# What is the AI cost per token? Definition and business implications > Cost per token is the elementary economic unit of an AI deployment. Providers bill input tokens (your prompt) and output tokens (the model's response) separately, with a typical ratio of 1 to 5. Mastering this cost requires distinguishing unit prices, consumed volumes, and optimisation levers. ## Definition Cost per token is expressed in dollars per million tokens (MTok). Anthropic prices as of October 2026 illustrate the orders of magnitude. Claude Haiku 5.5 (lightweight model, released on 7 October 2026): $0.10 input, $0.50 output per MTok for prompts up to 100,000 tokens, $0.50 / $2.50 beyond. Claude Sonnet 5.5 (quality/cost balance): $2 / $10. Claude Opus 5.5 (advanced model): $4 / $20. Claude Fable 5.1 (flagship model): $10 / $50. The output/input ratio is 5:1 across the whole Anthropic line-up. OpenAI applies the same ratio: GPT-6.1 Sol at $2 / $10, GPT-6 Astra at $10 / $50, with a higher rate for very long requests. DeepSeek V4.1 Flash goes down to $0.30 input and $1.20 output at peak hours, half price off-peak. Three optimisation levers exist. Batch API: minus 50% on input and output, for non-interactive uses. Prompt caching: context read back from the cache costs 10% of the normal input price, or even less on the most recent models (repeated contexts, system prompts). Adapted model choice: a Haiku 5.5 costs forty times less than an Opus 5.5, and a hundred times less than a Fable 5.1 (prompts up to 100,000 tokens), for classification, extraction, or routing use cases. ## Example A 90-employee SME uses an AI assistant via the Claude Sonnet 5.5 API for 3,000 requests per day. A typical exchange consumes 800 input tokens and 400 output tokens. Monthly calculation at September 2026 prices: 3,000 × 22 days × 800 input × ($2/MTok) = $106 for input. 3,000 × 22 × 400 × ($10/MTok) = $264 for output. Gross total: $370 per month, or about $4,400 per year. Activating prompt caching on the 600 tokens of common context: those tokens are read back at 5% of the price on Sonnet 5.5, the input bill drops by more than two thirds, an annual saving of about $900 (excluding cache write costs, marginal at this volume). Switching to Haiku 5.5, twenty times cheaper than Sonnet 5.5, for 40% of simple requests (request classification): about $1,350 additional annual saving. ## For a leader Cost per token is deceptively simple. Three implications for the executive. First, the unit price displayed by a provider is only an indicator. The actual cost depends on the effective volume (input and output tokens), the model chosen, and the use of optimisation levers. A frontal comparison between providers on unit price alone is misleading. Second, the 5:1 output/input ratio at Anthropic as at OpenAI implies that outputs weigh more in the bill than inputs. Long prompts are cheap; long generations are five times more expensive. Third, choosing the model use case by use case lowers the bill: per token, Haiku 5.5 costs forty times less than Opus 5.5 and a hundred times less than Fable 5.1 (prompts up to 100,000 tokens). The golden rule: Haiku for simple tasks, Sonnet for standard uses, Opus only for critical cases. ## FAQ **How is cost per token billed?** In dollars per million tokens (MTok), and above all by distinguishing input (your prompt) from output (the model's response), billed separately. The typical ratio is 1 to 5: at Anthropic as at OpenAI, an output token costs five times an input token. **Is comparing providers on unit price alone reliable?** No, it is misleading. The displayed price is only an indicator: the actual cost depends on the effective token volume, the model chosen and the use of optimisation levers. A head-to-head comparison on unit price alone leads you astray. **Which levers reduce the bill?** Mainly three. The Batch API, which takes 50% off input and output for non-interactive uses. Prompt caching, where context read back from the cache costs 10% of the normal input price, or even less. And choosing an adapted model: a Haiku 5.5 costs forty times less than an Opus 5.5 for classification, extraction or routing tasks. **Why do long responses cost more than long prompts?** Because of the 5-to-1 ratio between output and input. Long prompts are cheap; long generations are five times more expensive. Outputs therefore weigh more in the bill than inputs. **How do we choose the right model on cost?** The golden rule: Haiku for simple tasks, Sonnet for standard uses, Opus only for critical cases. Per token, Haiku 5.5 costs forty times less than Opus 5.5 (prompts up to 100,000 tokens): keep the most expensive models for the tasks that need them. ## See also - [Token](https://wyp.agency/en/glossary/token/) - [AI TCO](https://wyp.agency/en/glossary/ai-tco/) - [Inference](https://wyp.agency/en/glossary/inference/) - [LLM (Large Language Model)](https://wyp.agency/en/glossary/llm/) - [AI API](https://wyp.agency/en/glossary/ai-api/) - [Reasoning model](https://wyp.agency/en/glossary/reasoning-model/) ## Sources - Anthropic documentation, Pricing: Claude rates, Batch API and prompt caching, October 2026 (https://platform.claude.com/docs/en/about-claude/pricing) - OpenAI API Pricing: GPT-6 Astra, GPT-6.1 Sol, GPT-6 Luna, September 2026 (https://developers.openai.com/api/docs/pricing) - DeepSeek API Docs, Models & Pricing: DeepSeek V4.1 Flash and V4 Pro, September 2026 (https://api-docs.deepseek.com/quick_start/pricing) --- Entry from the WYP.agency AI glossary · https://wyp.agency/en/glossary/cost-per-token/ · Last updated : 2026-10-14