LLM Dilution: Why a cheap AI model generates higher costs?

A business owner sees Kimi K3 priced at $3 per million input tokens and Claude Fable 5 at $10 – chooses the cheaper one, and after a month, the bill turns out to be higher, tasks take longer, and text quality declines.

A decision based solely on the per-token rate ignores three layers of costs. They appear later: longer processing time and more output tokens. Additionally, there's the risk that the gateway serves a different model than the one paid for.

Kimi K3 is a model with 2.8 trillion parameters and a 1 million token window, released on July 16, 2026 by Beijing-based Moonshot AI. It is the largest open model on the market. Its scale explains why the $3 price should raise the question of what is actually being served. The model scored 88.3 points in Terminal-Bench 2.1 – only 0.5 points behind GPT-5.6 Sol – and leads in SWE Marathon (42.0) and Program Bench (77.8) benchmarks. MXFP4 quantization compresses weights from 5.6 TB in FP16 to 1.4 TB. The LatentMoE architecture manages 896 experts, but only 16 are active per token, yielding a 2.5-fold improvement in scaling efficiency over its predecessor, K2.

Comparing $3 and $10 loses meaning when a lower-priced model generates responses twice as long or requires additional corrective prompts. When a task passes through the OpenRouter gateway, the IRIS protocol detected a 40% routing dilution among 17 commercial providers at 85% power and an estimation error fraction below 0.04. IRIS operates without access to logprob and metadata – it fingerprints the backend by generating random character strings. Backend verification achieves an AUROC curve of 0.99 and a false alarm rate of 0.017. A company can cyclically fingerprint the endpoint and demand TEE attestation in the SLA.

PreNorm Dilution as a Structural Cause of Substitution

Deep networks waste layers due to PreNorm dilution. Each layer receives an equally weighted sum of all previous outputs. The hidden state grows uncontrollably with network depth. Important information from early layers gets lost under massive accumulation. The model must be wider and parametrically heavier to achieve any progress. Inference costs rise, and gateway operators' margins shrink. Substitution is not maliciousness from a random provider – it is a predictable consequence of cost mathematics. The model name in a contract does not guarantee model behavior, therefore SLA must pertain to behavior, not the label.

Three Layers of LLM Dilution

The first layer is dilution in the price list. A low price per token often results from aggressive quantization or a smaller number of active experts. The model sounds similar, but in tasks requiring precise formatting or strict adherence to legal instructions, it generates errors that need manual correction.

The second layer is dilution in the API gateway. Providers use classifiers that recognize benchmark prompts and direct them to the real model, while regular business traffic goes to a cheaper substitute. Additionally, they mask logprobs and quantize FP16 weights to INT8 or FP4, making detection difficult. This is a emissions-style trick, like Volkswagen's. Public rankings do not protect the client. The model wins tests, but business traffic gets a diluted version. With a substitution fraction of 30%, a system designed for the flagship model suddenly generates 30% catastrophic errors. The company pays for B30 concrete, but every third delivery is B20.

The third layer is dilution over time. Even with direct access to the provider, model parameters undergo RLHF updates and safety fine-tuning. Pleiotropy means that improving one feature degrades other tasks. A Stanford and UC Berkeley study showed that GPT-4 dropped from 97.6% to 2.4% accuracy in recognizing prime numbers within three months. During the same period, GPT-3.5 jumped from 7.4% to 86.8%. Drift is chaotic and bidirectional, so it cannot be budgeted. The percentage of directly executable code decreased from 50% to 10% because the model started adding markdown tags that broke automatic parsers. Responses to sensitive inputs dropped from 21% to 5%. A pipeline based on a strict output format stops working overnight.

Fourth Hidden Layer: Brand Identity

LLMs statistically average a company's image from everything written about it. They flatten unique positioning into a generic summary. Even a perfectly functioning model will dilute a brand if the communication is generic. In B2B high-ticket, this is a direct loss of authority. This is the mechanics of AI-slop.

Regulatory Risk with a Diluted Model

A diluted model misses legal instructions and generates generic text. Sending unsolicited B2B commercial information without consent is subject to penalties of up to 3% of global revenue or 1 million PLN (Art. 398 PKE). Exceeding the spam threshold of 0.30% (optimally below 0.10%) collapses inbox placement from 90%+ to zero. Cheap automation turns into regulatory risk and a burned sending domain.

When Kimi K3 Actually Becomes Cheaper

Kimi K3 was trained with quantization-aware training (QAT) in MXFP4 format, so the quality degradation typical for post-training quantization was eliminated. The British AI Safety Institute and the American CAISI tested K3 in the cyber range The Last Ones. For the first time, state institutes formally assessed an open-weight model as frontier-class creation. When governments treat Chinese open-weight as a frontier model, the difference between $3 and $10 is a pricing strategy, not a technological gap. Self-hosting on an 8-16 H100/B200 node cluster becomes feasible. Mooncake's infrastructure separates prefill and decode, achieving 90% cache hits for repetitive tasks. Self-hosting eliminates query limits, makes model substitution mathematically impossible, and physically guarantees data sovereignty under GDPR. Multi-model routing further reduces TCO: cheap models handle classification and scraping, while critical reasoning, compliance with Art. 398 PKE, and B2B logic go to a self-hosted K3. An IRIS audit and demanding TEE attestation in the SLA provide verification that exactly the declared weights were used.

TCO Arbitrage at AGAPE

A traditional salesperson costs over 8000 PLN per month (base salary, ZUS, tools, implementation) versus a system from 2340 PLN. This entire calculation relies on model predictability. Dilution destroys the foundation of savings because a cheaply purchased task has to be performed twice. At AGAPE, we calculate the cost of a completed task. The token rate is just one component of it.

Components of the Cost of a Completed Task

• input and output tokens multiplied by the rate
• processing time
• man-hours for corrections after model errors
• cost of re-executing the task
• cost of critical errors (regulatory penalty, burned domain)

Sum up your API bill and man-hours for corrections from the last month, then divide the result by the number of tasks completed without error. This is your first approximation of the cost of a completed task. We will show the full formula in the second part of the series.

Write in the comments if you calculate the cost of a completed task or just compare price lists.