Test reasoning settings to find a lower-cost way to run your existing AI automation

If you want to reduce an existing automation’s running costs without handing more work back to staff, test its reasoning setting against completed business outcomes. Adopt a lower setting only when predefined quality limits hold and total cost per accepted outcome improves—not simply when the model call becomes cheaper.

Reasoning effort is an operating setting that guides how much internal reasoning a model applies to a task. Supported levels depend on the model; lower effort favors lower token usage and speed, while reasoning tokens themselves are billable. That makes the setting worth evaluating within your existing automation. OpenAI reasoning documentation

Reduce operating cost without shifting work to staff

Cost per accepted outcome gives you a way to evaluate potential savings while accounting for the work needed to finish each task. Define an accepted outcome as a completed business task that meets your agreed criteria, then calculate:

Cost per accepted outcome = total operating cost for the test batch ÷ accepted outcomes in that batch.

Include model and tool charges for every attempt, retries, human review and remediation. Count each accepted outcome once, even when it requires several attempts. Keep unsuccessful work in the numerator; if nothing passes, the setting fails rather than producing a usable unit cost.

Quality is a gate before comparing costs: a lower unit cost cannot compensate for a prohibited error. Also compare test and implementation expenditure with the expected operating benefit over your planning period.

FAIS’s white-glove operation of existing agents and consulting would connect configuration testing to business acceptance rules and staff effort. The opportunity is to improve the economics of automation you already use, without requiring a model switch or software replacement.

What does the Vals benchmark tell buyers?

The Vals results justify testing a lower setting, not assuming equivalent business performance. Research published on November 3, 2026, with a chart dated October 27, reports that 4 of 7 models retained at least 80% of their Vals Index score at their lowest setting for half the cost or less. None matched its leaderboard-setting aggregate score at its lowest setting. Vals research

Vals tested lowest, middle and leaderboard settings across 8 benchmarks covering finance, coding, legal and tax work, using the same fixed subset—approximately half the tasks—at each setting. Costs used list prices and a plain mean across benchmarks, with Code Migration estimated from tokens. Vals methodology

Most benchmarks had a single run per setting; older leaderboard runs and October lower-setting runs could also reflect provider changes. Aggregate score retention therefore cannot establish acceptable accuracy or review burden for your workflow. Use the evidence to select a test, not a savings target. Vals limitations

Keep the workflow fixed for a useful comparison

A controlled comparison changes reasoning effort while holding the model, prompts, tools, permissions and business rules fixed. Check your model’s supported effort levels, then follow this test sequence:

1. Select representative historical cases, including routine work and consequential exceptions.
2. Define acceptance criteria, prohibited errors and review limits before running either setting.
3. Replay the same cases at the current and candidate settings without permitting live business changes.
4. Record outcomes, attempts, charges, review time and remediation consistently.
5. Repeat enough cases to assess variability, then compare quality gates and total cost per accepted outcome.

Task-specific tests, production-representative datasets, human-calibrated scoring and ongoing evaluation follow OpenAI’s evaluation guidance. Choose sample sizes and repetition counts for your workflow’s variability and consequences rather than importing a universal threshold.

Make acceptance rules reflect your business

Consider an illustrative Taiwan manufacturer whose existing automation prepares supplier delivery-change assessments. Before testing, it reads a supplier message, looks up affected purchase-order lines and production requirements, then prepares an assessment for a planner to approve.

After introducing the test, production stays unchanged. A separate replay compares settings on saved inputs. Each assessment must identify the correct purchase-order line, preserve the revised delivery date, identify affected production requirements and escalate missing information. Record the planner’s checking and correction effort before acceptance.

FAIS would model those relationships in a business ontology: shared definitions of business objects, their connections and permitted actions. A supplier message refers to particular purchase-order lines; those lines supply materials needed by production requirements. Acceptance rules specify what an assessment must establish, when information is insufficient and who approves it.

FAIS would connect agents and automation to those definitions through suitable connectors to existing systems. Purchase orders can remain in the ERP and planning records in their current software, preserving that investment. Reusable object definitions and relationships provide a foundation for future automation without repeatedly rebuilding the business meaning. Any modelling changes would be established before the comparison and held constant across both settings.

When should you adopt the lower setting?

Adopt the lower setting only when it passes predefined quality and review limits and improves total cost per accepted outcome. Keep the existing setting when evidence is inconclusive; monitor outcomes and review burden after an approved change, consistent with continuous evaluation guidance.

Start with an existing automation whose outcomes you can verify and review effort you can record. Those acceptance rules give you a concrete operating decision to examine with FAIS.