AI token market: November 2026
Headline prices barely moved this month. The changes that matter are in cached-input and batch pricing, so the cheapest model for you depends on how much of each prompt repeats and whether you can wait for results.
Sample post. This shows the layout of the monthly market post. The providers are anonymised and every price on this page is an illustrative figure, not a real quote. A live edition would name each provider and link to its pricing page.
How to read the table
Every price is per million tokens at the provider's public list price, with no committed-use discounts. "Cached input" is the price for tokens the provider has already seen in a recent request, such as a long system prompt. "Batch input" is the price when you accept results within about a day instead of straight away.
The workload column prices out one sample month, roughly what a busy coding agent gets through, so you can compare models on a bill instead of on a single rate. Select any column heading to sort by it.
Prices per million tokens, November 2026
Sample figures · USDSwipe the table sideways to see every column. The model name stays pinned.
| vs October | |||||||||
|---|---|---|---|---|---|---|---|---|---|
| Provider A · Large | Frontier | 5.00 | 0.50 | 25.00 | 2.50 | 400K | 5.0× | 870 | No change |
| Provider A · Small | Budget | 0.25 | 0.03 | 1.25 | 0.13 | 400K | 5.0× | 44 | No change |
| Provider B · Pro | Frontier | 3.00 | 0.30 | 15.00 | 1.50 | 1M | 5.0× | 522 | Cached input −50% |
| Provider B · Flash | Mid | 0.30 | 0.08 | 2.50 | 0.15 | 1M | 8.3× | 79 | No change |
| Provider C · Max | Frontier | 10.00 | 1.00 | 40.00 | 5.00 | 200K | 4.0× | 1,540 | No change |
| Provider C · Mini | Mid | 0.60 | 0.15 | 2.40 | 0.30 | 200K | 4.0× | 105 | Batch tier added (−50%) |
| Provider D · Open 70B, hosted | Open weights | 0.60 | Not offered | 0.80 | Not offered | 128K | 1.3× | 136 | Output −10% |
| Provider D · Open 8B, hosted | Open weights | 0.10 | Not offered | 0.15 | Not offered | 128K | 1.5× | 23 | No change |
| Provider E · Reasoner | Mid | 1.10 | 0.28 | 4.40 | 0.55 | 200K | 4.0× | 193 | Cached input −30% |
A dash means the provider doesn't offer that price, and the workload column then uses the full input price. Workload = 60M uncached input + 140M cached input + 20M output tokens. "Output ÷ input" shows how much more a token costs to write than to read. That ratio drives agent bills, because agents write a lot.
What changed
- Cached input got cheaper at two providers. For prompt-heavy work such as coding agents and retrieval, that matters more than any headline cut, because most of the input tokens are repeats.
- A new batch tier. Provider C now halves the input price for jobs that can wait. Overnight summarising and evaluation runs should move to it.
- Open-weight hosting is still the floor. The hosted 8B model is the cheapest way to run simple classification. It has no caching discount, though, so the gap closes on long, repetitive prompts.
My take
Don't switch providers this month on list price alone. Pull last month's usage, split input into cached and uncached, and price it against the two or three models that pass your own evals. That takes an hour and usually saves more than any headline change.