Future AI Pulse
Quick answer: Choosing a cheaper model is only useful if the completed work still meets its requirements. LegalOn’s case study pairs model selection with budgets and review practices. For a smaller team, cost per accepted deliverable is often a more useful comparison than the price of a single response.
What LegalOn’s account reports
OpenAI’s October 8, 2026 case study says LegalOn matched Astra, Sol and Luna models to different task complexity, introduced usage limits and restricted Fast mode by default. It reports an approximately 65% reduction in estimated daily costs while maintaining development performance. That is an estimated customer outcome, not a guaranteed saving for other teams.
The account also says LegalOn is building a measurement pipeline connecting feature releases with customer value and AI costs. That future pipeline should not be mistaken for a completed proof of return on investment. Future AI Pulse has not audited the company’s costs or tested its model combinations.
Define the deliverable before routing the task
A task with explicit inputs and a checkable output is easier to route than an open-ended request. A documentation correction might have a known source and a short list of changed paragraphs. A design decision might require exploring uncertain tradeoffs and consulting several constraints.
Use those differences to describe task categories, then test representative examples. Avoid treating a model’s position in a product lineup as sufficient evidence that it handles every task in a category. Keep a clear escalation path when the initial output fails its acceptance check.
Include the second attempt in the cost
Record the cost of the initial run, retries, review and corrections. An inexpensive first response can become expensive if it needs several attempts or substantial human repair. Conversely, a more capable model may be unnecessary for a simple well-specified transformation.
Compare completed work against the same criteria. Count an output as accepted only after it satisfies the required checks. Keep incomplete and failed runs in the record so the comparison does not hide expensive exceptions.
For each accepted deliverable, record: Task category and input version. Models used and the reason for any escalation. Initial run cost, retry cost and review effort. Acceptance checks and unresolved issues. Time from request to usable result. Do not invent usage amounts or convert an estimate into an audited result.
Give budget exceptions a purpose
A limit should identify when work stops and who can authorize additional spending. If a task reaches its limit, preserve the partial artifact and explain the remaining work. An exception can then be assessed against a concrete deliverable rather than an unspecified request for more resources.
Different projects may justify different limits. A discovery task may have more uncertainty than a routine maintenance edit. Record that reason and review it later; a generous starting allowance does not need to become an indefinite entitlement.
Keep speed and quality beside cost
Measure elapsed time to a usable result, not just response speed. Check whether the model selected for a task changes error rates, missed requirements or the amount of review. A lower bill accompanied by more regressions needs to be described as a tradeoff.
For a small pilot, use a handful of typical tasks and one difficult exception. Review the evidence before changing the default route for everyone. The goal is a practical policy based on your own work, not copying another company’s reported percentage.
Review the policy as the work changes
Save the task examples and acceptance criteria so later model or workflow changes can be compared. If the project begins handling longer inputs or new tools, rerun the relevant checks. A route that worked for yesterday’s task may no longer fit its new scope.
Related: our model-training brief covers explicit spending boundaries, and job-state guide examines avoidable retry and polling overhead.
Illustrative workspace photograph: Yen Vu / Unsplash, under the Unsplash License. Not a LegalOn product screenshot.
Sources
Source checked October 14, 2026. Written with AI assistance; measurement templates are original editorial suggestions.