acceptodds
Withdrawn from ICLR 2027

Strategic Pricing in LLM Routing Markets

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Routing systems for large language models choose among models and collaboration modes to balance answer quality against cost. They take supplier prices as given. When suppliers set prices strategically, the router defines a game with two difficulties. Its equilibrium need not be pure, and a single trained router can reward a supplier for raising its own price. We compute equilibria by fictitious play and accept a profile only after checking every unilateral deviation on the bid grid. We learn answer quality once, independently of price, and add payments through an economic layer, so a supplier's demand cannot rise with its own price. We then study two platform controls. A committed outside option gives the customer side bargaining power. Under a router that picks the best mode, it raises customer surplus whenever its committed value exceeds baseline surplus, and repricing cannot remove that gain. In simulations with model accuracies calibrated to three open-weight models, its benefit rises with its capability and falls with its price. A capability-blind near-tie rule changes of allocations at a moderate threshold and raises the least-used supplier's share in every run, at a bounded direct cost. Menu expansion buys contestability. Redistribution within the menu buys inclusion.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
large language model routing, Bayesian Nash equilibrium, algorithmic pricing, AI service markets

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_qpktOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies strategic supplier pricing in LLM routing markets, where suppliers submit prices under a fixed routing rule that selects among single-model and collaborative execution modes. It recovers approximate mixed Bayesian Nash equilibria using fictitious play and certifies them through exhaustive unilateral-deviation checks on a finite bid grid. The paper also proposes a price-independent quality predictor coupled with an explicit economic allocation layer that guarantees non-increasing expected billable demand with respect to a supplier's own price. Two platform interventions are investigated: a committed outside option (anchor) and near-tie redistribution. The evaluation uses simulations calibrated to three open-weight models and examines equilibrium residuals, customer surplus, supplier allocations, and sensitivity to selected market parameters.

Critical strengths
  1. Economically coherent routing architecture. Separating quality prediction from payment accounting provides an explicit own-price monotonicity guarantee for expected billable demand while accommodating learned quality estimates. This is a useful design principle for analyzing strategic pricing without confounding price changes with economically inconsistent routing responses.

  2. Careful distinction between equilibrium computation and certification. The paper directly checks unilateral deviations rather than treating fictitious-play convergence as evidence of equilibrium. Its explicit restriction of the certificate to the finite-grid game, together with the decomposition of intervention effects into direct allocation and strategic-response components, makes the numerical conclusions more interpretable.

Critical weaknesses
  1. Limited theoretical novelty and evidence for competition-driven gains. The strong-anchor theorem follows directly from introducing an outside option whose standalone value already exceeds baseline customer surplus. It does not characterize when the anchor improves outcomes through competitive repricing. In Table 7, nearly all gains under the hard, soft, and structured routers come from the direct allocation effect, leaving the role of strategic price competition unclear.

  2. Limited empirical validation. The experiments rely on a parameterized simulator. While single-model abilities are calibrated to benchmarks, it is unclear how query-level chain and ensemble accuracies are constructed or how correlations between model errors are handled. The paper also lacks quantitative held-out evaluation of the structured quality predictor. Moreover, all feasible bids exceed all private costs, excluding pricing near marginal cost.

Questions for authors
  1. How are query-level chain and ensemble accuracies constructed from the single-model benchmark calibration, including dependencies between model errors?

  2. Additional evidence needed for the current claims: Do the main intervention findings persist when the bid grid permits prices near marginal cost and is refined beyond the reported configurations?

Reviewer_5MRaOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies strategic pricing among LLM providers competing through a routing platform. It models provider bidding as a Bayesian game, computes approximate mixed equilibria using fictitious play, and examines two interventions: introducing a committed outside provider and redistributing traffic among nearly equivalent routing options. The results are evaluated primarily through simulations calibrated using three open-weight models.

Critical strengths

The paper considers a relevant question about pricing incentives in LLM routing markets. Separating model-quality prediction from pricing is a reasonable design choice, and the authors provide equilibrium diagnostics rather than relying solely on solver convergence.

Critical weaknesses
  1. Limited novelty and theoretical contribution.

The central research gap is not convincing. Strategic provider pricing has already been studied in LLM routing and modular API marketplaces, including Guo et al. (2026) and Bhawalkar et al. (2025). The paper does not clearly establish what substantial new economic or methodological insight its formulation provides.

The theoretical results are also rather elementary. Theorem 1 follows directly from the existence of mixed equilibria in finite games. Theorem 2 states that introducing an outside option whose standalone value exceeds baseline customer surplus guarantees an improvement, which follows immediately from the allocation rule. Proposition 1 is a standard softmax derivative identity, and Proposition 2 bounds the loss from reallocating among near-equivalent choices by their maximum value difference.

Although these statements are valid under their assumptions, I do not see sufficient theoretical depth or novelty for the claimed contribution.

  1. Artificial experimental setup and weak practical significance.

The empirical evidence is entirely simulation-based, with market parameters largely imposed rather than estimated. The main experiments use only three strategic suppliers with identical list prices, fixed token counts, discretized cost distributions, and bids restricted to be profitable by construction.

The anchor intervention is particularly unconvincing as evidence of a new mechanism. The strongest anchor has substantially higher assumed capability than the incumbents, so its improvement in customer surplus is unsurprising. Moreover, the results show that most of the gain under the price-coherent routers comes directly from adding a better option, rather than from changes in strategic pricing.

The equilibrium conclusions are also sensitive to discretization: the number of mixed supplier–cost cells changes substantially across five-, seven-, and nine-point bid grids. The certificates establish approximate equilibrium only on the corresponding grids, not in the continuous market.

Finally, the near-tie intervention produces very small improvements in supplier inclusion, with no convincing demonstration of meaningful practical benefits.

The presentation further weakens the paper. The writing is repetitive, the organization makes the central contribution difficult to identify, and several figures have very small labels and dense captions. The paper devotes considerable space to standard theoretical observations without developing a sufficiently clear research story.

Questions for authors
  1. What is the substantive theoretical novelty beyond existing strategic-pricing models, particularly Guo et al. (2026) and Bhawalkar et al. (2025)? Which results could not be obtained from standard auction theory and outside-option arguments?

  2. Why should the simulated market be representative of real LLM provider competition? Can the authors demonstrate that their conclusions hold with heterogeneous provider prices, realistic cost distributions, and actual provider behavior?

  3. How much of the reported customer-surplus improvement comes from the proposed strategic mechanisms rather than simply adding a higher-quality option? Can the authors establish meaningful incremental benefits under controlled comparisons?

Guo et al. (2026). Pricing Online LLM Services with Data-Calibrated Stackelberg Routing Game.

Bhawalkar et al. (2025). Equilibria and Learning in Modular Marketplaces.

Varian (1980). A Model of Sales.

Reviewer_iq3BOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper discusses pricing strategies for query routers in an LLM market. Authors model a first-price auction design where the router picks a collaboration mode among offering a single model or a combination of their responses. They then simulate interactions between the router and LLM providers via fictitious play and provide a characterization of the fixed points they find under a Bayesian Nash Equilibrium concept. The paper proposes an anchor pricing strategy as a non-strategic outside option to displace the incumbents in the market.

Critical strengths

The question of mechanism design (and in particular market design) for LLMs for routers is quite timely. I also appreciated the idea of the anchor option as an intervention (although I have reservations about how realistic that is).

Critical weaknesses

I found the paper rather unreadable. Very frequently, in the main matter, the authors give a 2-sentence description of a subsection in the appendix. This includes major results. For example, Theorem 2 on page 5 reads

Theorem 2 (Sufficient strong-anchor condition). Suppose the augmented router uses the hard allocation in Eq. equation 37, every baseline mode is retained, and the anchor's capability and bid are fixed. If 1{ }^1, CSI (b0)\left(b^0\right), then the anchor raises equilibrium customer surplus. More precisely, CS⁡D(bD)−CS⁡0(b0)≥vD−CS⁡(b0)>0.\operatorname{CS}^D\left(b^D\right)-\operatorname{CS}^0\left(b^0\right) \geq v_D-\operatorname{CS}\left(b^0\right)>0 .

(and yes, the "Eq. equation 37" is reproduced verbatim). Equation 37 appears on page 18 in Appendix C.5. This is just an example; there are many more to the point where I cannot parse section 3 onwards without bouncing back and forth between formal statements in the appendix and the claim in the main matter.

This is not a simple writing issue. The paper's main matter, as it stands, is not self-contained at all. All major formalizations and algorithms appear in the appendix.

There are, of course, many other typographical, writing, and organizational issues. To name a few:

  • Bayesian Nash Equilibrium suddenly pops out as the notion of equilibrium in Section 3.3. Why? Where did this come from? Why is this a Bayesian game? What is the type? prior. etc?
  • 3.1 "Notation" paragraph is a table of notation compressed into a paragraph and makes no sense since most of the terms are not defined at this stage to be notarized.
  • There are empirical evaluations in the middle of the method presentation in Section 5.2: "On the discretized grid, the hard rule, both soft-rule temperatures, and the structured router pass all 144 adjacent ownprice comparisons in both markets. The learned cache passes 101 of 144 without the anchor and 108 of 144 with it."

As a result, this paper, in my opinion, does not pass the basics of good scientific writing for even a 2nd tier conference, let alone ICLR. I recommend that authors withdraw the paper and focus their technical presentation on a standalone subset of the results in the appendix.

Questions for authors

I have no questions that would materially affect my assessment of this paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.