Most of your credits go to your agent's replies, but they aren't the only thing that draws on your balance. A few training and testing tools use credits too, and knowing which ones helps you read your usage ledger and avoid surprises.
This article lists everything that uses credits today, how each is charged, and what you can do to keep it in check. For plans, allowances, and auto-recharge, see Plans, Credits & Usage.
Agent replies
What it is
Every turn where your agent's model runs in a live chat - and every reply you test in Playground.
How it works
Each turn costs the credits shown for the agent's AI Model: GPT-5.6 Luna 1, GPT-5.6 Terra 3 (the default), Claude 4.6 Sonnet 3, Claude 4.6 Opus 5, Claude 4.8 Opus 6, and GPT-5.6 Sol 10. A turn still costs credits when the agent looks at a notify-moderators interaction card and decides to stay silent, because the model ran.
Best practices
Model choice is the biggest lever. See Choosing an AI Model before moving a high-volume agent to a pricier model.
Conversation imports
What it is
Reading past chats in Training: Conversations and turning them into suggested Q&A pairs. (Conversations aren't training data - they only produce suggestions, and only the Q&A you approve becomes training.)
How it works
Imports are charged at what the processing actually cost, plus 10%, with a minimum of 1 credit per import run. For eWebinar imports, Max credits to use sets a budget (it defaults to your remaining credits) and the import stops when that budget is used, working newest conversations first. Ongoing eWebinar sync imports newly archived, human-helped conversations in batches rather than one at a time, which keeps the per-run minimum from adding up.
Best practices
Start with a shorter window and a modest budget, review the suggestions, and widen Since later if they're useful.
Evaluation runs
What it is
Scoring your agent against past conversations in Evaluate.
How it works
Each run is charged at the measured cost of generating and judging the replies for test cases that succeeded, plus 10%, minimum 1 credit per run. Failed test cases aren't charged.
Best practices
Keep evaluation sets focused on representative questions; a few dozen good conversations usually tell you what you need.
Check training on Needs answers
What it is
The Check training (or Generate answer) button on a Needs answers card, which asks eChat whether your current training already answers that question.
How it works
Each check uses 1 credit when it actually runs. If a result was already worked out for that card, showing it again doesn't charge again.
What doesn't use credits
Adding web pages, uploading documents, writing quick docs, writing or importing Q&A, and reviewing suggestions don't use credits. They use training storage instead - and suggestions only start using storage once you approve them. Raw imported conversations never count toward storage.
Seeing it in your ledger
Every charge above shows in your usage ledger under Subscription, labeled with where it came from - the eWebinar chat, an evaluation, a conversation import, or an AI answer check - and with the model that ran. On paid plans, all of them count toward auto-recharge.