Changelog

• API: correction — GET /flows/{flowId}/stats now returns rates.successRate and rates.failureRate as percentages from 0 to 100, as documented. They were returned as fractions from 0 to 1, so a flow where every call completed reported 1 instead of 100. If your integration multiplied these values by 100 to work around this, remove that step.
• Dashboard: the Success rate figure on the overview shows the real percentage. It used to show 0% or 1% whatever the real rate was.

• Account: sign in with a passkey — Touch ID, Face ID, Windows Hello or your device PIN, saved in iCloud Keychain, Google Password Manager or your password manager. A passkey sign-in skips the 2FA code. Add one under User menu → Security → Passkeys; you get an email whenever a passkey is added or removed. See Passkeys.
• Account: Continue with Apple on sign-in and sign-up, with Hide My Email support. Existing users link Apple from User menu → Security; accounts are never merged by email. See Sign in with Apple.
• Security: Google sign-in now asks for the two-factor code on users with two-factor authentication turned on.

• Models: Eleven v4 Turbo (eleven_v4_turbo) is available — ElevenLabs’ newest real-time voice, with Turkish and 90+ other languages and inline audio tags. It is now the default voice for new agents and templates. See Choose models and voices and Write prompts.
• Models: model retirements are now announced in advance. Every model scheduled for retirement carries a deprecation object (retiresOn, replacedBy) in the models endpoint. Until that date it works as before; from that date calls on it run on the named replacement from the same provider, without republishing. See When a model is retired.
• Models: scheduled retirements — ElevenLabs Flash v2.5, Turbo v2.5, Multilingual v2, Turbo v2 and v3 Conversational move to Eleven v4 Turbo on 13 October 2026; Cartesia sonic-2 and sonic-turbo move to sonic-3 on 20 October; OpenAI gpt-4.1-nano moves to gpt-5.6-luna on 23 October; gpt-5, gpt-5-mini and gpt-5-nano move to the GPT-5.6 models on 11 December; gpt-5.1 and gpt-5.4-nano retire on 1 April 2027. Models their providers already shut down (gpt-5.2-chat-latest, Grok 4.1 Fast, Rime Arcana) now run on gpt-5.6-sol, grok-4.3 and Rime Coda.
• Models: new in the catalog — grok-4.3 and Rime Coda.
• Billing: prices follow each provider’s current public list price — ElevenLabs Flash, Turbo and v3 Conversational $0.04 per 1,000 characters (was$0.05), Multilingual v2 $0.08 (was$0.10); ElevenLabs Scribe v2 Realtime $0.0065/min; Deepgram Nova-3$0.0077/min; Amazon Transcribe $0.010/min; Cartesia Sonic$0.038 per 1,000 characters. See Credits and pricing.
• Account: signing in with Google now opens every account you belong to, including organization accounts.
• Docs: dashboard paths updated to the current menu — Voice agents, Calls, Contacts, Campaigns, Schedules, Leads, Phone & WhatsApp, Settings → Compliance, and the user menu for Security and Sessions.

• Builder: a new Transfer node hands a live call from the AI to a person — warm (the AI briefs your colleague while the caller holds) or cold. Ring a phone number, your team, a team the AI picks from the conversation, a destination group, or your own app. Every transfer carries a handoff case (reason, summary, your own fields). If nobody answers, or your colleague drops right after connecting, the agent comes back to the caller and carries on. See Transfer calls to people.
• Settings: Settings → People & numbers holds transfer destinations (phone numbers, people in the dashboard, your app) and destination groups that ring together.
• Integrate: Your app destinations send your server a signed transfer.offer webhook. Your server accepts it with an API key that has the new transfer_offers scope, then joins the call audio over a websocket. See Receive transfers in your app.
• Calls: take transferred calls in the browser. Turn on Ring in this browser under the status button in the top bar, then accept from the incoming-transfer window with the handoff case on screen. Listen and Take over let you join a live call. See Take calls in the dashboard.
• Account: a new Team page lists members, invitations, and a Roles & permissions table. Members carry teams (tags), weekly available hours and a status (Available, Away with an end time, Offline). Owners and admins can set a member’s status. See Members and roles.
• Account: a fourth role, Agent, only takes transferred calls. Admins now manage API keys, the webhook signing secret, account settings, phone number configuration, webhooks, integrations and transfer destinations. Money, deletion, ownership and role changes stay with the Owner. Members keep the daily work but no longer change account settings or number configuration.
• API: new endpoints for transfer destinations and groups, member tags, hours and availability, and transfer offers (transfers and transfer_offers scopes). Also new: GET /roles and GET /permissions. Embedded web calls use POST /public/calls, GET /public/calls/{callId}, POST /public/calls/{callId}/offer and POST /public/calls/{callId}/end; the older /public/calls/calls/… paths keep working, and @talkif/webrtc 0.1.6 switches to the new ones.
• Billing: AI usage stops when the agent leaves a transferred call. The outgoing leg to a phone-number destination is billed as telephony from answer. See Credits and pricing.

• Telephony: SIP trunk connections now handle outbound calls end-to-end without an extra telephony hop, and new SIP numbers go active immediately with no verification call. You can also set the outbound caller-ID format per SIP trunk provider. See Bring your own SIP trunk.
• Calls: Ending a call from an embedded web widget now hangs up right away instead of waiting on a network timeout. See Browser SDK.
• Builder: The AI Assistant keeps one conversation thread as you move between the dashboard and the flow canvas.
• Account: Adding a first phone number to your account no longer requires setting a password, including for Google-only sign-ins. See Security.
• Account: Data from a deleted account is now permanently erased 60 days after the deletion grace period ends. See Data export and deletion.
• Billing: Fixed a display bug where a real-time balance update could briefly show the wrong amount.

• Calls: Ending a web call now takes effect immediately instead of waiting for the connection to time out. See Browser SDK.
• Account: Adding your first phone number no longer requires setting a password — useful if you signed up with Google. See Security.
• Billing: Low-balance notifications are far less frequent, and accounts that have never added funds no longer get a false “balance depleted” warning. See Notifications.
• Billing: A phone number’s monthly rental now recovers correctly after a temporary payment failure — topping up reliably reactivates a past-due number instead of leaving it stuck. See Buy a number.
• Builder: The AI assistant keeps one continuous conversation as you move between the dashboard and the flow canvas, instead of starting over. See Assistant.

• Telephony: Searching for available phone numbers now returns full, correctly-typed results through the SDKs — previously the Python SDK could fail to parse the response and only 6 of 22 search filters (area code, region, postal code, capabilities, and more) were usable from the SDKs. See Buy a number.
• Builder: The assistant now keeps a single conversation as you move between the dashboard and the flow canvas, so following a “take me there” suggestion no longer starts a new, empty thread. See Assistant.

• SDKs: @talkif/sdk 0.1.1 (npm) and talkif 0.1.1 (PyPI) are published. Both auto-page list endpoints and expose typed errors — see SDKs.
• API: list endpoints now return typed responses (calls, contacts, campaigns, … plus meta) instead of untyped arrays; offset pagination starts at 0.
• Calls: the call list is now GET /calls (replacing /calls/history and /calls/active); use status to narrow to live calls. It gains filters — status, direction, source, phoneNumber, flowId, contactId, campaignId, scheduleId, providerType, isLead, search, and sorting.
• Integrate: Agent Skills for Claude Code, Codex, Cursor and any Agent-Skills-compatible tool — npx skills add Talkif-ai/skills — covering every job in these docs, verified against the public API. See Build with AI agents.
• Docs: docs.talkif.ai restructured around tasks — Getting started, Concepts, Build agents, Telephony, Run calls, Integrate, Account and billing, Reference. Every previous URL redirects.

• Builder: Prompts are now two clearly separated fields. The System Prompt is the call’s persona, set once on the Main Agent and inherited by every other agent; the Agent Prompt is each agent’s own job. Existing flows were migrated automatically. See Write prompts.
• Builder: Flows built by the Assistant now include spoken-output rules on the main agent by default (short replies, no lists read aloud).
• Models: Models no longer offered by any template were retired; ElevenLabs Multilingual v2 removed in favour of the v3 and Flash models.
• Calls: A contact skipped mid-campaign (Do Not Call, opt-out, manual skip) no longer keeps the campaign in running after every other contact is done.

• Calls: Campaign call windows are evaluated in each contact’s local time, so a list spanning time zones is called at sensible hours everywhere. See Campaigns.
• Calls: Campaign pacing is budgeted against your concurrent-call limit, and busy agents are protected during platform scaling — fewer mid-call drops under load.
• Builder: Campaign screens rebuilt around the new analytics; raw call errors no longer appear in the campaign summary.
• Models: ElevenLabs v3 conversational voices are available for text-to-speech.
• Calls: A call now ends cleanly, with a recorded reason, when a speech or language service becomes unusable mid-call, instead of hanging silently.
• Telephony: Meta Lead Ads form mappings can declare the phone country for their leads, so numbers submitted without a country code are normalised correctly.

• Builder: Models that support a reasoning effort setting now expose it, and every model defaults to the least thinking it allows — the right default for live conversation. Models that reject a temperature (current GPT-5 and Claude families) no longer show one. See Choose models and voices.
• Models: Catalogue corrected from live-call results; retired models removed; the Claude 5 family added.
• Calls: Campaign analytics gain outcome, attempt-distribution, failure-cause and duration-histogram views, and campaign counters no longer drift. See Campaigns.
• Telephony: WhatsApp calls from callers Meta identifies by business ID rather than phone number are now answered; a single unrecognised WhatsApp event no longer drops the events around it.
• Calls: Releasing a phone number no longer interrupts Meta Lead Ads calls already in flight for that number.

• Meta Lead Ads: A dedicated qualification pass decides each lead’s funnel stage, with AI stage suggestions behind a confidence gate, a hold window, and manual override from a needs-review queue.
• Meta Lead Ads: Per-form qualification criteria, a user-configurable call window and concurrent-call limit on the lead campaign; the campaign schedules its own lead calls.
• Meta Lead Ads: Drill-down navigation (Overview → Page → Form → Leads), source and integration filters, and per-lead actions: retry the Conversions API send, add to Do Not Call, schedule a call.

• Telephony: WhatsApp Business Calling — connect a WhatsApp Business Account (Embedded Signup or credentials), register numbers, and your agents answer WhatsApp calls with no telephony charge. The same digits can exist as a phone line and a WhatsApp number. See WhatsApp Business Calling.
• Builder: The Talkif Assistant — a chat in the dashboard that answers questions about your account, proposes flow edits as reviewable diffs, and builds a complete flow from a description.
• Models: Six new providers — Google Gemini TTS, Google Cloud STT, xAI Grok STT and TTS, Amazon Transcribe, Rime TTS — and the LLM catalogue refreshed with official pricing. ElevenLabs voice browser gains gender, age, language and accent filters; shared-library voices can be added to your account.
• Integrate: Web calls — put a voice agent on any web page with a publishable key and the browser SDK; and a real-time events WebSocket for account-wide and per-call events. See Real-time events.
• Calls: Failed calls carry the carrier’s real failure code and reason instead of a generic failed.
• Calls: Honest interruption accounting — the transcript shows what the caller actually heard, and wasted (generated but unspoken) TTS characters are reported per call. Per-contact call timeline; call history gains phone/name search, provider filter, lead filter and server-side sorting.
• Calls: Timezone-honest schedules — pick an IANA zone, preview in local time, default to the contact’s zone. See Schedules.
• Builder: End Call as a shared canvas node; provider nodes auto-connect to the Main Agent when dropped; per-flow turn-taking configuration.
• Billing: Trial credit, coupons and promotional balances; Assistant usage with a free monthly allowance, then billed to balance.
• Meta Lead Ads: Leads grouped under auto-created campaigns; dataset picker instead of pasting an ID; per-lead attempts, manual recall and a clickable funnel; moved to Call Center → Meta Leads.

• Builder: Silence handling (“are you still there?”, then a polite hang-up) on by default, and a per-flow maximum call length with a graceful goodbye. end_call now lets the farewell finish instead of hanging up instantly. See Transfer and end calls.
• Integrate: Webhook signing — every request to your endpoints carries a Talkif-Signature (HMAC-SHA256) once you create a per-account secret; secrets are rotatable with an overlap window. See Webhooks.
• Integrate: Meta Lead Ads — connect a Facebook Page, map lead forms to flows, and Talkif calls every new lead within seconds, reporting funnel stages back to Meta’s Conversions API. See Meta Lead Ads.

• Calls: Voicemail detection on outbound calls — recognise an answering machine and leave a message (fixed or composed) or hang up, with its own cost line. See Transfer and end calls.
• Calls: Single call state machine with an automatic reconciler — a call can no longer get stuck in an in-progress state; stuck calls are ended with reason reconciled.
• Calls: Admission is region-aware, so a call is only accepted where an agent can actually serve it.
• Calls: Lower browser-call connection latency.
• Developer: Account ID and Organization ID shown on the Developer page; link to API documentation.

• Builder: Flow functions — the agent calls your HTTP endpoints mid-conversation. Structured requests (path, query and body parameters as JSON Schema), with each parameter bound to the model, to call context (caller, call, flow) or to a static value; values bound to context are never shown to the model. Attach a function to any agent with the HTTP Request node. See Call your backend.
• Builder: Functions (called by the model) and actions (run automatically around an agent’s turn) are now distinct concepts.
• Models: Settings a model doesn’t support — a temperature on a reasoning model, for example — are stripped automatically rather than failing the call; the builder hides them.

• API: The public REST API at api.talkif.ai with OpenAPI documentation and an error catalogue endpoint.
• API: API keys with scopes, IP allowlists, expiry, usage tracking and error monitoring, managed under Developer → Credentials. See Authentication.
• Builder: Flow versions — every publish is an immutable version with notes, a semantic diff between versions, rollback, and phone numbers that follow latest or are pinned to a version. See Flows and versions.
• Builder: Voice browser with filters for TTS nodes; ElevenLabs shared voices; a prompt editor dialog on agent nodes.
• Calls: Unified call timeline with every speech event, function calls per turn, and the model used per message; redesigned call detail and cost breakdown; end reasons shown in the UI.
• Billing: Charge history — every billable event (call usage, number purchase and rental, analysis, recording storage, adjustments) with status and line items, alongside the transaction ledger. Payment-method health — cards show declined / expired / blocked with the decline reason, auto-recharge skips unusable cards, and cards can be labelled. See Credits and payment.
• Telephony: Redesigned number purchase; phone-number change with SMS verification.
• Calls: Faster browser test calls.

• Calls: Call recording with playback and download, per-recording storage billing, and a retention period you set. Delete a recording on demand. See Transcripts and recordings.
• Calls: Do Not Call list — numbers on it are never dialled, by any path; campaign contacts are skipped with the reason. See Do Not Call.
• Calls: Legal calling hours enforced by the recipient’s jurisdiction (US: 8 am – 9 pm local, TCPA), with timezone derived from area code when the contact has none.
• Builder: Dynamic prompts — {{contact.*}}, {{call.*}}, {{system.*}} variables with helpers, resolved at call time; AI-assisted prompt authoring; per-flow analysis instructions. See Write prompts.
• Builder: Start from a blank flow, or a template.
• Calls: Campaign contacts are snapshotted at enrollment — a contact edited mid-campaign is still called with the data the campaign started with, so every retry sees the same context.
• Telephony: Two regions (EU and US) — calls are served from the closer one; SIP numbers activate immediately (no verification call).
• Billing: Phone-number prices fetched live from the carrier per country; direction-based (inbound/outbound) telephony pricing; public pricing endpoint.
• Account: Data export — everything as one JSON file, email-confirmed, GDPR Article 20; account name and deletion from Settings → General; legal document acceptance with change notices.
• Dashboard: Onboarding checklist; toast notifications for every action.
• Security: SIP/Twilio credentials and two-factor secrets encrypted at rest; personal data redacted from logs.

• Calls: Campaigns — call a contact list with a flow, with concurrency, call windows, an end time and per-contact outcomes; skipped contacts are previewed with reasons before you start. See Campaigns.
• Calls: Two-layer call status: the telephony leg (status) and the agent leg (agentStatus) reported separately. See How a call works.
• Builder: Flow templates to start from; an End Call function for graceful termination; structured validation errors that name the field.
• Builder: OpenAI language models; Cartesia text-to-speech; ElevenLabs voice filtering; Soniox languages. Multi-provider STT/TTS per flow.
• Contacts: Import from CSV, JSON and vCard with phone-number normalisation and a country-code selector; tags with multi-tag filtering; export.
• Telephony: Caller ID on SIP calls; automatic carrier-edge selection from your carrier’s location; timezone detection from phone numbers.
• Billing: Auto-recharge; Stripe-hosted payment methods and 3-D Secure; cost breakdown per call with cached and reasoning token pricing for every LLM provider; telephony billed at the carrier’s actual per-call price.
• Account: Sign in with Google; organization invitations; personal and organization accounts.
• Calls: Faster first response — the greeting is prepared while the call connects; idle calls end automatically.

The current Talkif platform, rebuilt from the ground up. Everything below this entry is the product as it exists today; earlier history isn’t listed.

• Calls: Every call runs through a capacity-managed queue — calls beyond your concurrent limit wait and are placed automatically as slots free, in priority order. Call source (direct, scheduled, campaign) is tracked on every record.
• Calls: Post-call insights — summary, sentiment, intents, action items, outcome — generated automatically, with contacts created or enriched from what callers say.
• Calls: Full call details with related contact, campaign and flow, and per-service cost breakdown priced dynamically per model.
• Telephony: Bring your own Twilio account, or a SIP trunk; phone-number search with full filters (area code, city, state, ZIP, pattern).
• Account: Organizations with members and roles; account deletion with a grace period; multi-factor authentication.
• Dashboard: Real-time updates for call status, balance and notifications; analytics dashboard with call quality.
• Builder: Flow status lifecycle (draft → testing → published → archived) and flow tags.