How to Track AI Token Costs in Real Time

Your expensive AI session happened because you couldn’t see it happening.

Prices fell this year. In October 2026 Claude Opus 5.5 costs $4 per million input tokens and $20 per million output, and OpenAI’s top model, gpt-6-astra, costs $10 and $50. A long coding session with a big context window still adds up without you noticing. The problem isn’t the price list. It’s that nobody shows you the meter.

What real-time tracking means

Logging how many tokens you used yesterday is accounting. Tracking means seeing, while the conversation is happening:

• What the model is reading. How much of each call is system prompt, files, tool output and conversation, and how close you are to the window’s limit.
• What each answer cost. Which model replied, at what effort, for how much.
• Where the money went across providers when a session switches between Claude, GPT and the rest.
• Warnings before the limit, not after it.

Step 1: see what fills the window

Every call resends everything in the window, so the window is where the money is. TinkerClaw’s Context Window panel draws it as one coloured bar, split by kind, against the model’s limit. In this shot the tab carried 45,000 tokens: 17,000 of files, 9,500 of tool definitions, under 6,000 of actual conversation.

That split is the first thing to look at. If tools and files outweigh your conversation, the cheapest fix is to load fewer of them, not to pick a cheaper model.

Step 2: see what each call cost

The EEG draws each model call as a stroke whose height is its cost in euros, coloured by provider, in the column of the effort it ran at. A yellow line marks each prompt you sent, so the expensive question stands out.

Two warnings from using it daily. It counts output tokens only for now, so it underestimates. And a thin but tall stroke, a cheap model that ran for a long time, can cost more than one short call to an expensive model.

Step 3: know which model was worth it

Once you can see the cost, the next question is whether a cheaper model would have done. The Smartness × Cost map plots every model’s intelligence score against its price, across effort levels. I publish it, updated as new models appear, as the AI Analysis page.

Four ways to spend less

• Shrink what you resend. In March we cut the agent’s standing instructions from 23.5 KB to 12 KB without losing capability. Every kilobyte saved is saved again on every call.
• Route by task. Routine checks on Claude Haiku 5.5 ($0.10 per million input) instead of Opus 5.5 ($4): a 40× difference for the same answer. TinkerClaw’s router does this per turn.
• Set budget pressure. Above 85% of the week’s allowance, cap at a mid-range model. Above 95%, drop to the cheapest. Automatic, no surprises.
• Put automated work on a subscription. Scheduled jobs and background loops run on flat-rate plans; per-token billing is for the cases that need it.

Should you buy hardware, rent GPUs, pay per token or subscribe? That math has its own post: AI Costs in 2026.

Stop guessing what your AI costs. Start seeing it.