LLM API Cost Guides

Working notes on estimating and budgeting LLM API spend, built on the same verified pricing data as the table.

GuideWhat it covers
How to estimate LLM API costs The cost formula plus the four corrections that make budget estimates real: tokenizer differences, invisible tokens, caching, deprecations.
Sonnet 5 price increase (Sept 1, 2026) Claude Sonnet 5 introductory pricing ends Aug 31, 2026 — what +50% does to a real workload, and your options.

The numbers these guides are built on

As of 2026-09-07, the pricing table tracks 39 models (30 active) across 3 providers, each row verified against the provider's own pricing page. The spread is the whole story: Gemini 2.5 Flash-Lite at $0.1/$0.4 per MTok and GPT-5.5 Pro at $30/$180 differ by roughly 420× on combined rate — which is why "which model" dominates every budgeting question these guides answer. The deprecation tracker adds the other axis: 129 announced retirement dates that can force a migration mid-budget-cycle.

Recent pricing events

DateProviderWhat changed
2026-12-11OpenAIgpt-5 (all snapshots), gpt-5-pro, o3, o3-pro shut down — Migrate to GPT-5.5 family.
2026-12-01OpenAIgpt-image-1-mini, gpt-image-1.5, chatgpt-image-latest shut down — Migrate to gpt-image-2.

Events feed from the same verified dataset as the table. Run your own numbers in the calculator.