The AI token market, October 2026: what I actually paid
Five ways to buy model access, one month of real invoices, and the one plan I cancelled.
- For coding agents that run all day, a flat monthly plan beat pay-as-you-go every time I measured it.
- Free tiers are real, but they are a queue, not a service. Budget for the day one disappears.
- Put one gateway in front of everything. Switching provider became a one-line change.
Every month I look at what it costs to run the agents that do most of my typing, and every month the answer has moved. This is the fourth report. The format is the same: what I bought, what it did, and whether I'd buy it again. I pay for all of it myself, no provider has seen this post, and there are no affiliate links.
The short version is above. The long version is a table, an invoice and a config file.
Five ways to buy tokens
| Route | Best for | How it's billed | Verdict |
|---|---|---|---|
| Flat coding plan | Agents that run all day | Fixed monthly fee | BUY Predictable. My default. |
| Vendor subscription | Hard problems, interactive work | Fixed fee + monthly API credit | KEEP Best quality. Credits are a reserve. |
| Token plan | Code review, bulk drafting | Fixed fee, shared pool | KEEP Cheap second opinions. |
| Free tiers | Experiments, fallbacks | £0, rate-limited | USE Never as the only route. |
| Pay-as-you-go | Spiky, low-volume jobs | Per million tokens | AVOID One runaway loop ruins the month. |
Predictability is a feature. I would rather overpay by a little than be surprised by a lot.
What I actually paid
I log every request through the gateway, so the numbers below come from my own logs, not from the providers' marketing pages. Across the month the agents made a little over thirty thousand requests; roughly four in five went to the flat plan, one in ten to the free tiers, and the rest to the vendor subscription for the problems that deserved it.
The surprise was not the total. It was the shape: with pay-as-you-go pricing my two noisiest agents would have cost more than everything else combined. On a flat plan they cost nothing extra, which is why I stopped rationing them and started getting more out of them.
The gateway config
One LiteLLM instance sits in front of every provider. Agents ask for a role name, never a vendor, so a change of supplier is a config edit and a restart.
# agents ask for "coder"; the router decides who answers model_list: - model_name: coder litellm_params: { model: openai/glm-coding, api_base: os.environ/FLAT_PLAN_BASE } - model_name: coder litellm_params: { model: openrouter/some-model:free } router_settings: routing_strategy: simple-shuffle fallbacks: [ { coder: [reviewer] } ]
The plan I cancelled
One pay-as-you-go account went unused for five weeks while its balance quietly expired. That is the real cost of having a provider "just in case": you still have to remember it exists. It's gone, the key is revoked, and the gateway fallback covers the gap.
Prices move monthly; I re-measure on the last Tuesday. If you think I got something wrong, tell me and I'll correct the post and say so at the top.