> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.talkif.ai/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.talkif.ai/_mcp/server.

# AI turn-taking, the Talkif MCP server, charge history and number usage

* **Calls**: agents now make small, fast AI decisions about interruptions and turn-taking, on every call:
  * When the caller talks over the agent, the agent lowers its voice while their first words are checked. A short "mm-hm" or "right" lets it carry on; a real interruption stops it.
  * When the caller cuts the agent off with something that needs no answer of its own, the agent goes on with what it was saying instead of starting over. A clarifying question gets a short answer first. If the caller cuts in with "one moment", the agent waits quietly for up to a minute, then checks in.
  * The end of the caller's turn is judged from their words as well as their voice. A caller who stops mid-sentence gets 2 seconds to go on. A caller who ends their own turn asking for a moment gets up to 10 seconds.
* **Calls**: the agent's speech is generated about one sentence ahead of what the caller hears, instead of the whole reply at once. When the caller interrupts, only the sentence or two already generated is billed as voice, not the rest of the reply.
* **Calls**: when a call reaches its maximum duration, the agent stops mid-sentence and says its closing line. Before, the call could end without a goodbye.
* **Calls**: a call in which neither the caller nor the agent says anything for 2 minutes now ends after 2 minutes, not 5.
* **Calls**: browser calls, team-member rings and WhatsApp calls that fail while being set up end with the real reason — `network_error`, `canceled`, `no_answer`, `rejected` or `system_error` — instead of `bot_unavailable`. They no longer count toward the **Agent is not picking up** alert.
* **Builder**: fixed a bug where a test call on a slow network could stay silent until it timed out. It now waits up to 8 seconds to find a network path, and fails with a reason if there is none. See [Test and publish](/build/test-and-publish).
* **Billing**: AI decisions are billed per call as their own **AI Decisions** charge, by the tokens they use (charge type `ai_decisions`, cost category `ai_decision`). See [Cost breakdown](/account/cost-breakdown#ai-decisions).
* **Billing**: in **Billing → History**, a charge expands to show its line items with usage, rate and provider. Filter by date range, charge type and status. **Export all** exports every matching charge, not only the page on screen.
* **Billing**: a call is priced with the models of the flow version it ran on — for a builder test call, the version you tested — instead of the flow's current published models.
* **Models**: new — Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5 (`claude-opus-5-5`, `claude-sonnet-5-5`, `claude-haiku-5-5`), GPT-6 Sol and GPT-6 Luna (`gpt-6-sol`, `gpt-6-luna`), Cartesia Sonic 3.6 (`sonic-3.6`) and xAI Grok Voice Transcribe 2.0 (`grok-voice-transcribe-2.0`). New Claude nodes in the builder start on Claude Haiku 5.5. Post-call analysis now uses Claude Haiku 5.5 by default, at a tenth of the per-token price of Haiku 4.5. See [Choose models and voices](/build/choose-models-and-voices).
* **Models**: superseded models were retired on November 1, 2026 and are no longer offered. Flows that still use one keep working: at call time they run on its replacement. Saving, test calls, publishing and assistant edits refuse a retired model with `invalid_model`, so to edit such a flow, first choose a current model on each node that names a retired one. Older GPT-4o, GPT-4.1 and GPT-5.x models move to GPT-5.6 Luna, Terra or Sol; older Claude models to the Claude 5.5 model of the same tier; Deepgram Nova-2 to Nova-3; Cartesia Sonic 3 to Sonic 3.6; `grok-3-mini` to `grok-4.3`. This brings forward the later dates announced on October 3 for `gpt-5`, `gpt-5-mini`, `gpt-5-nano`, `gpt-5.1` and `gpt-5.4-nano`. The models endpoint names each retired model's replacement. See [When a model is retired](/build/choose-models-and-voices#when-a-model-is-retired).
* **Models**: Google Gemini models and Rime voices are not available for new selections.
* **Billing**: on Deepgram Nova-3, a flow's vocabulary is now sent as keyterms; Nova-3 doesn't accept the older keyword list. Deepgram charges keyterm prompting as an add-on, so a Nova-3 call on a flow with vocabulary costs \$0.0013 more per minute on its speech-to-text line.
* **Telephony**: a number's details show what uses it: the agent that answers it, the Meta lead forms that use it as caller ID, and the campaigns and schedules that call from it. Releasing a number shows the same list first and says what will stop working. A number that is a lead form's caller ID can't be released until no form uses it. API: [`GET /phone/numbers/{id}/usage`](/api-reference/phone-numbers/get-phone-number-usage). See [Buy a number](/telephony/buy-a-number#what-uses-a-number).
* **Telephony**: new number limits replace the limit of 2 numbers for accounts younger than 30 days. An account can hold up to 2 rented numbers until its first payment, and up to 10 after. At most 5 numbers can be bought in any 24 hours, counting numbers you have since released (`429 phone_purchase_rate_limited`). See [Buy a number](/telephony/buy-a-number#who-can-buy-and-how-many).
* **Billing**: free and bonus credit can pay for phone numbers once the account has made a payment. Before that, a number purchase that the paid balance doesn't cover is refused before any number is bought. See [Buy a number](/telephony/buy-a-number#paying-for-it).
* **Telephony**: number search offers only the number types a country sells. The [available-countries](/api-reference/phone-numbers/list-available-countries) response lists each country's `number_types`, and searching for a type the country doesn't sell returns `422 phone_number_type_unavailable` instead of a server error.
* **Telephony**: fixed a bug where releasing some numbers in the EU region failed, and so did the automatic release of unpaid numbers there.
* **Telephony**: numbers from your own Twilio account stay in your Twilio account when you delete your Talkif account; before, they were released. Removing such a number from Talkif, or deleting the account, also removes the voice webhook and status callback Talkif set on it, if they still point at Talkif. See [Bring your own SIP trunk](/telephony/bring-your-own-sip-trunk#instead-your-own-twilio-account).
* **Account**: **User menu → Security → Delete my login** deletes your own login — you, not a workspace — after 7 days, and you can cancel until then. If you own an organization, make someone else its owner first. The account deletion screen now says billing records are kept, as the law requires.
* **Dashboard**: the **Voice agents** list searches and pages through all your agents (name, description, tags) instead of filtering the first 20. The agent pickers for campaigns, schedules, quick calls, calling a contact and Meta lead form mappings list up to 100 agents.
* **Account**: fixed account deletion: a deletion that stopped part-way now finishes, a cancelled deletion is never carried out later, and the owner keeps their sign-in until the account's data has been deleted. See [Data export and deletion](/account/data-export-and-deletion).
* **Integrate**: the Talkif API is available as an MCP server at `https://api.talkif.ai/api/v1/mcp`, so AI assistants and agents can work with your account using an API key. The API operations the key's scopes allow become tools; `?read_only=true` and `?toolsets=…` narrow the list. Each tool says whether it only reads, and tools that delete, spend money or reach people outside the account are marked, so the AI app can ask before using them. Each key can make up to 60 tool calls per minute. See [MCP server](/integrate/mcp).
* **Integrate** (experimental): apps and AI agents that support MCP sign-in can reach your Connectors by signing in with Talkif (OAuth) instead of using an API key. You approve the app in the dashboard and pick the account; Connectors must be on for that account. A connected app acts for the person who approved it, with that person's role: it reaches the connected sessions they can reach (for owners and admins, every connected session in the organization), and can ask them questions and give them tasks. See [Sign in with Talkif](/integrate/mcp#sign-in-with-talkif).
* **Integrate** (experimental): connected apps are listed under **Developer → Integrations → Connected apps**. The person who connected an app can rename it, change who may use it and choose how it is woken (webhook or polling). Owners and admins see every app connected to the account and can disconnect any. See [Connected apps](/integrate/mcp#connected-apps).
* **API**: in the dashboard, an API key with **Full access** can now also get the **Connectors** scope, which Full access doesn't include. See [Authentication](/integrate/authentication).
* **API**: an unknown, revoked or expired API key gets `api_key_invalid`, `api_key_revoked` or `api_key_expired` (401) instead of `token_missing`. See [Error codes](/api-reference/errors).
* **API**: [`GET /flows`](/api-reference/flows/list-flows) takes a `search` parameter: a case-insensitive match on a flow's name, description or any of its tags. The total counts only the matches.
* **API**: [`GET /calls/history`](/api-reference/calls/list-call-history) (same result as `GET /calls`) and [`GET /calls/active`](/api-reference/calls/list-active-calls) (live calls only) are still served and are back in the public API definition. For new integrations, use [`GET /calls`](/api-reference/calls/list-calls).
* **API**: [`POST /calls/{callId}/end`](/api-reference/calls/end-call) on a call that has already ended returns the call's state instead of an error.
* **API**: the [`POST /calls/webrtc`](/api-reference/calls/create-webrtc-call) response can include `offerDeadlineAt`, the time until which the agent waits for your SDP offer. Send the offer well before it.