Guide · 2026-09-19 · 10 min read

AI Token Pricing in 2026: The Three-Tier Market Explained (Per 1M Tokens)

LLM API pricing has settled into three distinct tiers — ultra-low-cost nano/flash models, mid-tier workhorses, and frontier reasoning models. Here are the current per-million-token rates for every major model, why output costs 3-5x input, and how caching, off-peak and batch discounts change the math.

TL;DR

AI token pricing is no longer a single curve — it is a three-tiered market: ultra-low-cost nano/flash models, mid-tier workhorses, and frontier/reasoning flagships.
APIs bill separately per 1 million input tokens (what you send) and per 1 million output tokens (what the model writes). Output typically costs 3x to 5x input because generation is sequential compute.
The spread across tiers is roughly 500x: ~$0.03/M input at the bottom, $12.50/M input at the top.
Three discounts reshape every bill: prompt caching (80-90% off cached input), off-peak billing (up to 50% off), and batch processing (~50% off).
Most teams overpay not by picking the wrong vendor, but by running tier-1 work on tier-3 models.

!xAI

1. How Token Billing Actually Works

Every major API meters the same two quantities:

The asymmetry is physics, not policy. Input tokens are processed in a single parallel prefill pass across the whole prompt. Output tokens are produced one at a time, each requiring a full forward pass through the model with the entire KV cache resident in GPU memory. Generation occupies expensive hardware for far longer per token, so it is priced accordingly.

Practical consequence: a 50,000-token prompt returning a 200-token answer is *cheap*. A 500-token prompt triggering a 20,000-token reasoning trace can cost ten times as much. Watch output volume before you optimize prompt length.

2. Tier 1 — Ultra-Low Cost and Lightweight

Fast, cheap, built for volume. This is where document parsing, classification, routing, extraction, moderation and background batch jobs belong.

DeepSeek layers further off-peak discounts on top of already-low rates — see the V4.1-Flash architecture and pricing breakdown.

When it wins: anything where a well-specified prompt and a schema-constrained output do the job. At $0.03/M input, a million-document classification pass costs less than lunch.

When it fails: open-ended multi-step reasoning, long agent loops with tool-use recovery, and code that has to be right the first time.

3. Tier 2 — Mid-Tier Workhorses

Balanced intelligence and cost. This is the default tier for general assistants, everyday chat, RAG pipelines and standard code generation.

Note the outliers. Mistral Large 3 at $0.50 / $1.50 is priced like a tier-1 model while competing in tier 2, and Grok 4.6 carries an unusually flat input-to-output ratio (2x rather than 5x), which makes it disproportionately cheap for generation-heavy work like drafting and summarization.

When it wins: virtually all user-facing product traffic. If you are not sure which tier a workload belongs in, start here and measure quality before moving up.

4. Tier 3 — Frontier and Complex Reasoning

Maximum capability, priced accordingly. Reserve for heavy multi-step agent workflows, advanced math, deep research and critical enterprise decisions.

At $75 per million output tokens, a single verbose agent run with a high reasoning-effort setting can cost several dollars on its own. Frontier models are not expensive per call — they are expensive per loop, because agent loops multiply output tokens by iteration count.

Rule of thumb: a frontier model is justified when the cost of being wrong exceeds the cost of the tokens. For a legal review or a production migration plan, $3 of tokens is trivial. For a support-ticket classifier, it is malpractice.

5. The 500x Spread, Visualized

The same monthly workload swings from $2 to $275 purely on model selection. Routing — sending each request to the cheapest tier that can handle it — is by far the highest-leverage cost control available, ahead of prompt trimming and ahead of vendor negotiation.

6. Three Industry Trends That Change the Math

Prompt caching discounts

Providers discount input tokens heavily when the prompt prefix has been seen recently. Cached input often costs 80% to 90% less — for example $0.20/M dropping to $0.02/M.

This rewards a specific prompt architecture: put everything stable (system instructions, tool schemas, reference documents, few-shot examples) at the front, and everything variable at the end. Teams that restructure prompts this way routinely cut input spend by two-thirds without changing models. Full mechanics in Prompt Caching Explained.

Off-peak and time-of-day billing

Providers like DeepSeek dynamically lower prices during low-traffic windows — up to 50% off during off-peak hours. This is real yield management arriving in AI infrastructure, and it makes *when* you run a job a pricing decision. Nightly re-indexing, evaluation suites, synthetic data generation and backfills should all be scheduled into the discount window.

Batch processing discounts

Asynchronous batch endpoints — where non-urgent requests complete within a 24-hour window — are discounted by roughly 50% across OpenAI, Anthropic and Google. Anything without a human waiting on the response belongs in batch. See the Batch API pricing guide.

Stacking them

The discounts compose. A batched, cached, off-peak job on a tier-1 model can land an order of magnitude below the naive list price of the same job run synchronously on a mid-tier model.

7. A Practical Routing Policy

1. Classify every route by failure cost, not by how impressive the task feels.

2. Default to tier 1 for extraction, classification, routing, formatting and retrieval reranking.

3. Default to tier 2 for anything a user reads.

4. Escalate to tier 3 only on trigger — low confidence, retry after failure, or an explicitly high-stakes path.

5. Cap output tokens per route. Output is where the money goes; an unbounded `max_tokens` on a frontier model is an open tab.

6. Log tokens per request with the route name attached. You cannot route what you cannot attribute.

8. Where to Check Current Rates

Posted prices move monthly. Our live pricing table tracks input, output, cache and batch rates across every provider in this article, and the LLM leaderboard ranks models by value score so you can see which ones deliver the most capability per dollar rather than simply the lowest sticker price.