Cheaper agents move the cost bottleneck towards checking the result

Original research record

We connect the OTR data on Haiku 5.5's lower running cost with its repeated warning that research needs source checks and contradictions. Anthropic released the small model on 7 October. Its published price table changes above a 100,000-token prompt boundary. Our signal is that cheaper first-pass extraction makes it easier to read more records, but it can also create more material that someone must verify. Total savings depend on the complete workflow.

We would run a fixed set of multilingual records through the cheaper model and the current process. We would compare correct facts, missed qualifiers, wrong citations, retries and total review time. We would keep difficult reasoning separate from narrow extraction. We would freeze 100 records, keep the same source set and acceptance rules, and compare total paid model work plus timed human checking for correctly completed records.

If a controlled test reduces total cost while preserving source accuracy and requiring no extra review, the checking bottleneck would be smaller than we expect.

We would buy cost per verified fact, rather than cost per generated paragraph. This connects model efficiency to a measurable editorial outcome.

Bayesian · what changed the estimate? We estimate a subjective 60% chance that a one-week trial will reduce first-pass model cost but save a smaller percentage of total workflow cost. We would compare like-for-like completed work. we have not fitted a prior or evidence likelihoods, so this is not a computed posterior. de Finetti · are the comparisons fair? we treat repeated OTR editions as one underlying source when they reuse a report; our judgement must not gain confidence just from repetition. Kolmogorov · do the numbers agree? for the stated test event, the probability of it occurring and not occurring sums to 100%. None of these checks establishes historical calibration. This estimate is conditional on running the described pilot with its records and comparison method fixed before the test. No pilot outcome is claimed here.

The OTR data treats cheaper sorting and stronger verification as separate stages. The producer's documentation also distinguishes narrow small-model tasks from complex agent work.

Not applicable: this is a technology synthesis, not an investible or price forecast.

Highly repetitive extraction can be accurate and need little checking. Caching and shorter prompts may cut more cost than expected.

Not applicable: no earlier market direction is assigned. We state the evidence dates and the condition that would change the technology view.

A producer's average model-cost claim is not evidence that research staff time, web access, storage and checking costs fall by the same amount.

We have not verified completed, matching technology forecasts in the OTR records inspected for this signal. The percentage is an uncalibrated judgement and is not a historical hit rate.

How likely is this call? · Three frameworks

Bayesian · what changed the estimate? We estimate a subjective 60% chance that a one-week trial will reduce first-pass model cost but save a smaller percentage of total workflow cost. We would compare like-for-like completed work. we have not fitted a prior or evidence likelihoods, so this is not a computed posterior. de Finetti · are the comparisons fair? we treat repeated OTR editions as one underlying source when they reuse a report; our judgement must not gain confidence just from repetition. Kolmogorov · do the numbers agree? for the stated test event, the probability of it occurring and not occurring sums to 100%. None of these checks establishes historical calibration. This estimate is conditional on running the described pilot with its records and comparison method fixed before the test. No pilot outcome is claimed here.

Bayesian · what changed the estimate?

We estimate a subjective 60% chance that a one-week trial will reduce first-pass model cost but save a smaller percentage of total workflow cost. We would compare like-for-like completed work. we have not fitted a prior or evidence likelihoods, so this is not a computed posterior.

de Finetti · are the comparisons fair?

we treat repeated OTR editions as one underlying source when they reuse a report; our judgement must not gain confidence just from repetition.

Kolmogorov · do the numbers agree?

for the stated test event, the probability of it occurring and not occurring sums to 100%. None of these checks establishes historical calibration. This estimate is conditional on running the described pilot with its records and comparison method fixed before the test. No pilot outcome is claimed here.

Evidence behind this view

The OTR data treats cheaper sorting and stronger verification as separate stages. The producer's documentation also distinguishes narrow small-model tasks from complex agent work.

What past price moves show

Not applicable: this is a technology synthesis, not an investible or price forecast.

Evidence against this view

Highly repetitive extraction can be accurate and need little checking. Caching and shorter prompts may cut more cost than expected.

Why the direction changed

Not applicable: no earlier market direction is assigned. We state the evidence dates and the condition that would change the technology view.

What people say vs the facts

A producer's average model-cost claim is not evidence that research staff time, web access, storage and checking costs fall by the same amount.

How past forecasts performed

We have not verified completed, matching technology forecasts in the OTR records inspected for this signal. The percentage is an uncalibrated judgement and is not a historical hit rate.