---
title: "Changelog"
description: "New models, API features and fixes for the Speko Router, Gateway and hosted Agents."
canonical: "https://speko.ai/changelog"
last-updated: "2026-10-15"
---
> ## Speko page index
> The complete index of every page on this site is at: https://speko.ai/llms.txt
> Read it before exploring further. It lists the exact Markdown URL for every canonical HTML page.
# Changelog
Product updates
New models, API features and fixes for the Speko Router, Gateway and hosted Agents, newest first.
34 entries
[RSS](https://speko.ai/changelog.xml) [JSON](https://speko.ai/changelog.json)
Nov 3, 2026 Improved Available
Agents Console
## The voice library lists about 800 voices with filters
The Voices page lists about 800 voices from 27 providers. You can filter them and sort them by measured pace and clarity.
- Filter by gender, age, pace, accent, quality and provider. Each filter shows the number of voices that match.
- When you select a language, the library shows native voices apart from multilingual voices.
- The agent editor opens the same library to select a voice.
- For a new agent with no voice, Speko now selects a voice that matches the gender of the agent's name. This works only for names that Speko recognizes. Speko does not change a voice that you selected.
- MCP clients can clone a voice with the voice_clones tools. Give a name, a language and a public audio URL, and confirm consent.
[Voices](https://speko.ai/voices)
Nov 3, 2026 Added Available
Router Models
## The Router adds nine models, and the STT benchmarks add Vietnamese
The Router added nine models from October 21 to November 3. The STT benchmarks now measure Vietnamese.
- LLM: claude-haiku-5-5 from Anthropic, gpt-6-luna and gpt-6.1-terra from OpenAI, and gemini-3.5-flash-lite from Google.
- LLM: NVIDIA Nemotron 3 Ultra on Baseten. Automatic routing does not select it. Pin baseten:nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B to use it.
- STT: stt-rt-v6 from Soniox and transcribe-stream-1 from Fish Audio.
- TTS: simba-3.5 from Speechify and assemblyai-tts-1 from AssemblyAI.
- gpt-6-luna, gpt-6.1-terra, simba-3.5 and assemblyai-tts-1 have no benchmark results yet. The Models page shows them as not measured.
- 19 Router STT models now have Vietnamese results. paxa-tts-flash-v2 now has TTS results, and the S2S board has new results.
- Munsit now also serves the us-east-1 and us-west-2 regions.
[Models](https://speko.ai/models)
Nov 2, 2026 Added Available
Agents Billing
## You can see your rates before a call
The dashboard shows what an agent minute costs, what each realtime model costs, and what a call costs in each country.
- The agent header shows the agent's rate per minute. The rate changes as you edit the agent, before you save.
- Open Settings, then Billing, then Rates. The page shows the agent minute, each realtime model and telephony. It includes your own discounts and the models that run on your own keys.
- The agent header and the Rates page use the same calculation as billing.
- The Call rates page, next to Place call on the phone numbers page, shows the price for each country. You can search by country name or code.
[Call rates](https://docs.speko.ai/concepts/call-rates)
Oct 31, 2026 Added Available
Agents Telephony
## You can upload a call list, and call out without your own number
You can upload a CSV and let one agent call every row. A workspace with no number can place outbound calls from a shared Speko number.
- Open Call lists in the sidebar, upload a CSV, then select the agent and the caller number. Speko calls a few rows at a time and shows the result of each row.
- The CSV needs a phone column. Each other column becomes a variable, so a customer_name column fills {{customer_name}} in the prompt.
- If the workspace runs out of credits, Speko pauses the list. The rows that remain do not fail.
- Leave out from on an outbound call, and the call uses a shared Speko number. When you buy or import a number, calls use your number. Inbound calls still need your own number.
[Call a list](https://docs.speko.ai/guides/phone-agents#call-a-list) [Calling without a number](https://docs.speko.ai/guides/phone-agents#calling-without-a-number)
Oct 29, 2026 Improved Available
Router Agents Gateway
## One reasoning effort setting works for every model that reasons
Agents and sessions accept one reasoningEffort setting. Each model uses the nearest value that it supports, and a model that does not reason ignores it.
- Set llmOptions.reasoningEffort on an agent, or llm.reasoningEffort or s2s.reasoningEffort on a session.
- It applies to Gemini 3.5 to 3.8, OpenAI gpt-5, gpt-5-mini and gpt-5-nano, Baseten models, Grok Voice 3.0, GPT-Live and Gemini Live extended thinking.
- The agent editor shows a Reasoning effort control when the pinned model has one.
- Grok Voice 3.0 accepts none or high. Set none to skip reasoning before each reply.
- The Router accepts reasoning_effort none or minimal on LLMs that do not reason. These requests stay routable, and the Router does not send the setting to the provider.
- The LiveKit and Pipecat LLM plugins accept reasoning_effort.
[Agents](https://docs.speko.ai/concepts/agents)
Oct 29, 2026 Improved Available
Agents Console
## You can search the dashboard with ⌘K and turn off post-call analysis
Press ⌘K to go to any dashboard page or agent. Workspace owners and admins can turn off post-call transcript analysis.
- The command palette finds every page, including pages that are not in the sidebar. It also finds agents by name and the tabs of the agent that is open.
- It knows common words. For example, calls opens Sessions and card opens Billing.
- The dashboard settings have a switch for post-call analysis. When you turn it off, Speko keeps transcripts and existing reports and stops new analysis.
Oct 27, 2026 Improved Available
Agents API
## Phone calls are safe to retry, and transfers send webhooks
POST /v1/sessions/phone accepts an Idempotency-Key, so a retry does not place a second call. Transfers send webhooks, and you can redeliver failed webhooks by time range.
- POST /v1/sessions/phone accepts an optional Idempotency-Key. A repeat with the same key in 24 hours returns the first successful response.
- agentVersion on POST /v1/sessions and POST /v1/sessions/phone runs a numbered agent version: live, archived or draft.
- GET /v1/sessions accepts metadata[key]=value filters, up to 10 keys. Pages no longer skip sessions that Speko created at the same time.
- Each transfer sends call.transfer.initiated, then call.transfer.completed or call.transfer.failed.
- POST /v1/webhook-deliveries/redeliver sends deliveries again for a time range. By default it selects failed deliveries, up to 50 for each request.
- Set screening: false on a warm transfer to add a person to the call with no screening. In the dashboard, select Add to call (agent stays quiet).
[Sessions API](https://docs.speko.ai/api-reference/sessions) [Webhooks](https://docs.speko.ai/guides/webhooks)
Oct 27, 2026 Improved Available
Router Gateway API
## The Router accepts TTS settings per provider and Soniox endpoint settings
You can send a TTS provider's own settings through the Router. Live Soniox sessions accept endpoint settings.
- options.provider_options on POST /v1/tts/speech and on the session.configure frame carries a provider's own settings, keyed by provider name. Examples are Cartesia normalization and ElevenLabs stability.
- The LiveKit and Pipecat TTS plugins accept provider_options. On LiveKit, update_options can change them for each reply.
- Live Soniox sessions accept max_endpoint_delay_ms, endpoint_sensitivity and endpoint_latency_adjustment_level in provider_options.soniox. Use them to end a turn sooner after speech stops.
- Soniox now treats language multi as automatic language detection. Before, Soniox refused it. On the Gateway, each final transcript carries the language that Soniox identified.
- The realtime and translation WebSocket routes no longer need an Idempotency-Key header. The Router makes one when you do not send it. The OpenAI SDK, LiveKit and Pipecat can connect with only a new host and key.
[TTS provider settings](https://docs.speko.ai/relay/tts/speech#provider-settings)
Oct 25, 2026 Added Available
Telephony API
## You can set a caller name, and Canadian numbers can send SMS
You can set the name that people see when your US number calls them. Canadian numbers can send SMS to Canadian recipients.
- Workspace owners and admins set callerName on PATCH /v1/phone-numbers/:id, or open Caller name on the number page. It applies to US local numbers that you bought through Speko.
- A caller name has 15 characters or fewer: letters, digits and spaces. Carriers can take up to about 72 hours to show it.
- Canadian numbers send SMS to Canadian recipients without a 10DLC campaign. The dashboard shows these numbers as Canada only.
[Phone numbers API](https://docs.speko.ai/api-reference/phone-numbers)
Oct 20, 2026 Added Available
Router Models
## The Router adds Hopper, Hamsa and Paxa TTS Flash v2
The Router added six models from October 15 to October 20.
- Hopper: gemma-4-31b for LLM, nemotron-asr for STT and qwen3-tts for TTS. All three serve English.
- Hamsa: s3 for Arabic STT and default for Arabic TTS.
- Paxa Labs: paxa-tts-flash-v2 for TTS, at the same price as v1. Automatic routing continues to select paxa-tts-flash-v1. Pin paxa:paxa-tts-flash-v2 to use v2.
- xAI: grok-voice-3.0 now has benchmark results on the Models page.
[Models](https://speko.ai/models)
Oct 19, 2026 Improved Available
Agents Telephony Billing
## Calls that fail on Speko's side, or where the agent never speaks, cost nothing
A call where the agent never speaks, or that fails because of Speko, is not charged, telephony included. The refund is automatic.
- Anything such a call was already charged is refunded to your credits when it closes. You do not need to open a ticket.
- Outbound calls bill from answer to the carrier's hang-up. The seconds the room stays open after the hang-up are not billed.
- Set ringTimeoutSeconds (20 to 45) on an agent or a call to ring for less time. Ringing is never billed.
[Outbound call billing](https://docs.speko.ai/guides/phone-agents#how-outbound-calls-are-billed)
Oct 15, 2026 Added Available
Router Models
## The Router adds Munsit, Fish Audio drama-3 and Grok Voice 3.0
The Router added five models from October 8 to October 15.
- Munsit: munsit and munsit-en-ar for Arabic STT, and faseeh-v1-preview for Arabic TTS.
- Fish Audio: drama-3 for TTS.
- xAI: grok-voice-3.0 on GET /v1/realtime. Agents can use it in browser and SDK sessions.
[Models](https://speko.ai/models)
## Earlier updates
22 older entries, listed by date. The [JSON](https://speko.ai/changelog.json) and [RSS](https://speko.ai/changelog.xml) feeds keep the full text of each entry.
Date Type Update
- Oct 15, 2026 Improved You can buy Canadian numbers and see why an outbound call did not connect
- Oct 14, 2026 Improved Speko starts billing an outbound call when the callee answers
- Oct 14, 2026 Added Agent monitoring groups failed calls into issues
- Oct 10, 2026 Added Your backend can send messages to an agent during a live call
- Oct 9, 2026 Improved Auto routing supports languages that have no benchmark ranking
- Oct 8, 2026 Added Test runs can use 12 new caller personas
- Oct 6, 2026 Added The Router adds 12 STT and TTS models
- Oct 5, 2026 Added You can connect a Twilio account and import its numbers
- Sep 30, 2026 Improved Tools can return mock responses in test runs
- Sep 28, 2026 Added The Router adds live speech translation
- Sep 25, 2026 Added The Router adds Gemini 3.8 TTS and four Nari models
- Sep 23, 2026 Improved Transcription jobs accept compressed audio
- Sep 21, 2026 Added The Router publishes an AsyncAPI contract for its WebSocket routes
- Sep 18, 2026 Added The Router serves speech-to-speech models on their native protocols
- Sep 16, 2026 Added Gateway adds a Pipecat LLM service for the Router
- Sep 16, 2026 Improved Agents usage on your own provider keys has no Speko fee
- Sep 16, 2026 Added Speech-to-speech agents answer and place phone calls
- Sep 7, 2026 Added Agent edits stay in a draft until you deploy
- Sep 5, 2026 Added The Router returns word timings for speech and transcripts
- Aug 25, 2026 Improved The Router applies the routing policy stored on the API key
- Aug 24, 2026 Improved The Router publishes its machine-readable contract
- Aug 24, 2026 Added The public Gateway runs on your own infrastructure