LLM pricing landscape 2026

As of: 2026-07 (numbers go stale fast — structure survives)

The four tiers

1. Frontier premium / thinking (~$10-25 per M input tokens, more for thinking): the hardest debugging, architecture, final review. Use sparingly.
2. Frontier standard (~$1-5/M): daily driver for complex code and writing; this tier's quality in 2026 beats 2024's premium tier at ~10x lower price.
3. Fast tier (~$0.10-0.60/M): classification, extraction, routing, mechanical edits — 80% of agent call volume belongs here.
4. Open weights local: hardware-priced; break-even vs API around heavy/batch or privacy-bound workloads.

The levers that matter more than vendor choice

• Prompt caching — 50-90% discount on repeated context prefixes; agentic loops re-send system+tools every turn, so cache-friendly layout is the single biggest cost lever.
• Routing — classify first, escalate only failures. Simple router model → cheap tier → escalate on verification failure.
• Batch APIs — 50% off for non-latency-sensitive workloads (evals, backfills, enrichment).

Engineering rule: instrument cost per feature per week before optimizing anything; most teams optimize the wrong 5%.