Dashboard assistant, custom dashboards, per-request settings, and service tiers
New features
- Dashboard assistant. Open the assistant side panel to ask about your usage, spend, and logs, or to change settings in plain language. In Ask mode, you approve each change before the assistant applies it. In organizations, owners have access by default. Other roles need the assistant permission.
- Customizable dashboard. Add, arrange, and resize dashboard tiles, or start from a preset. Copy and paste tiles, undo your last change, pick the chart interval, and hover a chart to see values for a single bucket. The assistant can also build a layout for you.
- Unified Playground. Chat with one model or compare several side by side in one view. You can now attach PDFs, send prompts with no character cap, and send requests up to 50 MB. The Playground also shows which guardrails apply to your request.
- Per-request settings. Send
zdr,logging,redact, orroutingin the request body to apply them to a single request. A request can only make your key’s settings stricter. See Settings. - Service tiers. Set
service_tiertoflexfor lower prices orpriorityfor lower latency on supported models. See Service tiers. - Decisions API (beta). Get structured decisions from dedicated decision models with
/v1/decisions(OpenAI format) or/v1/systemone(TypeSafe format). See Decisions and System One. - API key expiration. Set an expiration date when you create or edit a key. The keys table shows when each key expires.
- Billing usage page. Billing → Usage charts your daily spend. The billing page now centers on your balance, merges all transactions into one filterable list, and shows auto top-up status on one line.
- Concentrate CLI (beta). Launch Claude Code, Codex, OpenCode, and pi through Concentrate with one command. See Concentrate CLI.
- More coding tool integrations. Set up OpenCode, Hermes Agent, and pi with a single script.
- Cost header. Every response now includes
x-concentrate-cost. Sendx-concentrate-compat-headersto also get LiteLLM or OpenRouter style headers. See Headers.
Updates
- New models and providers. Added providers including Baseten, Crusoe, Inception, Modal, Nebius, Relace, RunAnywhere, Runware, and Wafer, plus new models such as Kimi K3.1 Pro, GLM 5.3, DeepSeek V4.1 Flash, and Mercury 3. See Supported models.
- Clearer error codes. Errors now name the cause:
setting_enforced,model_not_allowed,provider_not_allowed,invalid_byok_credentials,byok_insufficient_credits,flex_capacity_unavailable, andupstream_rate_limit_exceeded. Unknown models return OpenAI-style404 model_not_found. See Errors. - Routing settings. Routing controls for keys, teams, and organizations include a provider picker and a live preview of where requests will go.
- Models pages. Models pages now show per-provider quantization, peak and off-peak pricing, a Flex/Priority filter, and a BAA-eligible filter.
- Keys and teams. Create and manage keys and teams in side drawers. Choose a key’s owner when you create it. The keys table adds a Created column.
- Invites. Invite emails appear as removable chips, and each invitee needs a role.
- Date range picker. The date range picker is redesigned across Dashboard, Logs, Errors, and Billing. Press Enter to submit forms.
- HIPAA. Web search runs in HIPAA compliance mode for HIPAA keys. In HIPAA scopes, the Playground lists only BAA-eligible models.
- Messages API. Anthropic’s advisor tool now works on
/v1/messagesfor supported Claude models.
Bug fixes
- Streams now end with a clear error when a provider fails mid-stream, instead of stopping silently.
- Model refusals are returned as refusals instead of
500errors. - Replies cut short by the provider are no longer reported as completed.
- An unknown
previous_response_idreturns400 previous_response_not_found, and earlier turns are replayed in order. tool_choice: "none"andparallel_tool_calls: falseare honored consistently across providers.- Requests with two or more tool calls on Bedrock no longer fail.
- Billing transactions and charts now show dates in your timezone. Auto top-up triggers only after billed requests.
- Editing a key no longer resets its usage. Disabled keys no longer appear in the Playground key picker.
- Teams, keys, and guardrails lists stay in sync after changes. A key reappears if deleting it fails.
- Modals and pages now render correctly in Safari.

