What is the cheapest LLM API in 2026?

Among widely used models, the cheapest options are small open-weight models served via providers — Llama 3.1 8B on DeepInfra ($0.02 / $0.04 per million tokens), DeepSeek Flash ($0.15 / $0.60 off-peak), and vendor small tiers like GPT-5.6 Luna ($0.20 / $1.20) and Gemini 2.5 Flash-Lite ($0.10 / $0.40).

How do I estimate how many tokens my text uses?

A rough rule of thumb for English text is 1 token ≈ 4 characters, or about 0.75 words. So 1,000 words is roughly 1,330 tokens. Exact counts depend on each model's tokenizer.

Do prompt caching and batch discounts stack?

It depends on the provider. It varies more than most comparisons admit: OpenAI, Anthropic and Google bill cache hits at about 10% of the input price, xAI at 16-25%, DeepSeek at about 3%. Batch is 50% off at OpenAI, Anthropic and Google but 20% at xAI. Some providers allow combining them; always check the provider's official pricing page for current rules.

LLM API 料金計算ツール 2026

OpenAI・Anthropic・Google・xAI・DeepSeek・Mistral・Metaの30モデルのトークン単価を比較し、プロンプトキャッシュ・バッチ割引込みの実際の月額を試算しましょう。

1 · ワークロードを入力

リクエストあたり入力トークン

リクエストあたり出力トークン

月間リクエスト数

プロンプトキャッシュ命中率 0%

バッチAPI (非同期・割引率はベンダー別)

2 · モデル別の月額コスト比較

ベンダー	モデル	コンテキスト	入力 $/1M	出力 $/1M	リクエスト単価	月額 ▲

𝕏 Share

表示価格はすべて Standard 料金・短いコンテキスト基準です — Batch・Flex・Priority ではなく、長いコンテキスト区間でもありません。各社はモデルごとに複数の単価を公開しているため、別の軸の値を見ると実際の請求額を2倍以上誤って見積もることがあります。キャッシュ命中分の入力は各ベンダーのキャッシュ単価(入力の10〜25%、ベンダーにより異なる)で計算されます。キャッシュ単価が非公開のモデルはスライダーの影響を受けません。バッチ割引はベンダーごとに入出力の両方へ適用されます — OpenAI・Anthropic・Google は−50%ですが、xAI は−20%です。2026年8月30日に各社の公式料金ページで確認済みです。DeepSeek のピーク時間は月〜金 UTC 01:00〜04:00 と 06:00〜10:00 で、オフピークは半額として別行に表示しています。Llama はオープンウェイトのため一次公式単価が存在せず、掲載値は DeepInfra のもので、他の推論プロバイダでは単価が異なります。

トークンカウンター — テキストを貼り付けてトークン数を推定

0≈ トークン (文字 ÷ 4)

0単語

0文字

英語の目安は1トークン ≈ 4文字 ≈ 0.75語。日本語・韓国語・コードは文字あたりのトークン数が多くなります。正確な値は各モデルのトークナイザーに依存します。

How LLM API pricing works in 2026

Every major AI provider bills by the token — roughly a word fragment. You pay separately for input tokens (your prompt, system message, retrieved context) and output tokens (the model's reply). Output tokens cost 2–6× more than input tokens, so long-form generation is what really drives your bill.

Two discounts matter more than the sticker price. Prompt caching bills repeated prompt prefixes (system prompts, few-shot examples, long documents) at a reduced rate — 10% of the input rate at OpenAI, Anthropic and Google, but 16-25% at xAI and about 3% at DeepSeek. A chatbot with a 3,000-token system prompt saves dramatically at scale either way. The Batch API gives 50% off both directions at OpenAI, Anthropic and Google (20% at xAI) when you can wait minutes-to-hours for results, which suits evals, ETL and bulk summarization.

The spread between models is enormous: as of August 2026, a million output tokens runs from about $0.04 (Llama 3.1 8B on DeepInfra) to $180 (GPT‑5.5 Pro) — a 4,500× range. Picking the smallest model that clears your quality bar is the single biggest cost lever, ahead of caching, batching, or prompt trimming.

Frequently asked questions

How much does the GPT‑5.6 API cost?

As of August 2026, per million tokens: GPT‑5.6 Sol $4 in / $20 out, GPT‑5.6 Terra $2 / $12, GPT‑5.6 Luna $0.20 / $1.20. Cached input bills at ~10% of the input rate, and the Batch API takes a further 50% off.

How much does the Claude API cost?

Anthropic prices Claude Fable 5 at $10 / $50 per million tokens, Claude Opus 5 at $5 / $25, Claude Sonnet 5 at $2 / $10, and Claude Haiku 4.5 at $1 / $5.

What's the cheapest LLM API right now?

For rock-bottom cost: Llama 3.1 8B on DeepInfra ($0.02 / $0.04), DeepSeek Flash ($0.15 / $0.60 off-peak), Gemini 2.5 Flash‑Lite ($0.10 / $0.40), or GPT‑5.6 Luna ($0.20 / $1.20). Use the calculator above to see which is cheapest for your input/output mix.

How many tokens is 1,000 words?

Roughly 1,330 tokens for English prose (1 token ≈ 0.75 words). CJK languages and source code usually tokenize less efficiently, so budget extra.

Do caching and batch discounts stack?

Provider-dependent. Some allow batch requests to also hit the prompt cache; others don't. This calculator applies both when toggled, as an optimistic estimate — check your provider's docs before committing.

Where do these prices come from?

Compiled from providers' public pricing pages and reputable trackers, last verified on the date shown in the header. Prices change often — always confirm on the official pricing page before signing a contract.

Go deeper

Claude API pricing
Every Anthropic model's current rates. OpenAI API pricing
GPT-5.6 Sol, Terra, Luna and the full lineup. Gemini API pricing
Google's tiers, incl. the long-context surcharge. Claude vs GPT
Tier-by-tier cost comparison with caching factored in. Cheapest LLM APIs
The 2026 budget leaderboard — and when cheap is enough. Prompt caching guide
The discount most teams leave on the table — 75% to 97%, by vendor. LLM API 価格改定の記録
日付つきの料金改定 — 私たちの誤りも含めて。 DeepSeek API pricing
Off-peak vs peak rates, and the cache discount. Gemini vs GPT
Which is cheaper for your token mix.