A model nobody priced was a model with no limit: an unpriced call was logged at zero and counted nothing against the user's monthly budget, and one such model was used daily by a script. This release closes that gap and, on the way, makes every AI call answer the question "who asked, and why".
Runs
- One run per cause A chat conversation, an agent run, a server script or job, a documentation action, or an external agent reaching in over MCP. The run carries its kind, its reference, the feature calling the model, the agent running and the tool call in progress, and survives thread hops.
- Parents and children A model call made by a tool is the child of that tool call; a sub-agent's calls are marked as such. Chat tool calls are audited with the statements they ran.
- Everywhere A script's AI calls keep the run that started it or open their own; page improve and translate in the manual are runs; chat, MCP and the compaction below all run inside one.
Price
- Cache and thinking priced Prompt-cache writes and reads and thinking tokens carry their own price, and the input count means one thing on every provider. Two providers include the cached part in their prompt count and would have been charged twice.
- No price, no call A billable call to a model with no price is refused with a message naming the model and asking an administrator to price it. Refusing it is what gets it priced. A call the provider does not charge is never asked.
Memory
- Compaction Past 60,000 tokens a conversation is folded into a summary row on a detached thread, keeping the newest 16,000 tokens verbatim, so the user never waits and the next turn opens on the summary.
The user console now shows AI and external agents in one section, with a flow diagram that lights the path a call takes and says who pays for it.