AI API Pricing for Small Business: How to Compare Costs

The short answer: a small business pays for an AI model by the token, so the real price is the input price plus the output price, times how many requests you send. For a typical customer-email reply (about 1,500 tokens in, 500 out), 1,000 replies cost $0.40 on Claude Haiku 5.5 or GPT-6 Luna, $1.70 on Gemini 3.5 Flash-Lite, $8.00 on Claude Sonnet 5.5 and $32.00 on Claude Fable 5.2, at each vendor’s own list price. Below: 15 models from 6 vendors, each row with the date we checked it and a link to the vendor’s page.

Prices are each vendor’s published list price, read on the vendor’s own pricing page. Last full check: November 10, 2026 (OpenAI: November 6, 2026).

How is AI API pricing calculated?

Every major vendor quotes 2 prices per million tokens (MTok): 1 for the text you send in (input) and 1 for the text the model writes back (output). Output costs more: 3 to 5 times the input price on most models below, and about 8 times on Gemini 3.5 Flash-Lite. A token is roughly 3/4 of an English word, so 1,000 words is about 1,300 tokens.

The formula for 1 request:

cost = (input tokens ÷ 1,000,000 × input price) + (output tokens ÷ 1,000,000 × output price)

Multiply by 1,000 and you have the cost per 1,000 requests, which is the number worth comparing.

What do 1,000 requests cost on each model?

Worked example: 1,000 customer-email replies, each with 1,500 input tokens (the customer’s email plus your instructions) and 500 output tokens (the reply). That’s 1.5 MTok in and 0.5 MTok out.

Two models carry an end date: Gemini 4 Argon’s price doubles on December 1, 2026, and Gemini 3.8 Flash’s on January 1, 2027. DeepSeek bills half price off-peak. A cheaper model that needs 2 tries costs more than a dearer one that gets it right the first time, so test before you commit.

What does each AI API cost per million tokens?

The full price list, grouped by vendor. Use it to price your own workload with the formula above.

• Promo prices end. Google lists Gemini 4 Argon at the lower rate through November 30, 2026, and Gemini 3.8 Flash through December 31, 2026. Both double after that. Budget at the later price.
• DeepSeek has peak hours. Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday through Friday, except Chinese public holidays. Every other hour, weekends included, bills at half the peak rate.
• OpenAI’s prices are for prompts under 272K tokens. OpenAI’s page says its listed prices are standard rates for context lengths under 272K. Its pricing page doesn’t list each model’s context window.
• Claude Haiku 5.5 has two rates. A prompt over 100,000 tokens costs 5 times as much per token as a shorter one.
• Claude Fable 5.2 replaced Fable 5.1 as Anthropic’s top listed model, at $8 in and $40 out per MTok (checked November 10, 2026).

Meta Llama: Meta’s developer site shows no per-token price for Llama (checked November 6, 2026). Each cloud or host that runs Llama sets its own price, so check the host you plan to use.

Why can the real bill come out higher than the math?

• Thinking tokens. Reasoning models bill their hidden thinking as output. Google’s price page says so for Gemini in plain words. A short answer can carry many more output tokens than you see.
• Long prompts. Some models charge more once a prompt passes a size limit. Claude Haiku 5.5’s price rises for prompts over 100,000 tokens.
• Promo prices with an end date. Gemini 4 Argon’s rate doubles on December 1, 2026, and Gemini 3.8 Flash’s on January 1, 2027. Budget at the later price.
• Tokens per task, not per request. An agent that loops through 10 steps sends 10 requests for 1 job. Price the whole job, and see our LLM cost tracking guide to find the workflow behind a rising bill.

How do you cut the cost?

• Batch work that can wait. Anthropic’s Batch API takes 50% off input and output for jobs that don’t need an instant answer, like overnight review summaries, and Mistral’s batch processing also cuts the price by 50%.
• Cache the instructions you repeat. Reused prompt text bills at a fraction of the input price on both Anthropic and Google (cache hits on Claude Sonnet 5.5 are $0.10 per MTok against $2 for fresh input).
• Route by job. Send sorting, tagging and short replies to the cheapest model that gets them right, and keep the expensive model for the few jobs that need it.

How do you compare models on quality, not just price?

Run your own 20 real requests through 2 or 3 models and count the answers you’d send without editing. Independent benchmarks help as a first filter, and some publish what a model costs to run on a whole test suite, which captures the thinking-token effect.

Where do you check each vendor’s current price?

• Anthropic (Claude): platform.claude.com pricing
• OpenAI (GPT): openai.com API pricing
• Google (Gemini): aistudio.google.com pricing
• xAI (Grok): docs.x.ai models and pricing
• DeepSeek: api-docs.deepseek.com pricing
• Mistral: mistral.ai pricing

How we keep this page current

We read every price on the vendor’s own page, and each row shows the date we checked it. We recheck the whole table on the first Monday of each month, and sooner when a vendor announces a change.

Prices change often, and the price page is the only number that counts. For what each price move means for a small business, see our daily signals.