Agents now remember
Work with files, without the friction
Mention any file or integration inline with @, browse all of an agent's files in one view with indicators for new and changed files, upload .zip archives, and download attachments with a single click.
Agents on less common cadences (monthly, several days a week, or every few hours) now show their real schedule everywhere, and tool approvals default to Always allow, with Allow once in the dropdown, so repeated approvals no longer interrupt your flow.
Agents remember across runs
Each agent now has its own durable, private storage that persists across runs and conversations, so a scheduled or repeated agent can keep lookup tables, cached data, and working notes between sessions instead of starting fresh. You can see what an agent has stored in the workspace tab inside a session, or in the agent's configuration. Runs can also continue for up to five hours, so longer pieces of work have room to finish.
A rough guide to what the Agent costs
We've run fresh benchmarks on typical requests. Sonnet is generally cheaper than Opus for most text work. To help you get a feel for credit usage, here's a rough guide per action.
Action | Opus | Sonnet |
Simple data question | ~150 | ~90 |
Median data question | ~300 | ~180 |
Deep analysis | ~1,000 | ~600 |
Generate image | ~100 | ~100 |
Generate video (8s) | ~2,000 | ~2,000 |
Map warehouses to storefronts
Warehouse-to-storefront mapping is now available to everyone in Settings, so you can align inventory locations with the storefronts they serve.
Other additions
Agent can now read and adjust your Causal Factor Attribution (CFA) configuration.
Sonnet 5 is now available as a faster, lower-cost option for most text work.
Files the Agent creates while working now show up in its file list automatically at the end of a turn, instead of only being reachable mid-conversation.
Question sets from the Agent can be collapsed, so a long set of options no longer crowds out the conversation.