The FinOps Directory AI spend tools →

Pricing

LLM API Pricing Comparison (2026)

LLM prices fell roughly 80% from 2025 to 2026, and they move quarterly. The sticker price per million tokens is the least interesting number on this page — the levers below (caching, batch, routing) swing your real bill far more than which frontier model you pick.

Read this first

Prices below are approximate, as reported in mid-2026, and change frequently. Always confirm against each provider's official pricing page before signing a volume contract.

Price table — cost per 1M tokens (input / output)

Representative frontier and budget tiers as reported mid-2026. Output is always dearer than input because generation costs more compute.

TierModel (example)Input / 1MOutput / 1M
FrontierOpenAI GPT-5 class~$2.50~$15
FrontierAnthropic Claude Opus class~$5~$25
FrontierGoogle Gemini Pro class~$2~$12
Mid / valueGemini Flash class~$0.50~$3
BudgetGrok / Gemini Flash-Lite class~$0.10–0.20~$0.40–0.50
BudgetDeepSeek V3 class~$0.14–0.27~$0.28–1.10

The spread is enormous: output pricing alone varies by more than 600x from the cheapest open models to top-end reasoning models. That spread is exactly why routing matters more than model choice.

The hidden costs the sticker price hides

Four levers that actually cut the bill

1. Tiered routing — the biggest single win

Route by query complexity: ~70% to a budget model, ~20% to mid-tier, ~10% to premium. This alone can cut average per-query cost 60–80% versus sending everything to a premium model. A gateway like LiteLLM or Portkey is what implements this in the request path.

2. Prompt caching

Cached input reads are commonly billed at ~10% of the base input rate, and some providers advertise up to 90% savings on cached reads. Huge for repeated system prompts and RAG context.

3. Batch processing

For work that isn't latency-sensitive, batch APIs typically cut rates by ~50%.

4. Prompt & context hygiene

Trim verbose system prompts, cap output length, and prune context windows. Silent bloat here is one of the most common sources of unbudgeted spend.

How to track LLM spend

You can't optimize what you can't see. A proxy such as Helicone or a gateway such as LiteLLM logs per-request cost, tokens, and latency with a base-URL change; for finance-grade allocation across teams and products, feed AI spend into a FinOps platform alongside the rest of your cloud bill. Full breakdown in the AI cost tooling guide.

FAQ

Why are output tokens more expensive than input tokens?

Generating output runs a forward pass per token, which is more compute-intensive than reading input. Output is typically 3–6x the input price, and for frontier reasoning models the gap can be far larger.

What is the single biggest way to cut LLM API cost?

Tiered model routing. Sending ~70% of queries to a budget model, ~20% to mid-tier, and ~10% to premium can cut average per-query cost 60–80% versus routing everything through a premium model.

How much does prompt caching save?

A lot on repeated context. Cached input reads are commonly billed at ~10% of the base input rate, and some providers offer up to 90% savings on cached reads. Batch processing typically cuts rates by another 50% for non-urgent work.

Run a cost tool finance leaders should know?

List it in the independent FinOps Directory — free, self-serve, and a verified backlink from a CFO-facing site.