LLM pricing shifts beyond token meters

LLM pricing is split between subscription products and pay-as-you-go APIs billed by token usage. Consumer and team plans increasingly bundle more than chat: OpenAI includes Codex, Sora, and Deep Research in paid ChatGPT tiers; Anthropic includes Claude Code and Cowork; Google adds Jules, Code Assist, NotebookLM, Flow, Veo 3.1, and Antigravity; Microsoft packages Copilot across productivity apps and Copilot Studio. xAI, Moonshot Kimi, MiniMax, Mistral AI, DeepSeek, and Meta take different paths across flat plans, credits, free access, private previews, and token-only API billing.

Cost control depends on input and output tokens, context windows, max output settings, prompt caching, and model choice. Long conversations can drive input tokens higher as history grows, while rolling windows, retrieval, and summarization help cap costs. Reasoning models such as GPT-5.5 Pro, Claude Opus 4.7 with extended thinking, and Gemini 3.1 Pro Deep Think can add internal reasoning tokens, making them more expensive for analytical workflows unless their accuracy gains justify the spend.

Pricing is also moving beyond simple token meters. Agentic workloads can trigger many sequential model calls, pushing providers toward task quotas, session limits, credits, and bundled usage pools. Enterprise buyers also need to budget for embeddings, vector databases, reranking, caching infrastructure, logging, monitoring, auditing, security controls, compliance features, and deployment options that sit outside headline API rates.