Agent Guardrails, Webhook Testing, Auto-Failover, Billing Transparency & more.
Safety Guardrails
Your agent can now refuse what it shouldn't engage with.
→ Block jailbreaks — prompt extraction, instruction bypasses, unauthorized tool calls. Caught and killed before they reach the caller.
→ Filter harmful output — harassment, self-harm, violence, gambling, regulated advice, sexual exploitation, child safety, and more. You choose which categories are off-limits.
→ Hard stops, not soft warnings — when a guardrail fires, the agent doesn't hedge. It won't produce that content. Period.
Enable per-agent under Security & Fallback in agent settings. Available on all plans.
Global Node Return Paths
Customers can now leave a global node and come back to exactly where they were in the conversation.
→ No lost progress — a caller asks about your callback policy mid-booking, gets the answer, and picks up right where they left off. No restart, no repeated questions.
→ "Actually, never mind" just works — if a customer triggers a global node then changes their mind, the flow resumes naturally, like a real conversation would.
Enable per-node in any global node's settings.
Webhook & Functions Test Tools
Stop deploying integrations blind.
→ Webhook testing — Hit Test, get a real payload sent to your endpoint. Same structure, same data your agent sends in production. Catch auth failures, bad URLs, and schema mismatches before a single live call.
→ Custom function testing — See variable substitution, test connectivity, preview the response. No live call required.
Available for Agent Webhooks, Alert Webhooks, and Custom Functions.
Always Edges — Guaranteed Transitions
Force any conversation node to move forward, no matter what.
→ Overrides everything — an Always edge that overrides other transitions on that node. One path, guaranteed.
→ You control timing — "After User Responds" waits for input first. "Skip User Response" transitions immediately.
Add via the transition dropdown → Always/Skip, then set your transition mode. This replaces the old skip-response toggle with a more flexible model.
Other Platform Upgrades
Audio Transcription Auto-Failover — Deepgram goes down? We switch to Azure. Azure goes down? We switch to Deepgram. Mid-call. No audio dropped.
Batch Testing API — Define test cases, run them in bulk, get results back — all via API.
Branded Caller ID (Custom Telephony) — Show your company name on outbound calls and reduce spam flags. Now works with your own telephony numbers, not just Retell-managed. U.S. only. Requires business profile.
Advanced Call & Chat Filtering — Filter by custom analysis data and dynamic variables. Pin specific agent versions to pull from.
Billing Transparency Changes
TTS & Voice Engine — Now Separate Line Items
Your invoice previously bundled TTS and voice engine costs together. They're now broken out — you can see exactly what each costs across voice engines, LLMs, and telephony.
Concurrency — Billed Upfront, Prorated by Day
Concurrency is now charged when you add it, not at the end of the month. Add capacity mid-cycle and pay only for the remaining days.
Model Migrations
Seven deprecated models have been automatically replaced. If your agents used any of these, they're already running the new versions:
- claude-3.5-haiku → claude-3.5-haiku-v2
- claude-3.7-sonnet → claude-4.5-sonnet
- claude-4.0-sonnet → claude-4.5-sonnet
- gemini-2.0-flash → gemini-2.5-flash
- gemini-2.0-flash-lite → gemini-2.5-flash-lite
- gpt-4o-mini-realtime → gpt-realtime-mini
- gpt-4o-realtime → gpt-realtime
GPT-4.1 is not being deprecated on Retell. OpenAI recently removed it from ChatGPT. We've seen some of you asking about this — GPT-4.1 is staying on Retell. Your agents running it continue to work as-is.