The first $0.034646 we can prove

The first GLM-5.3 bill Fourth Down Labs can prove down to every attempt is $0.034646. It bought five accepted specialist reviews. It also bought one response we rejected. The rejected response alone consumed $0.013816—39.9% of the observed bill.

That number is exact and narrow. It came from our five-case synthetic prompt-injection evaluation, not a full fantasy-football research bundle. Three earlier online research runs completed 13 GLM-5.3 role outputs before per-attempt ledgers existed. Their token use, latency, retries, and cost are unknown.

Six attempts, line by line

The statistician's first response reached the 3,000-token output cap, ended with finish_reason: length, and failed schema validation. That attempt cost $0.013816. The retry returned a valid structured review for $0.0034716. We paid for both because the provider performed both inferences.

The failed response was also the most expensive single attempt in the run. Hiding it would make the workflow look 60.1% as expensive as it actually was. That is precisely why the ledger flushes after each attempt instead of waiting for a final accepted artifact.

Where the bill accumulated

The statistician accounted for 49.9% of the bill because its two attempts used 4,529 of 9,580 recorded tokens. The editor was cheapest at $0.0016792. Those are observations from one frozen evaluation, not stable role prices.

Tokens are not as simple as one counter

OpenRouter reported 2,502 prompt tokens, 7,078 completion tokens, and 9,580 total tokens. Completion represented 73.9% of the recorded total.

The receipts also contain 5,326 reported reasoning tokens inside completion-token details. We do not add that field to totalTokens. In one failed attempt, the provider-reported reasoning count exceeded the completion count, so presenting it as a clean additive subset would create a reconciliation the source does not support.

That does not make the field useless. It makes its semantics provider-defined diagnostic metadata. Model and provider comparisons should use the same preserved fields and explicitly test reconciliation instead of silently normalizing whatever looks odd.

Modal and Together both answered

The first researcher attempt and failed statistician attempt were routed to Modal. The statistician retry and remaining three roles were routed to Together. OpenRouter returned those selected-endpoint labels in per-attempt metadata.

This is not a provider benchmark. The requests served different roles; even the statistician retry occurred after a failed sample and produced a radically shorter answer. We did not randomize provider assignment, repeat identical calls, or control endpoint conditions. The public ledger therefore reports both names without claiming one was faster, cheaper, or better.

96 seconds of calls, 105 seconds on the clock

The six transport measurements sum to 96.323 seconds. From the first request start through the last response, 105.013 seconds elapsed. The 8.690-second difference includes the gaps between sequential calls and local orchestration work.

Neither number is a concurrency benchmark. This evaluation ran roles sequentially because later review evidence and retry handling were part of the experiment. A different orchestration graph could reduce wall time without changing aggregate provider latency—and could also introduce new dependency or correlation risks.

The earlier bill stays blank

The three earlier online research runs are more representative of the publication workflow than this synthetic evaluation. They are also less economically auditable. Their 13 completed role artifacts preserve decisions and configured model identity, but no call ledger survives.

We do not multiply today's price by an imagined token count. We do not assume one attempt per surviving output. We do not put $0 in a null field. The known historical billed-cost floor is $0.034646, while the historical total remains null.

What this changes operationally

What $0.034646 does not mean

It does not mean a full Fourth Down Labs study costs three cents. Synthetic evidence was much smaller than a production bundle. It does not prove GLM-5.3 is cheaper or more expensive than a challenger. No challenger saw this frozen corpus. It does not include OpenRouter credit-purchase fees or reconstruct missing calls from list prices.

It does support a less glamorous, more useful claim: six recorded attempts cost exactly $0.034646 according to the response receipts, one rejected attempt represented 39.9% of that bill, and our earlier historical total cannot be calculated.

This economics artifact is a deterministic aggregation of already-preserved receipts. It did not call GLM-5.3 or send the unpublished Touchdown Regression bundle to OpenRouter.