Market report, 2 November 2026, about 6 minutes
AI token market, November 2026
Flagship prices held still this month. The money moved in the middle: small and fast models got cheaper again, and hosted open-weights models kept closing the gap.
Sample figures. The providers below are anonymised and the prices are illustrative, so the table shows how the monthly post reads. Do not budget from them.
How to read the table
Prices are US dollars per million tokens, as listed on each provider's public pricing page on the first working day of the month. "Blended" assumes three input tokens for every output token, which is close to what my own agent and chat workloads look like. If your ratio is different, the input and output columns are there for you to redo the sum.
Cached input is the price for repeated context such as a long system prompt or a codebase. For agent workloads it often matters more than the headline input price.
Price per million tokens, November 2026
Sample data; sorted by provider
| Model | Context | Input | Output | Cached input | Batch discount | Blended, 3:1 | Blended vs October |
|---|---|---|---|---|---|---|---|
| Lab A, flagshipClosed weights, API | 400k | $5.00 | $25.00 | $0.50 | 50% | No change | |
| Lab A, midClosed weights, API | 200k | $3.00 | $15.00 | $0.30 | 50% | No change | |
| Lab A, smallClosed weights, API | 200k | $1.00 | $5.00 | $0.10 | 50% | −20% | |
| Lab B, flagshipClosed weights, API | 1M | $2.50 | $15.00 | $0.25 | 50% | No change | |
| Lab B, fastClosed weights, API | 1M | $0.30 | $2.50 | $0.03 | 50% | −17% | |
| Lab C, flagshipClosed weights, API | 256k | $1.25 | $10.00 | $0.13 | None | +25% | |
| Open-weights 70BVia a hosted inference provider | 128k | $0.60 | $0.80 | — | None | −14% | |
| Open-weights large MoEVia a hosted inference provider | 128k | $0.90 | $2.40 | — | None | −10% | |
| Open-weights 8B, self-hostedHomelab GPU, electricity only | 32k | $0.04 | $0.04 | — | n/a | No change |
The table scrolls sideways.
What moved
Nothing at the top. Both large closed-weights labs kept their flagship prices where they were in October, which is now the third month running. The small and fast tiers are where the cuts landed, and those are the models most of my agent steps actually call.
Lab C went the other way and raised its flagship price. Its cached-input rate is still competitive, so for long-context work the increase matters less than the headline suggests.
What it means if you are buying
- Route by task. A flagship model for planning and review, a small model for the many small calls, and the bill drops more than any single price cut will give you.
- Turn on prompt caching before you shop around. On my own workloads it has saved more than switching provider would.
- Self-hosting an 8B model costs almost nothing per token, but you pay in setup time and in quality. I use it for classification and summaries, not for code.
Method and caveats
List prices only. No negotiated rates, no free tiers, no regional pricing. Context is the largest window offered on the standard tier. The self-hosted row is my electricity cost for one GPU at home, divided by throughput, and is a rough guide rather than a measurement anyone else should rely on.