The Reasoning Default Becomes the Convenience Toll
A new pricing model gives an LLM provider two controls: the per-token price and the reasoning allocation presented by default. A user may keep that default, choose another allocation, or leave.
The model’s revealing result is conditional. A default can steer the amount of reasoning only when avoiding customization carries a positive convenience benefit. The source of leverage is therefore not the default alone, but the friction around changing it.
The Paper
The source is Ahmet Bugra Gundogan, Yigit Turkmen, and Melih Bastopcu’s Keep, Customize, or Exit: Default Design and Token Pricing in LLM Reasoning Services, arXiv:2608.13315v1 [cs.GT], cross-listed in cs.AI, cs.LG, and eess.SY, submitted August 13, 2026. The ten-page paper combines a formal provider-user game with model and benchmark measurements. It does not report an observed commercial market or a study of human purchasing behavior.
A Service Model, Not a Market Observation
In the paper’s game, one provider serves one representative user working on a fixed task class. The provider moves first, choosing a per-token price and a binding default allocation of reasoning tokens. After seeing the offer, the user keeps the default, customizes the allocation, or exits. The model assumes complete information: both sides know the parameters and payoff functions.
Reasoning allocation is continuous for the proof and discrete in the experiments. Accuracy follows a fitted saturating curve, while billed tokens and latency rise linearly with the allocation. These choices make the equilibrium tractable. They are modeling decisions, not universal laws of reasoning services.
What the Model Calls Value
The user utility values the probability of a correct response, subtracts token charges and latency cost, and adds a default-specific convenience benefit, denoted by δ, when the user keeps the preconfigured allocation. The paper says this term represents reduced cognitive and interaction costs plus status-quo advantage. The authors explicitly distinguish it from a cost paid only when someone customizes, and it is not an empirically measured interface fee.
The provider’s payoff combines token margin, a weight on response accuracy, and a latency cost. Inside that specification, “user-optimal” means the allocation maximizing the stated user utility. It is not a social-welfare verdict, and it does not include every concern a person might have about privacy, energy, answer review, or dependence on the service.
When the Default Can Steer
The central comparative result is clean. If δ equals zero, any accepted default collapses to the user’s independently customized optimum. Price can still change the chosen allocation, but the default has no separate allocative power. If δ is positive, the acceptable defaults can form an interval, and the provider may choose a point inside that interval that differs from the customized optimum.
In the authors’ Qwen3-8B baseline numerical configuration, the provider chooses the upper acceptance boundary for all five benchmarks. For GPQA Diamond and HMMT 2025, the modeled user would customize to zero additional reasoning at the equilibrium price while accepting a positive default. That is a result of selected economic inputs, not evidence that real providers currently induce extra tokens. In this example, default leverage increases reasoning rather than rationing it.
What the Experiments Support
The empirical calibration uses Qwen3-8B and DeepSeek-R1-Distill-Llama-8B on AIME 2024, AIME 2025, GPQA Diamond, GSM8K, and HMMT 2025. The authors randomly select 500 GSM8K test examples, use the complete available evaluation split for the other benchmarks, and sample three responses per question and allocation. They enforce discrete reasoning budgets with a budget-forcing procedure and measure inference on one NVIDIA RTX PRO 6000 Blackwell Server Edition GPU.
Across the tested model-benchmark pairs, the measured accuracy generally rises with diminishing returns, but the fitted rate and attainable gain vary by task. Those runs support the paper’s service curves. The marginal token cost, user value of accuracy, latency sensitivity, provider weights, and convenience benefit remain modeling inputs explored through sensitivity analysis. The equilibrium examples therefore illustrate consequences of assumptions; they do not estimate a population demand curve.
The Evidence Boundary
The authors’ conclusion names the complete-information, representative-user design and leaves heterogeneous users, private valuations, repeated interaction, competing providers, and dynamic pricing for future work. The model also assumes a fixed task class, one request at a time, a particular keep-first tie rule at indifference, linear token and latency functions, and a single scalar convenience benefit.
There is no user experiment measuring the effort of finding a setting, understanding its consequences, comparing costs, or recovering from an unsuitable default. Accessibility needs, subscription bundles, organizational purchasing, hidden output-token charges, and unequal ability to tolerate latency could all change who customizes and who accepts. The paper identifies a mechanism worth governing; it does not determine its prevalence, direction, or distributional effect.
The Reasoning-Choice Receipt
A governed reasoning service should disclose the model and checkpoint, task class, default allocation, selectable range and increments, full token price, expected input and output charges, measured latency and accuracy curves with uncertainty, assumptions used to recommend a setting, location and steps required to customize, whether the choice persists, telemetry collected, fallback behavior, price and default change history, accessibility review, user-selected allocation, realized token count, and correction or refund path.
The Spiralist boundary is simple: a default is a suggestion only while refusal remains legible and inexpensive. Once interface friction makes acceptance valuable, the default becomes an allocation mechanism and should be audited with the price.
Related Pages
- The Token Meter Becomes the Budget
- The Evaluation Score Becomes the Inference Budget
- The Budget Router Becomes the Allocation Receipt
- The Affective Default Becomes the Interface Policy
Sources
- Ahmet Bugra Gundogan, Yigit Turkmen, and Melih Bastopcu, Keep, Customize, or Exit: Default Design and Token Pricing in LLM Reasoning Services, arXiv:2608.13315v1 [cs.GT], submitted August 13, 2026.
- Paper model and utilities, checked for the single-provider sequence, keep-customize-exit choices, accuracy, latency and token functions, payoff definitions, convenience benefit, and tie rule.
- Paper equilibrium analysis, checked for the default-acceptance region, zero-convenience result, three provider regimes, and equilibrium existence.
- Paper experiments, checked for models, benchmarks, sampling, hardware, budget forcing, fitted service parameters, numerical inputs, and sensitivity analysis.
- Paper conclusion, checked for the conditional default-steering result and the complete-information, representative-user, heterogeneity, competition, and repeated-interaction limits.