#ai-cost-optimization

Sonnet 5's $2/$10 intro price looks cheap until effort climbs. A decision matrix for Sonnet 5 vs Opus 4.8 vs Fable 5, with cost-per-task data by effort level.

Uber burned a year's AI budget in four months; Microsoft cut Claude Code. A 4-gate playbook that right-sizes model spend and cuts AI bills 60 to 80 percent.

Route each request to the cheapest model that can handle it. Model routing cuts real LLM bills 40-85% with no visible quality loss. The 2026 engineering guide.

The LLM gateway is now critical AI infrastructure. Compare LiteLLM, Portkey, Cloudflare, Vercel, and OpenRouter on caching, routing, and build-vs-buy economics.

Tokens, prefill, decode, KV cache, prefix cache, batch tier, reserved capacity, context-window discount — every LLM billing term, defined.

GPU spend, ops headcount, latency, and break-even volume for hosting Llama, Qwen, DeepSeek, and Mistral yourself vs API. With per-token cost curves at 4 scales.

Paged attention, prefix caching, MQA/GQA, MLA, and quant-aware caching — when each technique pays off and the inference-cost numbers behind it.

Cross-model quality regression, throughput lift, and VRAM savings at GPTQ-4, AWQ-4, INT8, and FP8 — benchmark data across 6 open-weight models.

When 1M context pays off — and when it bankrupts you. Token-spend math, prompt-cache strategy, and break-even tables for agentic Claude Opus 4.7 workloads.

Business guide to small language models for on-device deployment. Gemma 4 E2B/E4B, Microsoft Phi-4, and Qwen 3.5 compared for cost savings and privacy.