AI Cost Benchmarks: What Real Tasks Actually Cost

Published prices are quoted per million tokens, which tells you almost nothing about your bill. This page converts those prices into the thing you actually want to know: what does one real job cost? Eight common tasks, priced on all 17 major models.

Calculated from published list prices. Last verified: August 2026. Every token assumption is shown below — check our arithmetic.

Cost per task, by price tier

One model from each tier, so you can see the shape of the trade-off before reading the full matrix. Figures are for a single run of each task, except the last two, which are already at scale.

What these numbers actually tell you

The spread is enormous — and it is not subtle. Running that support chatbot costs $1.52 a month on Gemini 2.5 Flash-Lite, $80.00 on Claude Opus 4.8 in the table above, and $160 on Claude Fable 5, the dearest model we track. End to end that is roughly 105× — the same feature, the same users, a different line on your P&L.

Output is where the money goes. Every model charges several times more for output than input. Look at the two extremes: summarising a 10,000-word report sends 13,300 tokens in and gets 400 back, while drafting a blog post sends only 250 in and produces 1,350. The drafting job moves far less text, yet on Claude Opus 4.8 it costs $0.035 against $0.076 for the summary. Reading is cheap; writing is not.

Multi-turn conversations compound. The chatbot row looks disproportionate because it is: most APIs are stateless, so the entire conversation so far is re-sent with every turn. Eight turns does not cost eight times one turn — it costs closer to thirty, because turn eight carries the whole history. Teams budgeting from a single-message test are usually out by an order of magnitude.

The cheapest model is often enough. Extraction, classification, routing and summarising are largely solved at the budget tier. Frontier models earn their price on hard reasoning, long-horizon agent work, and anything where a subtle mistake is expensive. Paying frontier rates to reformat invoices is the most common way to overspend on AI.

All 17 models, four everyday jobs

Same arithmetic, every tracked model. The first three columns are per run; the chatbot column is a monthly total.

Claude Sonnet 5 is shown at list price; an introductory rate of $2.00 / $10.00 runs through 31 August 2026, which lowers every figure in its row by about a third. Need different assumptions? The AI API cost calculator takes your own numbers.

The assumptions, in full

These figures are arithmetic, not measurements: list price × token count. That makes them reproducible, and it makes the token counts the part worth arguing with. Here they are.

Two conversions do most of the work: English prose runs about 1.33 tokens per word, and source code about 14 tokens per line. Both vary with the tokeniser and the subject matter — expect ±15% on any single job. To count tokens in your own text exactly, use the token calculator.

Three things that move the real bill

The table above is the list-price ceiling. In production, three mechanisms move the actual number:

• Prompt caching (up to ~90% off repeated input). If a long system prompt or document is re-sent across calls, most providers will bill the cached portion at a fraction of the input rate. On the chatbot row — where the same history is re-sent every turn — this is the single largest lever available.
• Batch processing (~50% off). Work that need not be answered immediately, such as the invoice and meeting-notes rows, can be submitted as a batch and returned within hours at roughly half price.
• Long-context surcharges. Several models change rate above about 200,000 tokens of context. If a job routinely exceeds that, check the provider's tier boundary before assuming the headline rate applies.

Current list and effective prices for every model are on the AI price tracker.

How to use this page

Find the row closest to your workload and read across. If the cheapest tier is affordable, prototype there and only move up when you can point to a specific failure the bigger model fixes. If the frontier tier is affordable at your volume, the decision is easy and cost is not your constraint. The uncomfortable middle — where the cheap model nearly works — is where caching, batching and a smaller prompt are worth more than a model upgrade.

Then keep an eye on the date at the top. The price of a given level of AI capability has been falling roughly an order of magnitude a year. A budget built on this table will be pessimistic within months, which is the good direction to be wrong in.

The guide that goes deeper

You might also need