Home · Writing · Research

Test-Time Compute in the Enterprise: When Reasoning Tokens are Worth Money

Spending more compute at inference time, rather than only at training time, is now a deliberate lever an enterprise team can pull. It is also a lever with a real invoice attached, and the decision to pull it should be made the same way any other spend decision is made: against the value of the error it prevents.

TLDR

  1. Spending more compute at inference time, rather than only at training time, is now a deliberate lever an enterprise team can pull.
  2. Model choice used to be the main inference-cost lever: use a smaller model for routine work and a larger one for harder tasks.
  3. Best-of-n and self-consistency take a different route: generate several candidates and select one through voting or a verifier.
  4. Measure the latency distribution on the chosen region, model and prompt. A batch settlement check may tolerate a longer answer if accuracy gains justify the cost.
  5. The raw scenario produces 4,240 errors at 89.4 percent and 1,320 at 96.7 percent. That subtraction is not yet value.
Figure 1Incoming query to final answerCausal and control schematic
Incoming query to final answer6 declared states connected by 4 authored relations. The figure supports the section Reasoning tokens and the pricing mechanics. L0L1 01
Incoming query
02
Difficulty classifier
03
Low thinking budget
04
Medium thinking budget
05
High thinking budget
06
Final answer
Reading. The authored topology makes 4 declared relations across 6 states inspectable. Read it as the control structure for “Reasoning tokens and the pricing mechanics”, not as measured performance. Schematic derived from the paper's authored topology; no measured quantities.
On this page

A new variable on the cost sheet

Model choice used to be the main inference-cost lever: use a smaller model for routine work and a larger one for harder tasks. Test-time compute adds another lever. A team can vary the work spent on an answer through a reasoning budget, multiple samples or iterative refinement. The same model can therefore occupy different points on a cost-quality curve from one request to the next.

This creates an engineering opportunity and a new way to overspend. A larger budget can consume much more output-side compute for the same nominal task. The multiplier varies by model, configuration and response, so it must be measured rather than repeated as a universal ratio. This article develops a task-level decision rule for when the extra spend is justified.

The research case for spending more at inference time

Inference-time scaling research shows that additional computation can improve accuracy on multi-step reasoning tasks. The gain is usually larger on hard problems than on easy ones. That uneven benefit is the enterprise result that matters. If additional compute adds little to routine cases, a uniform high-budget policy wastes money and time.

Best-of-n and self-consistency take a different route: generate several candidates and select one through voting or a verifier. Cost grows with the number of candidates. The choice between a longer single pass and multiple samples remains task-dependent. In either case, added compute is useful, costly and unevenly valuable across the workload.

Reasoning tokens and the pricing mechanics

Reasoning-capable Gemini models expose controls for thinking, but the supported parameter and pricing depend on the model version. Reasoning or thought tokens can contribute to billable output and latency even when the visible answer is short. Use the response usage metadata and the current Vertex AI pricing page to calculate the actual unit cost. Do not infer it from visible words.

The available control may be an explicit budget or a qualitative thinking level. Test the options supported by the exact model in use. Plot accuracy, cost and latency together, then find where additional compute stops buying a material gain. The flattening point is specific to the task, model version and evaluation population.

The latency side of the ledger

Latency deserves equal weight with cost. A few extra cents may be tolerable while several extra seconds are not, especially when an employee or client is waiting. An accuracy gain has no value if the user abandons the interaction before reading it.

Measure the latency distribution on the chosen region, model and prompt. A batch settlement check may tolerate a longer answer if accuracy gains justify the cost. An interactive assistant needs an independent latency ceiling. The same underlying task can therefore require a different policy in synchronous and asynchronous channels.

Adaptive routing therefore needs two axes: consequence and interaction mode. A synchronous, human-facing request has a hard latency ceiling. An asynchronous request can trade more time for quality. Do not apply the same reasoning policy to live chat and overnight batch merely because both call the same service.

When reasoning tokens are worth the money

The decision framework compares the expected value of errors prevented with the marginal inference cost at production volume. Error cost is a distribution, not a single number. It ranges from a correction by a downstream reviewer to an irreversible financial action or compliance failure. Extended compute is worth buying when the task is difficult, the consequence is material and the measured improvement moves the error rate enough to matter.

A worked settlement comparison makes the method concrete. The task checks a proposed instruction against trade and counterparty data before release. Errors can have material financial and operational consequences, while the task requires several conditional reconciliations. The scenario assumes that a higher reasoning setting lifts accuracy from 89.4 to 96.7 percent on 500 labelled cases. At forty thousand monthly instructions and $0.014 incremental cost, the uniform high setting adds $560. These are calculation inputs, not reported production results; replace every one with local evidence.

When it is not

The mirror case is a high-volume FAQ classifier routing queries to well-separated categories such as billing, password reset or product information. The worked comparison assumes 97.8 percent accuracy at a minimal reasoning budget and 98.6 percent at a high budget. Misclassification is usually reversible at the next step. At millions of queries a month, uniform high-budget reasoning can add tens of thousands of dollars for less than one percentage point of modeled improvement. The figures are scenario inputs; the decision method is to measure the local accuracy-cost curve and reserve extended reasoning for cases where it changes a consequential decision.

Task characteristic Extended reasoning value Recommended default
High error cost, genuinely hard task, accuracy gain measured large High Extended thinking budget by default
High error cost, task is actually easy once measured Low Minimal budget, invest in verification instead
Low error cost, high volume, task is easy Very low Minimal budget always
Low error cost, but task is genuinely hard Moderate Adaptive, based on measured confidence

Adaptive test-time compute, routing by difficulty

An adaptive policy is often more efficient than either uniform extreme. Route each query from a cheap difficulty or consequence signal before running the expensive step. The signal may be a small classifier, calibrated confidence from a first pass, a familiar-versus-novel pattern check, or a business variable such as notional value. If a business rule already captures consequence, do not ask another model to infer it.

In the worked settlement design, low-notional instructions that match known patterns take the minimal path. Higher-notional or unfamiliar instructions take the high path. The policy should be accepted only after comparing its weighted accuracy, tail errors and blended cost with the uniform alternatives. Report confidence intervals rather than presenting a small point difference as certainty.

Figure 2Settlement instruction to validated instructionCausal and control schematic
Settlement instruction to validated instruction5 declared states connected by 3 authored relations. The figure supports the section Adaptive test-time compute, routing by difficulty. L0L1 01
Settlement instruction
02
Notional above threshold or unfamiliar pattern
03
Minimal thinking budget
04
High thinking budget
05
Validated instruction
Reading. The authored topology makes 3 declared relations across 5 states inspectable. Read it as the control structure for “Adaptive test-time compute, routing by difficulty”, not as measured performance. Schematic derived from the paper's authored topology; no measured quantities.

Measuring return on investment properly

A recurring mistake is measuring return only as percentage-point accuracy. Convert the improvement into avoided consequence that can be compared with inference cost. The estimate will be imperfect, but the exercise forces a concrete conversation about which errors matter.

The raw scenario produces 4,240 errors at 89.4 percent and 1,320 at 96.7 percent. That subtraction is not yet value. Downstream controls catch most errors before settlement, so the relevant number is the reduction in errors that would pass every other control. Multiply that residual by the institution's historical incident cost and compare it with the $560 inference increment. The residual error path, not headline accuracy, justifies the spend.

Planning for a moving target

Pricing and thinking controls change across model versions. A cost model tied to one version can become obsolete while the business process remains unchanged. Date every unit-cost assumption, name the model and region, and mark the estimate as provisional. Do not hard-code a current price into the design principle.

Recalibrate on every model-version change. Rerun the frozen evaluation set, locate the new flattening point and update the return calculation. Accuracy and price can move at the same time, so validating only one side can reverse the decision.

Present the business case as a range tied to a dated and sourced unit-cost assumption. Use a model-version change as an automatic review trigger rather than waiting for a calendar date or a surprising invoice.

Failure modes

The first failure mode is a generous uniform budget inherited from a reference implementation. It persists because the symptom is a rising bill rather than a visible incident. Make reasoning configuration part of the release review and expose it in cost telemetry by task.

The second is an evaluation set dominated by memorable hard cases. It overstates the value of extended reasoning across the full population. Use a representative sample for the blended decision and retain a separate challenge set for tail risk. Do not confuse the two.

The third is drift in the difficulty signal itself. New products or instruction formats can make a once-calibrated router misclassify hard work as easy. Monitor routing accuracy and the error rate within each route. A blended metric can hide deterioration in the smaller hard tail.

Worked example, underwriting triage with an adaptive reasoning budget

The worked underwriting scenario triages small and medium enterprise applications. It checks completeness, flags risk indicators and recommends either straight-through processing or human review. The volume assumption is 8,500 applications a month, of which 70 percent match well-understood industry and coverage patterns. These inputs exist to demonstrate the calculation and must be replaced during discovery.

A fast pre-classifier compares the industry code, requested limits and selected fields with a maintained set of straightforward patterns. Matches take the minimal path; everything else takes the high path. The classifier is biased toward escalation because unnecessary compute is cheaper than false confidence on a complex application.

Figure 3Application to human underwriterInteraction sequence
Application to human underwriter5 declared states connected by 5 authored relations. The figure supports the section Worked example, underwriting triage with an adaptive reasoning budget. t
Application
Pre classifier
Minimal budget triage
High budget triage
Human underwriter
01
Structured fields
02
Route
03
Recommend straight through
04
Route
05
Recommendation, cited evidence and concise rationale
Reading. The authored topology makes 5 declared relations across 5 states inspectable. Read it as the control structure for “Worked example, underwriting triage with an adaptive reasoning budget”, not as measured performance. Schematic derived from the paper's authored topology; no measured quantities.

For illustration, the router sends 64 percent of applications to the minimal path. On a 400-case held-out set, assume 95.1 percent accuracy at $0.021 blended cost, compared with 96.0 percent and $0.048 for a uniform high policy. The adaptive option saves 56 percent of inference cost for a 0.9-point accuracy difference. The approval question is where those difference cases sit: harmless routing friction or materially under-assessed risk.

The scenario also tests a second-order benefit. If the minimal path returns in 1.4 seconds and the high path in 4.8, routine cases clear the queue sooner. Underwriters can spend more attention on the complex minority. Include that human-flow effect in the trial rather than limiting the business case to token cost.

Turn the budget into an executable policy

A reasoning budget should be versioned beside the prompt, tool set and model. Leaving it as an SDK default makes a material cost and risk choice invisible. The policy needs five inputs: task class, consequence tier, evidence completeness, interaction deadline and the measured shape of the task's cost-quality curve. None of those inputs requires the reasoning model to judge itself.

The route can be deterministic when the business already knows the consequence. A payment release above a threshold, a vulnerable-customer indicator or a sanctions match should not depend on a model estimating “difficulty.” Evidence completeness can also be tested through required-field and retrieval-coverage rules. A learned difficulty classifier belongs only in the residual, where the hard cases cannot be identified reliably from business state.

Figure 4Case arrives to outcome and usage recordCausal and control schematic
Case arrives to outcome and usage record11 declared states connected by 6 authored relations. The figure supports the section Turn the budget into an executable policy. L0L1L2 01
Case arrives
02
Read task and consequence tier
03
Required evidence complete
04
Request evidence or human review
05
Deterministic high-consequence rule
06
Verified high-compute path
07
Difficulty router
08
Minimum viable compute
09
Moderate compute plus checks
10
Independent verification
11
Outcome and usage record
Reading. The authored topology makes 6 declared relations across 11 states inspectable. Read it as the control structure for “Turn the budget into an executable policy”, not as measured performance. Schematic derived from the paper's authored topology; no measured quantities.

The model must not be allowed to buy unlimited compute. Set a hard ceiling in the invocation policy and a trajectory ceiling across retries, candidates and verifier calls. A limit on one response does not control a loop that can call the response repeatedly. Record both figures so finance, service management and model risk see the same unit of consumption.

Calibrate the router, not just the reasoner

The router creates its own false-negative risk: a difficult case sent to the cheap path. Its evaluation set therefore needs more than overall accuracy. Measure hard-case recall, calibration and the consequence-weighted false-negative rate. A router that labels nearly everything hard may look safe but destroys the economics. One that labels nearly everything easy may look efficient while concentrating errors in the residual.

Use a three-way adjudication label during calibration: routine, compute-sensitive and human-dependent. The third class matters. Some cases do not improve with more model computation because the missing ingredient is external evidence, authority or judgement. Routing those cases to a larger budget merely produces a longer statement of uncertainty.

Router outcome Actual case Immediate effect Required control
routine route routine intended saving sampled regression monitoring
high-compute route compute-sensitive intended quality spend marginal-gain and latency ceiling
routine route compute-sensitive false economy hard-case recall threshold and sentinel set
high-compute route routine wasted cost route-rate and unit-cost alert
any compute route human-dependent polished non-resolution explicit abstention and evidence request

Calibrate using cases sampled from the traffic distribution and a separate tail set built from incidents, overrides and near misses. The representative set estimates blended economics. The tail set tests whether the router recognises danger. Combining them into one headline score conceals both questions.

Build a marginal-value curve

One high setting and one low setting are insufficient. Run at least several permitted budgets across the same frozen cases and preserve every output. For each task slice, plot consequence-weighted error, p50 and p95 latency, billable usage, abstention and verifier rejection. The useful point is where the next increment stops changing the decision or the checked outcome.

Figure 5Frozen labelled cases to approved routing policyCausal and control schematic
Frozen labelled cases to approved routing policy8 declared states connected by 10 authored relations. The figure supports the section Build a marginal-value curve. L0L1L2L3L4 01
Frozen labelled cases
02
Budget level 1
03
Budget level 2
04
Budget level 3
05
Budget ceiling
06
Independent adjudication
07
Cost, latency and error curves by slice
08
Approved routing policy
Reading. The authored topology makes 10 declared relations across 8 states inspectable. Read it as the control structure for “Build a marginal-value curve”, not as measured performance. Schematic derived from the paper's authored topology; no measured quantities.

The curve should use decisions, not prose quality. A longer answer may receive a better style rating without changing the proposed action, calculation or cited evidence. For a credit-refer-back task, score whether the right missing item was requested and whether the reason matches policy. For an investigation task, score entity resolution, evidence coverage and disposition routing. More tokens count as value only when they improve a checked decision.

Uncertainty belongs on the curve. Bootstrap intervals or repeated adjudication can show whether a one-point gain is signal or sampling noise. That is especially important for the difficult tail, where case counts are small and apparent gains swing sharply.

Operate it as a closed loop

The live control loop joins model telemetry, business outcome and cost. Usage metadata records what the service consumed. The trace records the route, evidence identifiers, tool and policy events, verifier results and a concise structured rationale for the proposed decision. It does not need, and should not claim to expose, hidden chain-of-thought. The business system records whether a reviewer corrected, overrode or accepted the outcome. A daily or weekly process joins those records by a case and trajectory identifier.

Figure 6Published budget policy to recalibrate, restrict or roll backCausal and control schematic
Published budget policy to recalibrate, restrict or roll back9 declared states connected by 9 authored relations. The figure supports the section Operate it as a closed loop. L0L1L2L3L4 01
Published budget policy
02
Live inference
03
Usage and latency telemetry
04
Trajectory and verifier evidence
05
Business outcome and override
06
Joined task-level ledger
07
Slice and marginal-value monitoring
08
Policy still inside bounds
09
Recalibrate, restrict or roll back
Reading. The authored topology makes 9 declared relations across 9 states inspectable. Read it as the control structure for “Operate it as a closed loop”, not as measured performance. Schematic derived from the paper's authored topology; no measured quantities.

Three changes should trigger recalibration without waiting for a quarterly review: a model version change, a material traffic-mix change and a shift in the router's hard-case rate. A pricing change may alter the economic threshold even when quality remains stable. A new product or document format may alter difficulty even when the model is unchanged.

What the finance view should contain

Finance does not need token mechanics without a business denominator. The monthly view should report cost per successfully verified task, cost per material error avoided and the proportion of spend assigned to each consequence tier. Separate base model cost, added reasoning, candidate generation, verifier calls and failed retries. A single “LLM cost” line makes an inefficient policy hard to diagnose.

Model the range, not only the mean. At minimum, show expected, high-volume and adverse-routing scenarios. The adverse case assumes more work reaches the high-compute path and more calls retry. That case often determines whether a service remains viable during an incident or product launch.

Test-time compute is a controllable operating expense only when route, budget and outcome are joined. Without that ledger, the enterprise has bought an invisible quality setting whose benefit cannot be defended and whose cost cannot be allocated.

Service levels and chargeback

The compute policy should sit inside the service-level design. A synchronous employee assistant might reserve most of its latency budget for retrieval and the first useful answer, with extended reasoning available only after the interface acknowledges a longer-running task. A batch control can spend minutes if it completes before the operational deadline. An event-driven investigation can pause for evidence without holding an inference session open. One business capability may therefore need several channel-specific budgets.

Internal chargeback should follow the task and consequence tier. Charging only by application hides a common pattern: one small high-compute route consumes more than a large routine route. Report successful verified tasks, high-budget route share, retries and human-dependent abstentions. The product owner can then see whether cost rose because volume changed, cases became harder or a release altered routing.

Do not turn chargeback into a reason to suppress legitimate escalation. The cost target belongs beside a hard-case recall target and a customer or risk outcome. A team that minimizes high-budget routing can produce a good invoice and a bad control. Balanced ownership requires finance, product and risk to approve the same policy version.

Capacity planning needs the tail of the budget distribution. Peak demand may coincide with unusual cases, which means both volume and high-compute route share rise together. Stress the joint condition. Reserve enough concurrency for priority cases, queue non-urgent work and define what degrades first. The system should reduce optional candidate generation before it removes required verification.

When an external model provider changes throughput or quotas, the policy needs a deterministic fallback. Options include a smaller approved budget, a different approved model, an asynchronous queue or human handling. The fallback should be evaluated as a product state, not written during an incident. A reasoning budget without a tested scarcity mode is not an operating policy.

Retain the usage and outcome record long enough to compare policy versions across seasonal cycles, within the institution's data-retention rules. A one-week trial may miss month-end peaks, annual reviews or product launches that change both case mix and consequence. Periodic comparison should therefore separate policy improvement from traffic change. If the high-compute route grows, the owner must be able to show whether customers became harder, the router drifted or the release moved its boundary.

That explanation belongs in the same monthly control pack as service performance, cost and material-error reporting, so corrective action follows one agreed account of the change.

It should also name the decision owner, effective date, approved fallback, evidence window and next review trigger. That turns a dashboard observation into a governed change rather than another unexplained cost variance.

Notes for practitioners

Treat test-time compute as a budget decision for each distinct task, not a default inherited from a reference design. Measure the gain on a representative production sample and keep the hard-case challenge set separate. Otherwise the evaluation will overstate the value of spending more on routine work.

Convert accuracy gain into avoided consequence and compare it with marginal cost at production volume. Use historical incident data where possible. Prefer an adaptive policy when a cheap, reliable consequence or difficulty signal exists. Then monitor the router as a model in its own right; quiet drift in that first decision can erase the value of the whole design.

A budget decision matrix

Task condition Compute policy Required evidence Stop condition
easy and low consequence minimum budget stable baseline accuracy no material slice regression
hard but verifiable larger budget or best-of-n verifier precision and marginal gain gain falls below marginal cost
hard and subjective bounded refinement plus human review reviewer agreement and calibration extra tokens add verbosity, not decisions
high consequence with weak verification human-led failure-mode coverage no autonomous path

Reasoning tokens are a spend category, not a quality certificate. The router needs its own benchmark. The verifier needs false-acceptance tests. The budget should fall when the evidence says more computation has stopped helping.

Sources and current-service check