Update
Credit costs, measured: what a task really costs on every model and for every kind of task
TL;DR
Credit costs are now published from a measurement rather than an estimate: 337 completed production tasks over 90 days, re-priced at today's token prices. Two new pages show the result by model and by task type, and the pricing page's per-task dollars were wrong and are fixed.
What the measurement found
- A median task uses 349 credits; the 75th percentile is 809 and the 90th 2,927. The mean is 1,091, pulled up by a few long research and video runs.
- The plan pages' 1,350–1,900 credits per task is a cautious budget, not the typical task: 82% of measured tasks used 1,350 or fewer and 87% used 1,900 or fewer. It stays the basis for "≈ N tasks a month", so nobody runs out sooner than a plan page says.
- The model changes the bill by up to 422×. The same median task costs 21 credits on GLM-4.7 Flash, 29 on GPT-6 Luna (the default), 1,397 on Claude Opus 5 and 8,872 on GPT-5.5 Pro.
- Input tokens are most of an agent's model cost, not output. The median task read 32 input tokens for every output token; output was under half of the model cost on all 49 models (about 29% on the median one). A model's input and cache-read prices matter more than its output price.
- Tools were 15% of spend (search, the Python sandbox, a hosted browser, image and video); 43% of input tokens came from the prompt cache at the cache-read rate.
What changed on the site
- Credits by model prices the measured median, small and large task on every selectable model, recomputed daily from each model's live token price.
- Credits by task type shows the measured distribution for changes, autonomous runs, reports, multi-platform analysis, SEO research, dashboards and video, plus every tool's price per call.
- The pricing page said a 1,350–1,900-credit task "works out to roughly $0.40–$0.80" at $0.001 a credit — a third of its own arithmetic. It now multiplies: $1.35–$1.90 on Starter, $1.13–$1.58 at Agency's rate, $1.62–$2.85 on top-up credits.
- /docs/model-pricing called its dollar column the provider's own price. It is credits × the Starter price per credit, and now says so.
The measurement script is in the repository (scripts/diag-credits-by-task-type.mts): model rows are re-priced from their own tokens, tool rows converted from the credit basis they were billed at, BYOM rows left out.
Frequently asked questions
- How many credits does an Agent Planners task use?
- Measured on 337 production tasks: 349 at the median, 809 at the 75th percentile and 2,927 at the 90th. The plan pages budget 1,350–1,900 per task, which 82–87% of tasks stay within.
- What does a task cost in dollars?
- Credits × your plan's price per credit — $0.001 on Starter. A median task is about $0.35; the plan pages' 1,350–1,900-credit budget is $1.35–$1.90 on Starter.
- Do output tokens dominate the cost of an AI agent?
- Not for agent work. Measured on production tasks, the median task read 32 input tokens per output token, and output was under half of the model cost on all 49 models — about 29% on the median model. Input and cache-read prices matter more.
- Which model is cheapest for an agent task?
- GLM-4.7 Flash at 21 credits for the median measured task, then Qwen3.5 Flash and GPT-6 Luna (the default) at under 30. Frontier models cost 1,000–9,000 for the same task.