Token Inflation: How Dishonest Providers Can Overcharge for Large Language Model Usage
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Security Radio. Generated commentary on the latest security and cryptography papers.
Nadia: We're looking at the paper: "Token Inflation: How Dishonest Providers Can Overcharge for Large Language Model Usage".
Elias: Here's the gist of the paper. Per-token billing is now the standard pricing model for commercial large language models (LLMs), so the honesty of reported token counts directly affects what users pay Page 1>.
Nadia: The paper demonstrates that current auditing frameworks are structurally flawed because they rely on artifacts controlled by the provider, leading to systematic overcharging through low-cost manipulation Page 2>.
Elias: Structural Diagnosis of Auditing Frameworks The core finding is a "trust paradox" where every audit must trust some artifact, but current frameworks trust exactly the ones a financially motivated provider is incentivized to manipulate Page 2>.
Nadia: This diagnosis applies across three recent auditing frameworks: CoIn Sun et al.
Title and authors: Nadia: We’re talking about this paper called "Token Inflation: How Dishonest Providers Can Overcharge for Large Language Model Usage" and its authors, Shahinul Hoque, Jinghuai Zhang, Jinyuan Sun, and Fnu Suya >
Elias: I read the introduction again. It sets up this whole problem by saying that per-token billing is now the standard way people pay for using these large language models >
Priya: And they immediately point out that because providers hide their model, tokenizer, and execution to protect things like intellectual property or user privacy, an auditor can only look at artifacts the provider gives them >
Nadia: It’s like the audit is forced to be a consistency check on reports the provider already made instead of checking what actually ran >
Elias: They call this a trust paradox because every audit has to trust some artifact, but currently, those frameworks trust exactly what a financially motivated provider has the strongest reason to manipulate >
The paper's summary: Nadia: The paper goes on to study three specific auditing frameworks and their weaknesses, and they show that producing evidence that looks honest is actually cheap for a provider trying to overcharge >
Elias: They give us some concrete numbers. For instance, in the CoIn framework, a provider can inflate reported token counts by up to one thousand four hundred sixty-nine percent on average without anyone detecting it in the most permissive setting > <ref:2605.30040#pg2,inflate reported token counts by up to>
Priya: That level of inflation is massive. It means if you rely only on a consistency check, you could be looking at usage that is over fifteen times what's actually happening >
Nadia: And in the PALACE setting, where the auditor trusts data supplied by the provider, inflation can reach up to two hundred forty-seven percent >
Elias: They also look at the Velasco et al. framework and found that even when a user can see the full reasoning string, just tokenization ambiguity alone allows for over reporting of fifty point eight five percent below detection thresholds >
The paper's improvements: Nadia: So what’s the fix they suggest? It seems like they argue that we need to move away from these consistency checks over artifacts and towards things that tie the reported token counts to evidence the provider doesn't control >
Elias: That means we need something like trusted execution attestation, cryptographic proofs of inference, or third-party re-execution to prove what was actually computed >
Priya: I wonder if that really works. The paper mentions alternative solutions like trusted execution environments or cryptographic proofs of computation coupled to the reported token count >
Nadia: The authors caution that those mechanisms apply to the stated computation and counter, not necessarily to the provider’s actual computational cost, which is a key limitation they point out >
Elias: They conclude that exposing more artifacts isn't enough if those artifacts are still under the provider's control >
Conclusion: Nadia: To wrap up, the main point is that current token-count auditing frameworks give limited assurance because they rely on provider-controlled stuff instead of independent proof of computation >
Elias: The paper "Token Inflation: How Dishonest Providers Can Overcharge for Large Language Model Usage" shows that reliable billing accountability needs to move from consistency checks over artifacts to provider-independent evidence of the actual work done >
Priya: I think what this means for us is that we have to demand mechanisms like trusted execution environments or cryptographic proofs of computation coupled with the reported token count >
Nadia: That’s right. We need those things to bind the reported counts to proof that the provider didn't shape, which is a big shift for how we think about verifying these systems.
Shahinul Hoque, Jinghuai Zhang, Jinyuan Sun, Fnu Suya
University of Tennessee, Knoxville · University of California, Los Angeles
cs.CR, cs.AI, cs.CL
Submitted: 2026-05-28
Updated: 2026-11-07
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 83/100
The gist: Per-token billing is now the standard pricing model for commercial large language models (LLMs), so the honesty of reported token counts directly affects what users pay Page 1>.
Key concepts
- Structural Trust Gap
- This is the core problem where auditing frameworks must trust artifacts that a financially motivated provider is incentivized to manipulate. Because audits rely on data the provider controls, they cannot independently verify if the reported token count accurately reflects the true computational cost.
- Provider-Controlled Artifacts
- These are pieces of data or reasoning strings supplied by the LLM service that are under the provider's control. Auditing frameworks currently check consistency against these artifacts, but since a provider can shape or reuse these artifacts, this check is easily gamed to hide inflated token counts.
- Provider-Independent Evidence
- This is the necessary future standard for billing accountability. Instead of checking the provider's output or reasoning strings, this requires external proof that the specific computation actually occurred. Examples include trusted execution environments or cryptographic proofs that bind the reported token count to an undeniable, verifiable process.
- Martingale Statistics
- This is a statistical tool used in auditing to detect anomalies over time. The paper shows providers can exploit 'natural slack' in these statistics—using small compensating underreports on benign samples—to hide significant inflation patterns that would otherwise be detected by the audit.
Terminology
Summary
Per-token billing is now the standard pricing model for commercial large language models (LLMs), so the honesty of reported token counts directly affects what users pay Page 1>. p. 0 The paper demonstrates that current auditing frameworks are structurally flawed because they rely on artifacts controlled by the provider, leading to systematic overcharging through low-cost manipulation Page 2>. p. 9
Structural Diagnosis of Auditing Frameworks
The core finding is a trust paradox
where every audit must trust some artifact, but current frameworks trust exactly the ones a financially motivated provider is incentivized to manipulate Page 2>. This diagnosis applies across three recent auditing frameworks: CoIn Sun et al. p. 1 [2025a], PALACE Wang et al. p. 1 [2025], and a martingale-based statistical auditor Velasco et al. p. 1 [2025a] Page 1>. p. 1 The authors argue that the same constraints that force providers to hide their computation also prevent the auditor from collecting evidence independent of the provider, so audits end up trusting exactly the artifacts a financially motivated provider has reason to manipulate Page 2>. p. 1
Empirical Attacks Against Specific Frameworks
The paper details systematic adversarial analysis against three frameworks, showing that producing evidence that looks honest is cheap for a provider that wants to overcharge Page 2>. For the CoIn framework, a provider can inflate reported token counts by up to 1,469% on average without detection in the most permissive setting Page 1>. p. 0 In the PALACE setting, inflation can reach 247% when the auditor trusts data supplied by the provider Page 2>. p. 1 Even when the user can see the full reasoning string, tokenization ambiguity alone allows for 50. p. 085% over-reporting below the detection threshold in Velasco et al. p. 0's framework Page 1>.
Mechanisms of Inflation Across Frameworks
The attacks exploit different vulnerabilities in each auditing mechanism Page 2>.
-
Against CoIn, a provider can reuse or fabricate semantically plausible reasoning blocks to inflate reported reasoning length by 1,469% on average without triggering detection Page 3>. p. 1
-
Against PALACE, a provider can shift the auditor’s estimate by rewriting the answer style or appending trigger tokens, and exploit reliance on training data to inject poisoning points for targeted backdoors Page 4>. p. 1
-
Against Velasco et al. p. 1's martingale-based auditor, tokenization ambiguity and selective inflation still permit 50.85% over-reporting even when the user can see the full reasoning string Page 1>. p. 0
Requirements for Future Verification
Restoring honest billing requires verification that ties reported token counts to evidence the provider does not control, such as trusted execution attestation, cryptographic proofs of inference, or third-party re-execution Page 1>. The authors conclude that exposing more artifacts is not sufficient when those artifacts remain under provider control Page 1>. p. 1
Alternative Solutions and Limitations
Alternative solutions proposed include trusted execution environments (TEEs) or cryptographic proofs of computation coupled to the reported token count Page 9>. However, these mechanisms apply to the stated computation and counter, not necessarily to the provider’s realized computational cost Page 11>. p. 8 The study evaluates representative recent auditors, showing weaknesses in current approaches rather than an impossibility result Page 12>. p. 9
Conclusion
Current token-count auditing frameworks provide limited assurance because their checks rely on provider-controlled artifacts rather than independent evidence of computation Page 13>. Reliable billing accountability will require mechanisms that bind reported token counts to provider-independent evidence of the computation actually performed Page 17>. p. 9
Pruning redundant reasoning tokens
The paper identifies the problem as a structural trust gap in audits over provider-controlled evidence Page 17>. The authors show that each audit replaces direct verification of billed computation with a consistency check over artifacts the provider can choose or shape Page 17>. p. 8
How the training signal is built
The PALACE framework trains a domain-adapted auditor model on auxiliary data supplied by the provider, and this data is used to estimate hidden reasoning-token count from (P, A) Page 4>. The provider controls both the answer (A) and the auxiliary data (DA) used to fine-tune that estimator Page 4>. p. 4
Where the gains come from
The gains come from exploiting behavioral assumptions that are plausible for cooperative providers but fragile against a financially motivated adversary Page 2>. The attacks succeed because they break assumptions such as Merkle commitments binding the audit, and matching heads distinguishing honest reasoning from inflated content Page 3>. p. 2
The paper shows that semantic similarity over learned embeddings tells you whether two pieces of text are close in meaning, not whether the block was produced as part of the model’s reasoning Page 3>. This demonstrates that even exposed artifacts do not solve the problem when they remain under provider control Page 1>. p. 1
The statistical auditor's failure stems from exploiting natural slack in the aggregate test, where a provider can simulate the audit before submission by searching for inflation patterns that fit inside it Page 9>. This allows for 50. p. 085% total token-count inflation over the audit horizon without crossing detection thresholds Page 19>. p. 7
The authors conclude that the common issue across all frameworks is that each audit replaces direct verification of billed computation with a consistency check over artifacts the provider can choose or shape Page 17>. Local hardening may reduce specific attacks, but it does not remove provider control over the audited evidence Page 17>. p. 8
The final conclusion is that reliable billing accountability requires moving from consistency checks over provider-controlled artifacts to provider-independent evidence of the computation actually performed Page 17>. This necessitates mechanisms like trusted execution environments or cryptographic proofs of computation coupled to the reported token count Page 19>. p. 8 p. 8
The paper evaluates representative recent auditors, showing weaknesses in current approaches rather than an impossibility result Page 12>. The results show that the problem is not a single model or hyperparameter, but a structural trust gap in audits over provider-controlled evidence Page 17>. p. 9 p. 9
The paper demonstrates that for smaller injections, such as 1,000 or 1,500 tokens, the resulting martingale statistic Mt closely resembles the benign patterns in Figure 11 Page 23>. This indicates that moderate inflation can remain hidden when negative deviations from non-inflated samples dominate the martingale trajectory Page 21>. p. 7
The paper shows that for larger injections, such as 10,000-token inflation, the martingale eventually crosses the threshold, with earlier detection as the injected amount increases Page 21>. This suggests magnitude-dependent detection in statistical auditing Page 21>. p. 7
The authors show that for optimal offsets, the net inflation is always positive even when using small compensating underreports on non-inflated samples Page 23>. This confirms that strategic reporting can amplify available negative slack to achieve significant overcharging Page 23>. p. 7 p. 7
The paper evaluates representative recent auditors, showing weaknesses in current approaches rather than an impossibility result Page 17>. Some attacks require provider control over the serving pipeline, auxiliary data, answer formatting, or tokenization Page 17>. p. 9
The final takeaway is that reliable billing accountability requires moving from consistency checks over provider-controlled artifacts to provider-independent evidence of the computation actually performed Page 17>. This necessitates mechanisms like trusted execution environments or cryptographic proofs of computation coupled to the reported token count Page 19>.
The paper evaluates representative recent auditors, showing weaknesses in current approaches rather than an impossibility result Page 17>. The results show that the problem is not a single model or hyperparameter, but a structural trust gap in audits over provider-controlled evidence Page 17>.
The paper demonstrates that for optimal offsets, the net inflation is always positive even when using small compensating underreports on non-inflated samples Page 23>. This confirms that strategic reporting can amplify available negative slack to achieve significant overcharging Page 23>.
The paper evaluates representative recent auditors, showing weaknesses in current approaches rather than an impossibility result Page 17>.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve token count auditing by moving from consistency checks over provider-controlled artifacts to provider-independent evidence of computation actually performed. This requires verification mechanisms that bind reported token counts to evidence such as trusted execution attestation, cryptographic proofs of inference, or thirdparty re-execution rather than clever checks on signals the provider controls.
-
Develop systems that incorporate verifiable computation methods like trusted execution environments (TEEs) or cryptographic proofs of computation coupled to reported token counts. These mechanisms could certify that a provider executed a specified computation and that the reported value is consistent with a formally defined counter, such as a tokenizer output length or a fixed number of decoding steps.
-
Implement agentic systems designed to detect subtle inflation patterns by analyzing the sequential behavior of per-sample deviations and martingale statistics across numerous reports. This allows detection even when strategic reporting exploits slack through small compensating underreports on non-inflated samples, as demonstrated in the 1000-token inflation setting where
the net inflation is always positive.
-
Enhance reasoning length estimation models by replacing reliance on answer style or auxiliary data with robust predictors that are insensitive to provider manipulation. This involves using a learned discriminator that looks for legitimacy of reasoning rather than topical co-occurrence, moving beyond the limitations of embedding similarity checks.
Sources
- Are You Getting What You Pay For? Auditing Model Substitution in LLM APIs
- Stealing Part of a Production Language Model
- HuatuoGPT-o1, Towards Medical Complex Reasoning with LLMs
- The Foundations of Tokenization: Statistical and Computational Concerns
- DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning
- Scaling Laws for Neural Language Models
- H-CoT: Hijacking the Chain-of-Thought Safety Reasoning Mechanism to Jailbreak Large Reasoning Models, Including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking
- Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting
- CoIn: Counting the Invisible Reasoning Tokens in Commercial Opaque LLM APIs
- Invisible Tokens, Visible Bills: The Urgent Need to Audit Hidden Operations in Opaque LLM Services
- SVIP: Towards Verifiable Inference of Open-source Large Language Models
- Slalom: Fast, Verifiable and Private Execution of Neural Networks in Trusted Hardware
- Auditing Pay-Per-Token in Large Language Models
- Is Your LLM Overcharging You? Tokenization, Transparency, and Incentives
- Predictive Auditing of Hidden Tokens in LLM APIs via Reasoning Length Estimation
- Emergent Abilities of Large Language Models
- Auditing Black-Box LLM APIs with a Rank-Based Uniformity Test
Related papers
- From Strategy Search to Policy Learning: MCTS-Guided Adversarial Co-Evolution for LLM Safety Alignment
- Privacy Leakage from Recurrent State in LLM Inference
- Incrementally Verifiable Oracle Aided Computation
- Privilege and confidentiality in generative AI workflows
- BadMoE: Backdooring Mixture-of-Experts LLMs via Routing-Aware Triggers and Task-Underutilized Experts
- CoVault: Secure, Scalable Analytics of Personal Data