Right-sizing models: routing between small and frontier models

1Introduction

A workflow in a Gulf distributor or logistics company is a chain of steps. An order that arrives on WhatsApp is classified, extracted into lines, checked against the catalogue, confirmed with a reply and, weeks later, reconciled against a delivery note and an invoice. Each step can go to a different model. Most first deployments send all of them to the most capable model available, because that is the surest way to make the first workflow work.

That default grows expensive with success. Vendors bill per tokenA piece of text, often part of a word, that a model reads or writes. Vendors bill per million tokens read and per million written., and on the pricing pages we read on 14 October 2026, OpenAI listed gpt-6-astra at $10.00 per million input tokens and $50.00 per million output tokens and gpt-6-luna at $0.10 and $0.50, while Anthropic listed Claude Opus 5.5 at $4 and $20 and Claude Haiku 5.5 at $0.10 and $0.50 [5OpenAI (2026). Pricing. OpenAI API documentation., 6Anthropic (2026). Pricing. Claude API documentation.]. In 2023, Chen, Zaharia and Zou noted that the fees of language model APIs could differ by two orders of magnitude [1Chen, L., Zaharia, M. and Zou, J. (2023). FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176.]. They still do.

The Gulf sharpens the question. The work arrives in Arabic and English, often in the same message, and English-centric tokenizers need several times as many tokens for the same Arabic text [3Petrov, A., La Malfa, E., Torr, P. H. S., et al. (2023). Language Model Tokenizers Introduce Unfairness Between Languages. Advances in Neural Information Processing Systems 36 (NeurIPS 2023)., 11Ahia, O., Kumar, S., Gonen, H., et al. (2023). Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models. Proceedings of EMNLP 2023.]. Some data cannot leave the country, which rules models out before price is discussed. And a reply in Gulf Arabic, the output a customer actually reads, is where published evaluations are thinnest and hardest to automate [12Mousi, B., Durrani, N., Ahmad, F., et al. (2025). AraDiCE: Benchmarks for Dialectal and Cultural Capabilities in LLMs. Proceedings of COLING 2025.].

We ask how a company can keep frontier-level quality at lower cost by sending each step to the smallest model that does it well. This is a review of published evidence with our own arithmetic, not a report on a measured deployment. It contributes:

• A review of routing, distillation, quantization and caching (section 2).
• A cost model and a calculator at list prices read on 14 October 2026 (section 3).
• The Arabic tokenizer premiums, quoted from the studies that measured them (section 4).
• A mapping of four common workflow steps to model size, offered as our starting judgement (section 5).
• A router design, and a procedure for testing any routing change on the customer's past cases (sections 6 and 7).

2Background and related work

Three lines of work bear on the question: choosing which model handles a request, making small models capable of more, and not paying twice for the same tokens. Tokenization follows in section 4.

Cascades and routers

FrugalGPT framed the problem in 2023 [1Chen, L., Zaharia, M. and Zou, J. (2023). FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176.]. Its cascade sends a query to a cheap model first and moves to a dearer one only when a small scorer, in its case study a DistilBERT regression model, judges the answer unreliable. Matched to the accuracy of the best single model, the cascade cost 98.3% less on HEADLINES, a financial news task, 73.3% less on OVERRULING, a legal one, and 59.2% less on the reading set COQA.

A router decides before any model is called. Hybrid LLM predicts from the query alone whether a small model's answer will come close to a large one's, and reports up to 40% fewer calls to the large model with no drop in response quality [13Ding, D., Mallick, A., Wang, C., et al. (2024). Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing. Proceedings of ICLR 2024, arXiv:2404.14618.]. The gain shrank as the gap between the pair grew: between Llama-2 7B and 13B, its deterministic router sent 40% of queries to the small model for a 0.2% drop in quality; between Llama-2 13B and GPT-3.5-turbo, the same share cost up to 4% in quality.

AutoMix lets the small model check itself through few-shot self-verification, and a router built on a partially observable decision process uses that estimate to decide whether to escalate. It reports more than 50% lower computational cost for comparable performance across five models and five datasets [14Aggarwal, P., Madaan, A., Anand, A., et al. (2024). AutoMix: Automatically Mixing Language Models. Advances in Neural Information Processing Systems 37 (NeurIPS 2024), arXiv:2310.12963.].

RouteLLM learned routers from 80,000 human preference battles in Chatbot Arena, augmented with comparisons judged by GPT-4, and sends each query to one of GPT-4 or Mixtral 8x7B [2Ong, I., Almahairi, A., Wu, V., et al. (2024). RouteLLM: Learning to Route LLMs with Preference Data. Proceedings of ICLR 2025, arXiv:2406.18665.]. Its best routers reached 95% of GPT-4's score on MT Bench [15Zheng, L., Chiang, W.-L., Sheng, Y., et al. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023 Datasets and Benchmarks Track, arXiv:2306.05685.] with a cost saving ratio of 3.66 against a random router, and its most expensive router added under 0.4% to the cost of GPT-4 generation.

Smaller models

DistillationTraining a small model to reproduce a larger model's outputs, so that the small one can do the task alone. trains a small model to reproduce a large one [16Hinton, G., Vinyals, O. and Dean, J. (2015). Distilling the Knowledge in a Neural Network. NIPS 2014 Deep Learning Workshop, arXiv:1503.02531.]. Gemma 2's 2B and 9B models were trained this way, and their authors describe them as competitive with models two to three times larger [17Gemma Team (2024). Gemma 2: Improving Open Language Models at a Practical Size. arXiv:2408.00118.]. FrugalGPT applies the idea to a single task: fine-tune a small model on an expensive model's answers, which also allows shorter prompts and faster answers [1Chen, L., Zaharia, M. and Zou, J. (2023). FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176.]. Read the vendor's terms first: Anthropic's prohibit using its services to train competing AI models without its approval [18Anthropic (2025). Commercial Terms of Service. Anthropic legal, section D.4, effective 17 June 2025.]. Small general models are often English-first: the authors of Phi-3-mini, a 3.8-billion-parameter model that scored 69% on MMLU, mostly restricted its language to English [19Abdin, M., Aneja, J., Awadalla, H., et al. (2024). Phi-3 Technical Report: A Highly Capable Language Model Locally on Your Phone. arXiv:2404.14219.].

QuantizationStoring a model's weights in fewer bits, for example 4 instead of 16, so that it needs less memory and runs faster, sometimes at a cost in quality. cuts the cost of serving any model. LLM.int8() ran a 175-billion-parameter model in 8 bits without performance loss [20Dettmers, T., Lewis, M., Belkada, Y., et al. (2022). LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale. Advances in Neural Information Processing Systems 35 (NeurIPS 2022), arXiv:2208.07339.], and AWQ keeps 4-bit models accurate by protecting about 1% of salient weights [21Lin, J., Tang, J., Tang, H., et al. (2024). AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration. Proceedings of MLSys 2024, arXiv:2306.00978.]. For Arabic the caveat is direct. Quantization hurt non-Latin-script languages most, and automatic metrics understated the damage: a 1.7% automatic drop in Japanese matched a 16.0% drop reported by human evaluators [22Marchisio, K., Dash, S., Chen, H., et al. (2024). How Does Quantization Affect Multilingual LLMs?. Findings of EMNLP 2024, arXiv:2407.03211.]. Arabic was evaluated automatically but not by people, so any hidden loss there is unknown.

Caching and serving

Both vendors we priced offer prompt cachingThe vendor keeps the start of a prompt that repeats across requests and bills those tokens at a reduced rate when they are read again.. Anthropic bills cache reads at 0.1 times the input price on most models and 0.05 times on Claude Opus 5.5 and Claude Sonnet 5.5 [6Anthropic (2026). Pricing. Claude API documentation.]; OpenAI discounts cached input by up to 95% when the whole rendered prefix matches [23OpenAI (2026). Prompt caching. OpenAI API documentation.]. An agent's instructions, catalogue and customer rules repeat on every call, so they go first and the incoming message last. GPTCache reuses whole answers to similar requests and reports responses 2 to 10 times faster on a hit [24Bang, F. (2023). GPTCache: An Open-Source Semantic Cache for LLM Applications Enabling Faster Answers and Cost Savings. Proceedings of the 3rd Workshop for Natural Language Processing Open Source Software (NLP-OSS 2023).]; that suits stable lookups, not answers that depend on today's stock or balance.

For models run on one's own hardware, serving matters as much as size. vLLM's PagedAttention raised throughput 2 to 4 times at the same latency [25Kwon, W., Li, Z., Zhuang, S., et al. (2023). Efficient Memory Management for Large Language Model Serving with PagedAttention. Proceedings of SOSP 2023, arXiv:2309.06180.], and speculative decoding, in which a small model drafts tokens that the large one verifies, ran T5-XXL 2 to 3 times faster with identical outputs [26Leviathan, Y., Kalman, M. and Matias, Y. (2023). Fast Inference from Transformers via Speculative Decoding. Proceedings of ICML 2023, arXiv:2211.17192.].

3What a step costs

A step costs the tokens it reads times the input price, plus the tokens it writes times the output price. Routing changes the prices; caching and shorter prompts change the input tokens; tokenization, in section 4, changes both token counts at once.

Within each vendor, the large model costs 40 to 100 times the small one per token, and for all four models output costs five times as much as input, so drafting a reply costs more per token than classifying a message. Prices across vendors are not like for like, because each model counts tokens its own way: Anthropic notes that the tokenizer it introduced with Claude Opus 4.7 produces about 30% more tokens for the same text than the one before [6Anthropic (2026). Pricing. Claude API documentation.].

At a 100-fold gap, routing a share of steps to the small model saves almost exactly that share: routing 70% saves 69.3%. A cascade also pays the small model on the steps it escalates, but at a 30% escalation rate and the gaps above, that adds under one percentage point to the bill, so the choice between a router and a cascade turns on latency and the quality of the check, not on price. As an illustrative example, 50,000 steps a month, each reading 3,000 tokens and writing 500, cost $2,750 on gpt-6-astra alone and $844.25 with 70% of them on gpt-6-luna; with Claude Opus 5.5 and Claude Haiku 5.5, the figures are $1,100 and $349.25.

The calculator leaves out three adjustments. Anthropic bills batch requests at half price [6Anthropic (2026). Pricing. Claude API documentation.], which suits steps that can wait, such as a nightly reconciliation. Caching cuts the input term. Regional processing costs more: OpenAI adds 10% for data residency on models released from 5 March 2026, and Anthropic charges 1.1 times the standard rate for US-only inference [5OpenAI (2026). Pricing. OpenAI API documentation., 6Anthropic (2026). Pricing. Claude API documentation.]. Prices also move on their own: Google lists Gemini 4 Argon at $2.00 per million input tokens through 30 November 2026 and $4.00 from 1 December 2026 [27Google (2026). Gemini Developer API pricing. Gemini API documentation.]. A routing decision should carry the date of the prices it was made on.

4The tokenizer tax on Arabic

A token is not a fixed amount of text. A tokenizerThe component that cuts text into tokens. One trained mostly on English cuts Arabic words into more, smaller pieces. learned mostly from English cuts Arabic into more pieces, so the same message costs more, takes longer and fills more of the model's context. Petrov and colleagues measured this on FLORES-200A benchmark of the same sentences, taken from Wikipedia, translated into about 200 languages, so that languages can be compared on identical content. as a tokenization premiumThe number of tokens a language needs for a text, divided by the number English needs for the same text. A premium of 3 means three times the tokens. [3Petrov, A., La Malfa, E., Torr, P. H. S., et al. (2023). Language Model Tokenizers Introduce Unfairness Between Languages. Advances in Neural Information Processing Systems 36 (NeurIPS 2023).]. On cl100k_base, the tokenizer of ChatGPT and GPT-4, Standard Arabic had a premium of 3.04, and so did Najdi Arabic, spoken in central Saudi Arabia; on GPT-2 the premium was 4.40. The authors tie premiums directly to cost, since vendors bill per token, and to latency, since processing time in their measurements grew linearly with the number of tokens.

The multilingual tokenizers in the study's main comparison sit between 1.14 and 1.40, and one trained for Arabic reverses the premium: Arabic BERT needs fewer tokens for the Arabic text than for its English translation. It also stays near parity across 25 Arabic dialects, at 0.91 for Jeddah, 0.92 for Doha and Riyadh and 0.94 for Muscat relative to Standard Arabic [3Petrov, A., La Malfa, E., Torr, P. H. S., et al. (2023). Language Model Tokenizers Introduce Unfairness Between Languages. Advances in Neural Information Processing Systems 36 (NeurIPS 2023).]. Ahia and colleagues, studying OpenAI's API in 22 languages, found languages with their own scripts needing up to five times as many tokens as English, and their speakers overcharged while obtaining poorer results [11Ahia, O., Kumar, S., Gonen, H., et al. (2023). Do All Languages Cost the Same? Tokenization in the Era of Commercial Language Models. Proceedings of EMNLP 2023.]. Their Arabic results are shown only in figures.

Tokenizers built for Arabic

Jais trained a byte-pair tokenizer of 84,992 entries on 10 billion words, half English and half Arabic; Table 2 gives its fertilityThe average number of tokens a tokenizer produces per word. Close to 1 means that words are rarely split. comparison [4Sengupta, N., Sahu, S. K., Jia, B., et al. (2023). Jais and Jais-chat: Arabic-Centric Foundation and Instruction-Tuned Open Generative Large Language Models. arXiv:2308.16149.]. ALLaM instead merged an Arabic vocabulary into the tokenizer of Llama-2 and reports that this lowers fertility in Arabic [28Bari, M. S., Alnumay, Y., Alzahrani, N. A., et al. (2025). ALLaM: Large Language Models for Arabic and English. Proceedings of ICLR 2025, arXiv:2407.15390.].

For routing, the price of a step in Arabic is the vendor's price times a premium that depends on each model's tokenizer, and the premium can differ between the frontier model and the small one. As an illustrative calculation, a premium of 3.04 on reading and writing would turn the $2,750 month of section 3 into $8,360 for the same work. We found no published measurement on Arabic of the tokenizers behind the models in Figure 1, and tokenizers change between versions of a model [6Anthropic (2026). Pricing. Claude API documentation.]. The premium has to be measured on the customer's own messages and documents, in both languages, with each candidate model's own token count.

5Matching steps to models

Table 3 takes four steps that recur in Gulf distribution and logistics and sets out where we would start each one. It is our judgement from the evidence above, and every row is a hypothesis to test on the customer's past cases (section 7).

Three principles sit behind the table. Give a step a contract where one exists: an extraction that code can check can start on a small model, and Anthropic itself describes its smallest current model as meant for tasks such as classification, extraction and routing [29Anthropic (2026). Models overview. Claude API documentation.]. Keep arithmetic in code, which does it exactly. And spend the frontier model on free text that a customer reads, especially in dialect.

AraDiCE found Arabic-centric models such as Jais and AceGPT ahead of multilingual ones on dialect tasks, with significant challenges remaining in dialect identification, generation and translation [12Mousi, B., Durrani, N., Ahmad, F., et al. (2025). AraDiCE: Benchmarks for Dialectal and Cultural Capabilities in LLMs. Proceedings of COLING 2025.]. Models did better on Gulf Arabic than on Levantine or Egyptian, which the authors attribute to its closeness to Standard Arabic; but they had to replace free generation with a multiple-choice task, because models often explained the prompt instead of replying in the dialect. Free text in dialect is where automatic evaluation is weakest, so people judge it.

6A router for operations

Figure 3 shows our design for one step. A router decides before the call where the step's type makes the decision easy, and a cascade's check catches what the router lets through.

Rules come first, because many steps announce their difficulty. An order from a known customer, in a known pattern, with items in the catalogue, is routine; a dispute, a legal notice or an amount above the customer's threshold is not. A learned classifier takes only what the rules leave undecided, and as section 2 showed, its cost is small next to generation.

The check should be as deterministic as the step allows: the schema, the catalogue and the customer's usual quantities for an extraction; the classifier's confidence and a list of labels that always escalate for a classification; a model only for free text, as in AutoMix's self-verification or FrugalGPT's scorer [14Aggarwal, P., Madaan, A., Anand, A., et al. (2024). AutoMix: Automatically Mixing Language Models. Advances in Neural Information Processing Systems 37 (NeurIPS 2024), arXiv:2310.12963., 1Chen, L., Zaharia, M. and Zou, J. (2023). FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176.]. A deterministic check also leaves a reason in the log when it fails.

Latency decides between routing first and checking after, because a cascade that escalates calls two models in sequence [2Ong, I., Almahairi, A., Wu, V., et al. (2024). RouteLLM: Learning to Route LLMs with Preference Data. Proceedings of ICLR 2025, arXiv:2406.18665.]. For a customer waiting on WhatsApp, known-hard steps should go straight to the frontier model; a nightly reconciliation can wait, and can go in a batch.

{
  "step": "extract_order",
  "routed_to": "small",
  "reason": "known customer, known order pattern",
  "model": "gpt-6-luna",
  "tokens_in": 2140,
  "tokens_out": 310,
  "check": "passed: schema, catalogue, quantity range",
  "escalated": false,
  "cost_usd": 0.00037
}

Logging each routing decision makes routing auditable. Every escalation is a case the small model could not do, and every pass that a person later corrects is a case the check missed. Both become past cases for the next test.

7Testing a routing change on past cases

Every figure in section 2 was measured on someone else's data, in English. RouteLLM's routers saved most on MT Bench, the benchmark closest to their training data, and far less on MMLU and GSM8K [2Ong, I., Almahairi, A., Wu, V., et al. (2024). RouteLLM: Learning to Route LLMs with Preference Data. Proceedings of ICLR 2025, arXiv:2406.18665.]; FrugalGPT's savings ranged from 59.2% to 98.3% [1Chen, L., Zaharia, M. and Zou, J. (2023). FrugalGPT: How to Use Large Language Models While Reducing Cost and Improving Performance. arXiv:2305.05176.]; Hybrid LLM's loss of quality grew with the gap between its models [13Ding, D., Mallick, A., Wang, C., et al. (2024). Hybrid LLM: Cost-Efficient and Quality-Aware Query Routing. Proceedings of ICLR 2024, arXiv:2404.14618.]. The only reliable estimate of what routing will save on a step, and at what cost in quality, comes from that step's own past cases.

The MMLU result carries the practical lesson. Routers trained only on Chatbot Arena conversations did no better than random on MMLU; adding about 1,500 labelled MMLU examples, under 2% of the training data, brought every router to roughly 20% fewer GPT-4 calls than random at the same quality [2Ong, I., Almahairi, A., Wu, V., et al. (2024). RouteLLM: Learning to Route LLMs with Preference Data. Proceedings of ICLR 2025, arXiv:2406.18665.]. Our reading is that a modest number of a company's own labelled cases may matter more to a router than any amount of public data.

The procedure

1. Write the step's contract with the team that does the work today: the fields of an extraction, the labels of a classification, what a good reply says.
2. Collect past cases as they arrived, including mixed Arabic and English, Arabizi and photographed documents. The team marks the correct output for each.
3. Run both models on every case with the same prompt and the same check, recording tokens, cost and latency, with tokens counted by each model's own tokenizer.
4. Score with the step's own measure: field match, label accuracy, or people's ratings for Arabic free text. GPT-4 judges agreed with people over 80% of the time on English chat [15Zheng, L., Chiang, W.-L., Sheng, Y., et al. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. NeurIPS 2023 Datasets and Benchmarks Track, arXiv:2306.05685.], but we found no such evidence for Gulf Arabic, and automatic scores can miss losses that people see [22Marchisio, K., Dash, S., Chen, H., et al. (2024). How Does Quantization Affect Multilingual LLMs?. Findings of EMNLP 2024, arXiv:2407.03211.].
5. Set the routing rule at the threshold where the routed step meets the quality bar the team chose, and read off the share routed, the escalation rate and the saving.
6. Run it in shadow mode beside the live step before it acts, and send every difference to the team.
7. Run it again when anything changes: a model version, a tokenizer, a price or a new kind of message.

8When the data cannot leave

Where a customer, or a rule in its sector, keeps data in the country, a model that runs abroad is not eligible for the steps that read that data, whatever it costs.

On 14 October 2026, Anthropic's API offered inference in two geographies, global and the United States, and kept data at rest only in the United States [8Anthropic (2026). Data residency. Claude API documentation.]. OpenAI listed ten data residency regions; the UAE was one of three with inference in the region, limited to named models such as gpt-5.6-luna and subject to additional approval [7OpenAI (2026). Data controls in the OpenAI platform. OpenAI API documentation.]. Neither page mentioned Oman, Saudi Arabia, Qatar, Bahrain or Kuwait. Models offered through the large clouds' marketplaces follow each cloud's own regional rules, which we did not review.

That leaves open-weight models run on hardware in the country, and there are now small ones built for Arabic. The Jais family spans 590 million to 70 billion parameters under the Apache 2.0 licence [9Inception (2024). Jais Family Model Card. Hugging Face.]. ALLaM-7B-Instruct-preview comes from the National Center for Artificial Intelligence at the Saudi Data and AI Authority [30National Center for Artificial Intelligence, Saudi Data and AI Authority (2025). ALLaM-7B-Instruct-preview. Hugging Face model card.]. Fanar Prime, a 9-billion-parameter model from the Qatar Computing Research Institute at Hamad Bin Khalifa University, was continually trained from Gemma-2 9B [10Fanar Team (2025). Fanar: An Arabic-Centric Multimodal Generative AI Platform. arXiv:2501.13944.], itself a distilled model [17Gemma Team (2024). Gemma 2: Improving Open Language Models at a Practical Size. arXiv:2408.00118.].

Where a workflow must keep personal data in the country, routing and residency meet. Steps that read personal data, such as a customer's name and phone number, can go to a model run in the country; steps without it may go abroad under paper agreed with the customer. That model then carries the routine volume for reasons of cost and the sensitive volume for reasons of law, and the test in section 7 has to show it is good enough for both. Self-hosting is billed in hardware and operations rather than tokens, and serving software and quantization decide how much it can carry [25Kwon, W., Li, Z., Zhuang, S., et al. (2023). Efficient Memory Management for Large Language Model Serving with PagedAttention. Proceedings of SOSP 2023, arXiv:2309.06180., 21Lin, J., Tang, J., Tang, H., et al. (2024). AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration. Proceedings of MLSys 2024, arXiv:2306.00978.].

9Implications for deployment

The evidence points to six rules.

• Start on the strongest model, then earn the move down, step by step, on past cases.
• Decide per step. Classification and extraction are the usual first candidates for a small model; free text in dialect is the last.
• Count tokens in Arabic, on the customer's own documents, before comparing prices.
• Let residency choose before price does. Decide which models may read which data, then route among those.
• Cache the stable part of every prompt and batch what can wait [6Anthropic (2026). Pricing. Claude API documentation., 23OpenAI (2026). Prompt caching. OpenAI API documentation.].
• Date every price. A routing decision is only as current as its price list.

This is how we approach it at Qantara. Our engineers work on site inside the customer's operation, write down how each workflow really runs and record a baseline. Each agent is run on the customer's past cases before it goes live, with their team marking what good looks like, and we treat a routing change as a new version of the agent that must pass the same cases. For each workflow we state in writing which model reads what and where it runs. The record and its log are kept in the customer's country, every read and write is a line in the log, including which model did each step, and above the customer's thresholds a person approves, whichever model did the work.

10Limitations

The paper is limited in seven ways.

• Every routing result we cite was measured in English with models of 2023 and 2024; none on Arabic, mixed-language messages or operational data.
• The Arabic premiums are for tokenizers published up to 2023. The tokenizers of the models priced here are unmeasured in the literature we found, and their premiums may be lower or higher.
• Prices are one day's list prices for two vendors' first-party APIs at the standard tier. Negotiated, marketplace and regional prices differ, and some are already scheduled to change [27Google (2026). Gemini Developer API pricing. Gemini API documentation.].
• Self-hosting costs are not modelled, and the calculator does not apply to them.
• Table 3 and Figure 3 are design judgements to be tested, not findings.
• The residency review covers two vendors' documentation on one date, and it is not legal advice.
• The evidence that automatic metrics miss quantization losses comes from languages other than Arabic [22Marchisio, K., Dash, S., Chen, H., et al. (2024). How Does Quantization Affect Multilingual LLMs?. Findings of EMNLP 2024, arXiv:2407.03211.].

A router that transfers to Arabic operational data without labelled cases from it, or frontier prices that fall to the level of today's small models, would weaken the case for routing step by step.

11Conclusion

The published evidence agrees on the direction: most requests do not need the largest model, and routers and cascades keep most of the quality for much less. It disagrees on the size of the saving, which depends on the task, on the gap between the models and on how closely the router's training data matches its traffic. In the Gulf, two more terms enter the arithmetic: Arabic text costs several times as many tokens on English-centric tokenizers, and some data cannot leave the country at all.

The answer is to decide per step, on the customer's own past cases, with tokens counted in Arabic and English, among the models the data may reach, and to log every decision. Done that way, right-sizing is a routine rather than a one-off saving: every new model, price or tokenizer is another run of the same cases.