Table14
AI token market, November 2026
Market · monthly price ledger · · 6 min read
- Verdict · price report
- Output tokens still drive the bill. The small tiers now cover most agent sub-tasks, so route by task, not by vendor.
- Models tracked
- 11 live, 1 retiredSix vendors, sample set
- Prices in
- US$ per million tokensList price, no batch or volume discounts
- Sample month
- 40M in, 4M out30M of the input served from cache
- Compared with
- October 2026Change is on the blended price
SampleVendor and model names are placeholders and every price below is sample data, written to show how the monthly table works. None of it is a current price.
How to read it
Each row is one model at its list price. Blended assumes three input tokens for every output token, which is roughly the mix chat and coding assistants produce. The last column is what a sample month of agent work would cost: 40 million input tokens, 30 million of them cache hits, and 4 million output tokens.
Sort by any column. Hover or tab into a cell to follow its row and column; arrow keys move between cells. On a phone, swipe the table sideways; the model name stays pinned.
The ledger
| tokens | $ / M | $ / M | $ / M | $ / M | blended | $ total | ||
|---|---|---|---|---|---|---|---|---|
| Vendor A · Largeclosed | Frontier | 400k | 3.00 | 0.30 | 15.00 | 6.00 | 0% | 99.00 |
| Vendor A · Mediumclosed | Mid | 400k | 1.00 | 0.10 | 5.00 | 2.00 | −20% | 33.00 |
| Vendor A · Smallclosed | Small | 200k | 0.25 | 0.03 | 1.25 | 0.50 | 0% | 8.40 |
| Vendor B · Proclosed | Frontier | 1,000k | 2.50 | 0.25 | 12.00 | 4.88 | 0% | 80.50 |
| Vendor B · Flashclosed | Small | 1,000k | 0.15 | 0.04 | 0.60 | 0.26 | −13% | 5.10 |
| Vendor C · Reasonerclosed | Frontier | 200k | 5.00 | 0.50 | 25.00 | 10.00 | new | 165.00 |
| Vendor C · Chatclosed | Mid | 200k | 0.80 | 0.08 | 3.20 | 1.40 | 0% | 23.20 |
| Vendor D · Open 70Bopen, hosted | Mid | 128k | 0.60 | — | 0.80 | 0.65 | −7% | 27.20 |
| Vendor D · Open 8Bopen, hosted | Small | 128k | 0.05 | — | 0.08 | 0.06 | 0% | 2.32 |
| Vendor E · MoE Largeopen, hosted | Frontier | 256k | 0.40 | 0.10 | 1.60 | 0.70 | +8% | 13.40 |
| Vendor F · Miniclosed | Small | 128k | 0.10 | 0.025 | 0.40 | 0.18 | 0% | 3.35 |
| Vendor C · Chat v1retired 31 Oct | Mid | 200k | 1.00 | — | 4.00 | 1.75 | retired | 56.00 |
What moved
Three cuts and one rise in the sample set. Vendor A's mid tier dropped a fifth on the blended price, Vendor B's fast model came down again, and the hosted 70B open model shaved a little off. The hosted mixture-of-experts model went up slightly.
One frontier reasoning model joined at the top of the range. In the sample month it costs about seven times the mid-tier model from the same vendor, almost all of it from output tokens.
Output is still the bill
For most rows, output costs four to five times input. An agent that writes long plans or verbose diffs pays for it here. Trimming what the model writes back is usually worth more than switching vendor.
Routing advice
- Sub-tasks such as classification, extraction and short summaries: the small tier. Check quality on twenty real examples before you commit.
- Day-to-day coding help: a mid-tier model with caching on. Long system prompts and repository context are exactly what caching is for.
- Hard planning steps only: the frontier tier, called rarely and with a cap on output length.
Running models yourself instead? Read the Ollama review before you buy a GPU.