diff --git a/skills/ms-ai-security/references/cost-optimization/small-language-models-economics.md b/skills/ms-ai-security/references/cost-optimization/small-language-models-economics.md index d2c7367..4b086ef 100644 --- a/skills/ms-ai-security/references/cost-optimization/small-language-models-economics.md +++ b/skills/ms-ai-security/references/cost-optimization/small-language-models-economics.md @@ -1,6 +1,6 @@ # Small Language Models: Economics and Use Cases -**Last updated:** 2026-04 +**Last updated:** 2026-06 **Status:** GA **Category:** Cost Optimization & FinOps for AI @@ -21,7 +21,7 @@ Microsofts Phi-serie (Phi-3, Phi-4) demonstrerer denne trenden tydelig: Phi-4-mi | Egenskap | Small Language Models (SLMs) | Large Language Models (LLMs) | |----------|------------------------------|------------------------------| | **Parameterstørrelse** | < 10 milliarder | > 10 milliarder | -| **Eksempler** | Phi-4-mini (3.8B), Phi-3-small (7B), Falcon-7B | GPT-4o (175B+), Llama-3.3-70B | +| **Eksempler** | Phi-4-mini (3.8B), Phi-3-small (7B), Falcon-7B | GPT-4o, Llama-3.3-70B | | **Inferenskostnad (Azure)** | 0,10–0,50 NOK per 1M tokens | 5,00–50,00 NOK per 1M tokens | | **Hosting-alternativ** | Cloud, on-premises, edge, sidecar | Cloud (primært) | | **GPU-krav** | Optional (CPU mulig), T4, A100 | A100, større clustere | @@ -440,9 +440,9 @@ az webapp create --name webapp-slm-phi4 --resource-group rg-slm-norway --plan pl | **Phi-4-mini** | SLM | 0,10 NOK / 1M tokens | 0,30 NOK / 1M tokens | 0,40 NOK | | **GPT-4o-mini** | Small LLM | 1,58 NOK / 1M tokens | 6,30 NOK / 1M tokens | 7,88 NOK | | **GPT-4o** | LLM | 26,25 NOK / 1M tokens | 105,00 NOK / 1M tokens | 131,25 NOK | -| **GPT-4** (utfaset, illustrativt) | LLM | ~105 NOK / 1M tokens | ~315 NOK / 1M tokens | ~420 NOK | +| **GPT-4** (legacy, illustrativt) | LLM | ~105 NOK / 1M tokens | ~315 NOK / 1M tokens | ~420 NOK | -**Besparelse:** Phi-4-mini er ~**20x billigere** enn GPT-4o-mini og ~**330x billigere** enn GPT-4o (frontier LLM) på ren token-kost. (GPT-4o/4o-mini verifisert juni 2026; GPT-4 er utfaset og oppgis kun illustrativt. Kanonisk prissannhet: `deterministic-cost-calculation-model.md` §1.1.) +**Besparelse:** Phi-4-mini er ~**20x billigere** enn GPT-4o-mini og ~**330x billigere** enn GPT-4o (frontier LLM) på ren token-kost. (GPT-4o/4o-mini verifisert juni 2026; GPT-4 er legacy og oppgis kun illustrativt. Kanonisk prissannhet: `deterministic-cost-calculation-model.md` §1.1.) ---