People keep asking me what it costs to use AI. And the honest answer is: it depends. On what you're doing, and what you're doing it with. Which is true, and completely unsatisfying when you just want a number. So I thought I'd give a Big Mac index a go and pin one down.
Hold one realistic coding task fixed, change only the model, and whatever's left in the price is the model and nothing else. I update it as prices and models move.
The PointFive Coding Task Index, current as of July 2026.
What it measures
The Big Mac index prices the same burger across countries to compare currencies. Here is what that actually looks like:
The same burger runs from $2.62 in India to $9.12 in Switzerland, and about $4.87 here in Australia. This index does the same thing with one coding task across models. The task is held constant, so the spread you see is purely the model.
My standard task is one realistic piece of agentic work: read a few files, reason, then write and revise a diff. I hold it fixed at 200,000 input tokens and 30,000 output tokens once you count tool calls and context. That ratio is deliberate, because the reading pile dwarfs the writing pile, which is exactly how these agents work in practice.
How AI agent input vs output pricing works
A lot of people don't get the input vs output pricing model. Put simply, input is everything the model has to read before it answers, and output is everything it writes back.
Let's picture a lawyer who bills at two rates. Everything you hand them to read, the contract, the case file, the pile of discovery, is billed at the reading rate. Everything they write for you, the brief, the opinion, is billed at the writing rate, and the writing rate runs about five times the reading rate, because writing is the slow, skilled work and reading is fast and cheap. Coding is the same shape. Input is the brand guide, the existing codebase, and the reference files you point the agent at. Output is the actual code it writes.
The twist that trips people up is that with coding agents the reading pile is enormous. To make one small change, the agent often reads a big slice of the repo, the file history, and the results of its own tool calls, easily 200,000 tokens of input, then writes back a 30,000-token diff. So even though writing costs about five times more per token, you can still pay more for the reading, simply because there's so much more of it. That's also why caching matters. When the agent re-reads context it saw a minute ago, that re-read bills at roughly 10 percent of the normal reading rate.
The index: cost of one standard task
Per-million-token rates, current as of July 2026. The task is fixed at 200K input plus 30K output.
Same task, a five-times spread from cheapest to priciest, and the only thing that changed is which model did the work.
Sonnet 5 is at an introductory $2/$10 through August 31, 2026, which prices this task at about $0.70, below Gemini and GPT, until it returns to the standard $3/$15, and about $1.05, on September 1. The chart shows the standard rate for continuity across quarters.
What a day and a team actually cost
One task is about a dollar, and that is the number that fools everyone. Nobody works in tidy one-dollar units. Real agentic sessions are long and messy. The agent reads more, retries, and reruns its own tool calls, so a real day burns through far more than ten clean tasks would suggest.
Independent field data from DX puts a real active engineer at $200 to $600 a month once you count actual usage, not the sticker. That works out to $10 to $30 a day, and the model you run is what moves you across the band.
For a team of ten that is $2,000 to $6,000 a month, and higher for teams that reach for a frontier model by default. But even those are averages, and the average is the wrong thing to trust. Sessions vary wildly, and a single all-out day on a frontier model can run past what any monthly average implies. One of my own days, priced at API rates, came to about $240, almost all of it one frontier model doing work a cheaper one could have handled.
That is where the bill actually runs away: the heavy user who lives in a frontier model, and the rare runaway session, the kind that turned into an $81,267 week. The per-task price is cheap. The habit is not.
How I built it
- The task is fixed. 200,000 input tokens plus 30,000 output tokens, meant to stand in for a single agentic coding task. Cost is (input divided by 1M times input rate) plus (output divided by 1M times output rate).
- Rates are published API rates for each model, the same basis Copilot, Cursor, and Claude Code now bill on, so the figure is tool-agnostic and isolates the model.
- One caveat on tokens. Tokenizers differ and some models are more verbose, so the same prompt produces slightly different token counts on each. The numbers above use a fixed token count for a clean comparison. Sonnet 5's newer tokenizer, for instance, can run 1.0 to 1.35 times longer for the same text.
- What this does not capture. Caching pulls real cost below these figures, since re-read context bills at roughly 10 percent of the input rate. Task size varies widely, from cents for a quick edit to several dollars for a big agentic refactor, so the fixed task is a midpoint, not a ceiling. And a subscription plan like Claude Code Max caps spend regardless of tasks, which makes the per-task API math an upper bound for those users.
Why this is the number I care about
The cheapest path is rarely a different editor, because at the token level the big tools all bill at the same published rates. The lever is the model. The rule of thumb: match the model to the difficulty of the task. Use a small, fast model for drafting, reformatting, and straightforward edits, and save the frontier model for the genuinely hard work that has to be right the first time. When I looked at a full day of my own use, the frontier model was under half my work but nearly all my cost.
The moment you can see one team running a frontier model for work a lighter one would handle fine, the saving stops being a mystery and turns into a decision. Getting that view, across every tool and every model, attributed to teams, is what we built TokenShift to do. The whole point is to see and price the layer nobody else is looking at.
FAQ
What is the PointFive Coding Task Index? Original data that prices one fixed coding task, 200,000 input and 30,000 output tokens, across the leading models at their published rates. It isolates the single biggest cost lever, which is model choice, and I update it as prices and models move.
How much does an AI coding task cost by model? On the standard task, about $0.35 on Claude Haiku, $0.76 on Gemini 3.1 Pro, $0.77 on GPT-5.3-Codex, $1.05 on Claude Sonnet 5, and $1.75 on Claude Opus 4.8. Same work, a five-times spread, set only by the model.
What does a day or a team actually cost? One task is about a dollar, but real days are not tidy. Field data from DX puts a real active engineer at $200 to $600 a month, which is $10 to $30 a day, and a team of ten around $2,000 to $6,000 a month. Model choice moves you across that band. And the average hides a tail: a single all-out day on a frontier model can run into the hundreds. One of mine came to about $240.
Why does the model matter more than the tool? Copilot, Cursor, and Claude Code all bill at the model's published API rate, so the same model costs about the same in each. The gap between tools is packaging. The gap between models is five times or more.