AI bills come down to token counts multiplied by per-token rates, adjusted by caching, batch discounts and reasoning tokens. This glossary defines 25 terms behind those line items, with a number from my price data where one helps. All prices are USD as of 2026-10-11.
Core pricing terms
- Token: The unit a model reads and writes. A token is a fragment of a word, so the same text can produce different token counts on different models.
- Input token: A token you send to the model: your prompt, system prompt, chat history and any documents. Claude Haiku 5.5 lists at $0.10 per 1M input tokens.
- Output token: A token the model writes back. Every model in my data prices output above input. Claude Haiku 5.5 lists output at $0.50 per 1M, five times its input rate.
- Price per 1M tokens: The standard way providers quote rates. Divide your token count by 1,000,000, then multiply by the rate.
- List price: The published rate before any discount, caching or volume deal. It is the starting point for any estimate, not the final bill.
- Cached input: Input the provider has recently seen, billed at a lower rate. GPT-6 Luna lists cached input at $0.01 per 1M, against $0.10 for regular input.
- Batch API: An asynchronous interface where you submit work and get results later at a lower rate. Official pages list a 50% batch discount at OpenAI, Anthropic, Google, Mistral, Alibaba and Amazon Nova. At xAI it varies by model.
- Tiered long-prompt pricing: A higher rate that applies once a single prompt passes a size threshold. OpenAI charges more past 272K tokens. Grok 4.7 and Gemini 3.1 Pro charge more past 200K tokens.
- Context window: The maximum number of tokens a model can handle in one request. Claude Haiku 5.5 and Claude Sonnet 5.5 both list 1,000,000. Command A+ lists 192,000.
- Tokenizer ratio: How many tokens a provider's tokenizer needs for the same English document, with OpenAI at 1.00. Anthropic's ratio is 1.39, so the same document uses more tokens there. Two models with the same per-token price can produce different bills.
- Preview model: A model released under a preview label. Gemini 3.1 Pro Preview lists at $2.00 input and $12.00 output per 1M tokens. Check the provider's page for terms that may change.
Reasoning and output terms
- Reasoning token: A token the model spends thinking before it answers. Reasoning tokens are billed as output tokens, so they can make a cheap-looking model expensive. DeepSeek V4.1 Flash used 5,461 reasoning tokens on one chat answer in my measured run.
- Reasoning mode: How a model handles thinking. "Always" means it reasons on every call, "default" means it reasons unless told otherwise, and "none" means it does not reason. Command A+ is listed as "none".
- Visible output: The part of the response you read. Your bill counts it plus the reasoning tokens.
- Output cap: A limit on output tokens per request. My measured runs were capped at 16,000 tokens.
Quality term
- LMArena score: A text-leaderboard rating built from human comparisons of model answers. Claude Opus 5.5 scores 1507 and Llama 4 Maverick scores 1327. A higher score does not mean a lower cost per task.
Task cost terms
- Cost per task: The total charge for one job, including input, output, reasoning and any caching.
- Cost per 1,000 tasks: Cost per task multiplied by 1,000, which is easier to budget with. GPT-6 Luna costs $0.59 per 1,000 chat answers.
- Median cost: The middle value across all 326 models I track. For a chat answer it is 0.24¢. For the 50-step coding agent it is 48¢.
- Measured basis: I ran the task once, with default settings, through OpenRouter and recorded the billed tokens. Twelve models carry this label, including GPT-6 Luna and Claude Haiku 5.5.
- Estimated basis: The cost is calculated from token counts and list prices, not a recorded run. The models outside the twelve measured ones carry this label.
- Modeled scenario: A fixed workload used to compare models on a long job. It is not a live run.
- System prompt: The standing instructions sent with every request. In my agent scenario it is 8,000 tokens and is billed again on every step.
- Agent step: One model call inside a multi-step job. My scenario runs 50 steps, adding 1,500 tokens of context and 300 visible output tokens per step.
- Without caching: The cost of the same modeled run if no input is served from cache. GPT-6 Luna's 50-step agent costs 4.1¢ with caching and $0.23 without.
Putting the terms together
Compare models on cost per 1,000 tasks, not list price. Reasoning tokens can erase a low rate. GLM 5.3 lists input at $0.039 per 1M, yet a chat answer costs 0.86¢. GPT-6 Luna lists input at $0.10 and costs 0.06¢.
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  GLM 5.3 FlashZ.ai | 0.04¢$0.36 / 1,000 | 1474 | 48 / 1,782 | measured best value | |
| 2 |  GPT-6 LunaOpenAI | 0.06¢$0.59 / 1,000 | 1443 | 45 / 1,179 | measured | |
| 3 |  Claude Haiku 5.5Anthropic | 0.07¢$0.66 / 1,000 | - | 61 / 1,299 | measured | |
| 4 |  Qwen3.8 FlashAlibaba Qwen | 0.08¢$0.76 / 1,000 | - | 88 / 1,598 | measured | |
| 5 |  Mistral Small 4Mistral | 0.08¢$0.83 / 1,000 | - | 40 / 1,377 | estimated | |
| 6 |  MiMo-V2.6-ProXiaomi | 0.09¢$0.90 / 1,000 | 1479 | 206 / 933 | estimated best value |
Caching matters most on long agent jobs, where the same system prompt repeats. Here is the gap in my modeled 50-step scenario:
| Model | With caching | Without caching |
|---|---|---|
| GPT-6 Luna | 4.1¢ | $0.23 |
| Claude Sonnet 5.5 | 84¢ | $6.52 |
| Claude Opus 5.5 | $1.68 | $13.04 |
| Claude Fable 5.1 | $3.46 | $32.60 |
See the 50-step agent task page for the full ranking. The calculator lets you enter your own token counts, and the methodology page explains how measured and estimated figures differ. If you are weighing subscriptions instead of per-token rates, see plans.
FAQ
Is the batch API always half price?
No. Official pages list a 50% batch discount at OpenAI, Anthropic, Google, Mistral, Alibaba and Amazon Nova, but at xAI it varies by model. Check the provider's page before planning a batch job.
Why can a model with a lower price per token cost more per task?
Reasoning tokens are billed as output, and some models use far more of them per answer. Per-task figures capture this. My chat answer task page shows the spread.
Should I trust measured or estimated figures?
Measured figures come from a recorded run. Estimated figures use token counts and list prices, so treat them as a close guide rather than an exact bill. The methodology page explains each label.