AI API Pricing Compared: Cost per Million Tokens (October 2026)

How AI API pricing works: tokens in, tokens out

A token is the unit AI models read and write. DeepSeek's pricing page describes it as 'the smallest unit of text that the model recognizes', which can be a word, a number or a punctuation mark. Every API call is billed on two counts:

• Input tokens: everything you send: your instructions, the customer's message, any documents or FAQ text you add for context, and the conversation so far.
• Output tokens: everything the model writes back. For reasoning models this includes 'thinking' tokens you may never see; Google's page says so explicitly for Gemini.

Because the instructions and context are resent on every call, input usually makes up most of the tokens, while output makes up a large share of the bill. That is why caching repeated input, covered below, is the biggest lever on cost.

API price per million tokens, October 2026

Standard list prices in US dollars per million tokens, from OpenAI's API pricing, Anthropic's Claude pricing, Google's Gemini API pricing and DeepSeek's pricing page, checked on 18 October 2026. Approximate, before tax; vendors change prices often.

Two details to read on the vendor pages before you commit. Google's Gemini 4 Argon and Gemini 3.8 Flash have introductory prices that double on 1 December 2026 and 1 January 2027 respectively. DeepSeek's off-peak rate applies outside its weekday peak hours (01:00 to 04:00 and 06:00 to 10:00 UTC), and its servers and terms are outside Singapore; our guide to DeepSeek, Qwen and open models covers the data questions.

What 1,000 customer replies cost

Per-million prices only make sense against a real task. Take a website assistant that answers a customer question. Each reply sends about 1,500 input tokens (instructions, the relevant FAQ text and the question) and gets back about 300 output tokens. Here is what 1,000 such replies cost at standard rates, with no caching:

For most small businesses, the model cost of a website chatbot is a few dollars a month. The bigger costs are building it properly, keeping its knowledge up to date and checking answers. Our guide to AI chatbots for websites covers those.

Prompt caching and batch: the two discounts that matter

Prompt caching means the vendor stores the repeated start of your prompt, such as your instructions and FAQ, and charges much less to reuse it. On the October 2026 list prices, a cached input token costs 5% of the normal rate for GPT-6.1 Sol and Claude Sonnet 5.5, and 5% for Gemini 4 Argon. Writing to the cache costs a little more than normal input at OpenAI and Anthropic (US$2.50 instead of US$2.00 for the mid-range models), and Anthropic's listed cache prices assume a 5-minute lifetime.

Batch processing is for work that does not need an instant answer: tagging 5,000 products overnight, summarising last month's support tickets. OpenAI's batch prices are half its standard prices, Anthropic says 'save 50% with batch processing', and Google lists its Batch API at a 50% cost reduction.

The order of savings is simple: first pick the smallest model that passes your own test, then put fixed instructions at the start of the prompt so they cache, then move anything that can wait to batch.

There are also premium tiers that cost more: OpenAI's Fast and Ultrafast modes, Anthropic's fast mode for Opus 5.5 at twice the standard price, and 10% uplifts for regional or US-only processing at OpenAI and Anthropic.

Quality per dollar: using Artificial Analysis

Price per token says nothing about how many tokens a model uses to finish a task, or how well. Artificial Analysis publishes an Intelligence Index alongside the cost to run its test set. A selection from its leaderboard on 18 October 2026:

The same model can cost very different amounts depending on how hard it is told to think. Note too that Haiku 5.5 and Luna share a token price but differ in cost per task. Our explainer on AI benchmarks and leaderboards covers what these scores do and do not measure.

Run your own 30-minute cost test

Leaderboards are a starting point. Before you build, test on your own work:

1. 1

   Collect 20 real questions or tasks

   Use real customer emails or the documents you would process, with personal data removed.

2. 2

   Write one prompt and keep it fixed

   Same instructions and context for every model, so the comparison is fair.

3. 3

   Run it on a small and a mid-range model

   Use each vendor's playground or console, on a paid tier so your data settings match production.

4. 4

   Record the token counts

   Each console shows input and output tokens per call. Multiply by the prices above for cost per task.

5. 5

   Score the answers blind

   Have a colleague rate each answer without seeing the model name. Pick the cheapest model that passes.

Repeat when prices change, which in 2026 has been every few weeks. Our guide to what AI agents cost shows how token costs add up when a model takes many steps per task.

Do it yourself or get help?

If you have a developer, the price table and the 30-minute test above are enough to choose a model. Watch the data settings too: Google says content on the Gemini API free tier is used to improve its products, while paid-tier content is not. Check each vendor's data terms before sending customer data.

If you want it built, our AI Website Assistant (S$790) answers customer questions from your own content with caching and a sensible model choice. Workflow Automation (S$390 per workflow) adds AI steps to email, forms or sheets, and a Custom Tool or Dashboard (from S$1,990) suits larger document-processing jobs.