Language-model price sheets look like a single number per model and are actually four prices wearing one label: input, cached input, output, and batch. In November 2026 the spread between the cheapest and the most expensive model on the sheet is a hundredfold on input and on output, so reading the sheet correctly is worth more than any prompt trick.
The four prices on every row
Take GPT-6.2 Astra, released on 17 November 2026. OpenAI lists $8 per million input tokens, $0.40 per million cached input tokens, and $40 per million output tokens, with batch requests at half those rates. Claude Fable 5.2 lists $8, $0.20 and $40. Gemini 4 Hydrogen lists $1.50, $0.075 and $7.50. The pattern across providers is consistent:
- Output costs four to six times input. Generation is sequential and memory-bound; reading a prompt is parallel.
- Cached input costs 2.5–10% of normal input on the current frontier tiers (older models are stingier: GPT-4o still charges 50%). Reusing a prefix is close to free compared to resending it.
- Batch costs 50%. Waiting hours instead of seconds halves the rate at every major provider.
What a token of each kind is
Input tokens are everything in the request: system prompt, history, retrieved documents, tool schemas, images converted to tokens. Cached input tokens are the part of that prefix the provider recognises as unchanged since a recent request, billed at the cached rate instead of the full one. Output tokens are everything the model writes, including, on reasoning models, the thinking tokens you never see. Batch tokens are the same tokens processed through an asynchronous queue with a 24-hour (or similar) turnaround.
Cache economics, with numbers
Caching is the largest lever most teams have and the one most estimates ignore. Suppose your system prompt and retrieved policy block are 4,000 tokens, and you send 1,000 requests a day on GPT-6.2 Sol.
Without caching, that prefix costs 4,000 × 1,000 × 30 = 120 million input tokens a month at $2 per million: $240. With caching, the same 120 million tokens are billed at $0.10 per million: $12. The saving is $228 a month from one design decision — keeping the prefix byte-identical and putting the variable part last.
Two cautions. Cache writes are not free: OpenAI charges $2.50 per million for a 30-minute cache write on Sol, Anthropic charges 1.25× the input price for a 5-minute write, and Google bills cache storage per hour. And a prefix only hits if it is identical, so a timestamp or a user name at the top of the system prompt destroys the hit rate for everyone.
Batch, and when 50% is the right trade
Batch pricing is straightforward: half the rate for work that can wait. Classification of yesterday's tickets, nightly summarisation, embedding backfills, evaluation runs and dataset labelling all belong in batch. Interactive chat does not. A team that moves its evaluation suite to batch typically cuts that line item in half without touching quality — and evaluation suites are usually the second-largest bill after production traffic.
Long context is not linear
Most current models have a short-context rate and a long-context rate. OpenAI's GPT-6.2 models double the input price and raise output by 50% once a request passes 272K input tokens. Gemini 3.1 Pro doubles input and raises output by 50% above 200K. Claude Haiku 5.5 multiplies its input rate fivefold above 100K. Sending a whole codebase "because the window is big" can therefore cost two to five times what the headline rate suggests. The simulator's token fields accept any number, but the price book applies short-context rates; if a request will cross the threshold, price it at the long-context rate by using a custom row.
What the sheet leaves out
An invoice is not a multiplication of the price sheet. The recurring extras:
- Reasoning tokens billed as output, often two to five times the visible answer on hard prompts.
- Tool schemas billed as input on every call, even when no tool runs.
- Retries and agent loops resending the same context, which multiply input cost silently.
- Cache writes, storage, and minimum cache lifetimes.
- Embeddings and vector storage for retrieval, priced separately per million tokens and per GB-month.
- Regional endpoints, which carry a 10% uplift at several providers, and priority tiers at 2× or more.
A budget built from the headline rate and the prompt length will be low by 20–50% for an agentic product. The calculator's buffer advice exists for that reason.
How to read a provider's pricing page in five minutes
Price pages are dense, and the four numbers you need sit in different places. A quick order that works on all three major providers:
- Confirm the tier. Standard, batch, flex, priority or fast modes are different sheets. The headline number is Standard.
- Find the context threshold. Look for the note under the table: "short context" is usually up to 200K–272K input tokens, and above it the rate steps up.
- Read the cached column, not just input. A cheap cache rate is worth more than a cheap input rate for any workload with a stable prefix.
- Check the effective date. Introductory pricing is common right now — Gemini 4 Carbon doubles on 1 December 2026, Gemini 3.8 Flash on 1 January 2027. A budget built on today's number is wrong next quarter by design.
- Note what is missing. Cache writes, storage, regional endpoints and priority processing are often on a separate page. If the sheet does not mention them, they still appear on the invoice.
TokenLedger's price book records the first three and the effective dates; the custom row covers the rest when a negotiated or regional rate applies.
Reading the November 2026 sheet
The current frontier row is crowded and unusually cheap at the small end. GPT-6 Luna at $0.10/$0.50 and Claude Haiku 5.5 at $0.10/$0.50 cost less than a hundredth of Astra or Fable 5.2 per output token. Gemini 3.8 Flash at $0.75/$3.75 sits in the middle and carries an introductory price until the end of 2026. GPT-6.2 Sol and Claude Sonnet 5.5 both land at $2/$10, which has become the default "serious work" price point; Gemini 4 Argon and Carbon match it until 30 November 2026, then double.
The practical reading: choose the smallest model that clears your quality bar on a real evaluation set, use caching for anything stable, batch anything asynchronous, and keep one frontier model for the cases that fail. Teams that do those four things routinely run at a fifth of the cost of teams that send everything to the top tier — with the same output quality on the tasks that matter.