are billed for are cache reads, priced far below fresh input
0
times the work stopped on a plan rate limit, across 41 runs

So the machinery is cheap to run: about a hundredth of the wall clock, mostly cache reads, and never once stopped by a plan limit. The expensive thing was never the structure. It was the disorganized default that most people already reach for: several sessions colliding on one folder, the setup that topped the cost table above.

The technical detail
The 408 ms ceremony is one controlled trial: worktree carve plus three merges, auditor-timed on the parallel run, about 1.2% of that run's wall clock. The 92.6% cache-read share and the zero rate-limit hits come from the base cell and the 41-run wave, on one account and a shared machine. This is the footprint of the orchestration steps themselves. It is not a with-machine-versus-without A/B, which we did not run.
∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿
Why the machine has to exist

Even a top model will not organize the work on its own.

If the best setup is isolated lanes, why not just let a smart model choose it? We tested exactly that. We handed the whole job to the model and watched what it picked.

Across nine runs, spanning three quality tiers of model (including the strongest), not one chose to run the work in parallel isolated lanes on its own. Every run reached for the slow, single-track shape by default. None split the work into lanes. None handed the cheaper parts to a cheaper worker.

Every run still finished the work correctly. The gap was not reliability, it was strategy: left to its own judgment, even the strongest model reached for a slower, less efficient shape. That is the job VibeMind's lane machinery does. The good strategy does not appear on its own, so the machine encodes it.

The technical detail
Nine runs total: three model tiers, three trials each, on the identical medium-repo three-task workload. Zero cheap-tier workers dispatched, zero parallel delegation, not once, at any tier. The finding is about strategy selection, not reliability: reliability was 3/3 on every trial at every tier. This is the small-batch regime. Delegation's compounding advantage is a many-task and large-task property, shown by the N-scaling and the large-feature anchor above.
∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿
Why you can trust the calculator

We wrote down the answer before we ran the test.

Before we ran the final benchmark, we wrote the cost model's prediction to a file with a timestamp. Then we ran the test. Here is what it predicted, next to what actually happened.

What we predicted, in advance
output tokens52,514
dollars$2.44
wall time~267 s
What the test measured
output tokens48,679
dollars$2.26
wall time275.6 s
tokens: off by 7.3% dollars: off by 7.4% wall: off by 3.2%

The prediction landed within 8 percent on all three measurements. That is why the calculator on the front page is an instrument, not an illustration: it is the same cost model, and we have shown it can predict a run it had never seen.

Honest scope, in one sentence: this was a prediction made between two points we had already measured, not beyond them, so the model is proven on the range it was tested on and we do not stretch it further.

The technical detail
The cost model was fit only on earlier rows. The prediction for a held-out N=4 isolated run (medium repo, sonnet) was written to disk and timestamped 2026-07-08T21:40:42Z, then never edited, and only then did the run happen. Predicted: 52,514 output tokens, $2.44, ~267 s. Measured median of 3: 48,679 tokens, $2.26, 275.6 s. Deltas: -7.3% tokens, -7.4% dollars, +3.2% wall, each inside the error ranges we wrote down in advance (tokens 41.5k-63.5k, dollars $2.14-$2.73, wall 205-384 s). This is an interpolation success on the N-axis: N=4 sits between the fitted N=3 and N=6. The model is deliberately simple and is not validated off that fitted axis (repo size, task complexity, coder tier).
∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿∿
What we did not run

The gaps are on purpose. We only claim what we tested.

A page like this is only worth reading if it also tells you where it stops. So here is the honest edge of the data.

  • We publish only what we ran. Small-repo and large-repo shared-checkout comparisons were not run, so we make no claims there.
  • The larger feature is a single run each way. It is a strong data point, not a five-run average, and we label it that way.
  • The cost model is proven between the points we measured, not beyond them. We do not stretch it across repo size, task complexity, or coder tier.
  • Wall-clock seconds come from a shared machine, so treat them as directional. The ratios between setups are the durable part.
  • We make no quality-degradation claim and no cheaper-worker economics claim beyond the specific numbers shown above.

If a number is not on this page, we did not measure it, and we would rather leave the space blank than fill it with a guess. That is the whole point of the receipts.