Changelog

The minimum amount for a custom credit recharge is now $10, up from $5. Preset recharge amounts are unaffected.

When you pick a voice, the preview now speaks a natural line in that voice's language instead of a mismatched sample, with localized lines for 42 languages and a more conversational Hindi line. Previews that were missing audio now regenerate automatically the first time you play them.

The picker is smoother too: previews load with a clear spinner, don't re-synthesize on replay, and stop when you close the voice picker.

We reworked the dashboard and site for small screens: a redesigned navigation drawer, larger tap targets, forms and dialogs that fit the viewport, loading skeletons that match the real layout, and no more clipped or overflowing content. Agent configuration, call logs, billing, and messaging are all now comfortable to use from a phone.

Invoices and statements now print the buyer's Tax ID or VAT number and region-appropriate seller details such as HSN/SAC, CIN, and PAN, alongside a running credit balance. Tax fields adapt to the customer's region instead of assuming one format.

Importing a SIP trunk is smoother: a redesigned dialog, automatic carrier selection, clearer guidance when something is wrong, and a clean rollback so a failed registration never leaves a half-added number behind.

Text and email conversations now show a clear "Not delivered" status when a message fails, so you are not left guessing. Contacts can opt out of email, and opted-out threads stop sending automatically.

Agents now let you choose how aggressively to clean up background noise on a call. Open an agent's Speech to text settings and pick Standard or Advanced noise reduction.

Standard suits most calls; advanced helps when the person on the line is in a noisy environment. Leave it on the default if you are not sure.

You can now apply a promo code when buying extra concurrent-call capacity for your organization, the same way you already can on plan upgrades.

The discount applies to your wallet charge; the listed price stays accurate for your records.

Your knowledge base can now pull accurate text out of scanned documents and tables, not just standard text-based files. Upload the same way you always have, under Knowledge Base > Upload.

Search and retrieval across your existing documents are also more accurate as a result.

Invoices are easier to read. Every invoice now shows your credit balance before and after it, and line descriptions spell out the plan, billing period, and top-up amount instead of a bare total.

Invoices still being generated appear as a processing row, so you know one is on the way instead of wondering where it went. Find them all under Billing > Invoices.

The OmniDimension MCP server, which connects your agents and calls to Claude and other MCP clients, picked up several upgrades:

• Tools are now labelled by what they do, so clients can tell read-only actions apart from ones that change or delete data before running them.
• Voice and language details are available as reference resources, making it easier to pick the right voice for each language.
• Failures are reported clearly, so a failed call no longer looks like it succeeded.

The TypeScript SDK types have also been refreshed to match the latest API.

Creating a WhatsApp message template that Meta rejected used to show a generic failure. The exact reason from Meta now appears in the error, so you can fix the template and try again without guessing.

Opening Phone Numbers used to slow down as you added more numbers. It now loads each number's connected agent in a single pass, so the page opens quickly no matter how many numbers you have.

A failed LLM attempt no longer triggers the pipeline kill path while failover is mid-recovery. Per-attempt failure signal and cooldown raised to 3 strikes for steadier recovery. Welcome-turn TTFT grace stops a pointless failover storm during the greeting.

Recording ingestion now runs as a dedicated background service. Recordings are stored immediately after a call ends, and slower carrier fetches are processed on a separate queue so they never delay anything else. If a recording arrives before the call record is ready, the system retries automatically.

All recordings are stored and accessible from the same place, regardless of carrier.

Post-call processing now runs as a dedicated background service. After a call ends, AI extracts a summary, sentiment, and any custom variables you've configured - then delivers the results to all connected destinations: Slack, email, webhook, CRM, WhatsApp, and Google Sheets.

Google Sheets delivery goes through a dedicated queue worker. Rows are written with full call metadata, new columns are added automatically as your data evolves, and token refreshes happen transparently. Any result that doesn't land is detected and retried automatically - no dropped data.

We reworked how bulk campaigns run. Calls now dispatch in parallel, so large campaigns start faster, and the system no longer falsely marks active calls as failed under load.

Calls that get stuck mid-flight are now detected and recovered automatically, and you can retry failed lines with a single button.

We rolled out the Sonic-3.5 voice model across all bots. The upgrade brings sharper intonation on emphasized words, lower time-to-first-byte for new utterances, and more natural pacing on longer responses.

No action needed. Existing bots pick up the new model automatically. Voice names, IDs, and the cloning pipeline are unchanged.

 Feature Details

• Notification tone
  
  
• Phone number banner

Feature Details

• Numbers shop documentation
  
  
• Number purchase banner
  
  
• Agent Management APIs for IVR
  
  
• Improved Asterisk templates in Omnidimension
  
  
• Responsive UI fixes (tabs)

 Feature Details

• UI fixes on Phone Numbers page
  
  
• Bug fixes for number purchase and release flows

Feature Details

• Phone number purchase