Cette charge de travail doit-elle tourner sur un LLM frontière ou un petit modèle de langage ?

It takes nine details about your AI workload — task, volume, token profile, accuracy tolerance, latency SLA, residency, language, current spend — and returns a side-by-side monthly cost across 10 models, an accuracy delta on your task, the right hosting mode, and a top-3 shortlist with fit scores. No login, runs in 90 seconds.

LLM Pricing Comparison compares token prices across models you pick. This tool picks models for a workload you describe. Same dataset, two lenses for two different buyer moments.

SLM ≈ Small Language Model, typically 1–10B parameters with task-specific accuracy that matches frontier models on narrow tasks at a fraction of the cost. LLM = frontier general-purpose models like GPT-5, Claude Opus 4.7, Gemini 2.5 Pro that are stronger on agentic and reasoning workloads.

Classification, extraction, summarization, translation. Cost-sensitive workloads at high volume. Residency-constrained deployments. Latency-critical paths where every millisecond counts. Anywhere accuracy on the specific task is good enough at much lower cost.

Monthly volume × average input tokens × published input price + monthly volume × average output tokens × published output price. Caching discount of up to 90% applied per cache-hit-rate; batch discount up to 50% applied when "Batch-tolerant" is selected. Self-hosted cost adds amortized setup + GPU monthly.

Up to 90% off the input portion when cache-hit-rate is 100% (rare). 50% off the total when batch mode is selected. Real workloads typically see 20–40% savings from caching, 50% from batch on async workloads.

They are public-benchmark proxies, not your workload. Strongly recommend a 100–500 sample PoC before committing. Benchmarks come from Artificial Analysis, HuggingFace Open LLM Leaderboard, Stanford HELM, HumanEval / MBPP, AgentBench, plus task-specific suites.

Use the matrix: under 100K queries/month → API. 100K–1M with EU residency → managed inference in EU. >1M with sub-second latency → self-hosted GPU. On-prem or air-gapped requirements → open-weight SLM on your hardware.

Typically past 1M–10M queries/month depending on token profile. The break-even chart on the results page shows the exact crossover for your inputs.

Use the min_vram_gb column on each model card. Phi-3.5 Mini fits on an L4 (24GB). Llama 3.x 8B + Mistral 7B comfortably on a single A100 40GB. Llama 3.3 70B needs 2× A100 80GB minimum at production throughput.

Frontier APIs offer some regional hosting (Anthropic EU, OpenAI EU via Azure, Gemini in EU/SG/IN). For strict on-prem only open-weight SLMs apply: Llama, Mistral, Phi, Qwen, Falcon, BharatGen.

Qwen for Chinese / Japanese / Korean. Mistral for European languages. Llama 3.x for broad multilingual baseline. GPT-5 / Claude Opus / Gemini 2.5 Pro for global coverage when budget allows.

Mistral (EU sovereign), Falcon (UAE / TII), Qwen (APAC), BharatGen (India). The tool surfaces these neutrally on cost + compliance + language merit when residency is selected — not by default.

Pricing — monthly vendor refresh + human review, with a daily snapshot cron catching mid-month moves. Benchmarks — quarterly. Sovereign-model coverage — quarterly + as new models ship.

No. No vendor sponsorships, no pay-to-play placement, every benchmark cited with source URL and capture date. We list all models we track and rank them on cost, accuracy, latency, residency — not relationships.