Skip to main content
A rusty split-flap departure board with blank tiles mid-flip on a foggy railway platform at dusk under amber lanterns, a metaphor for an ElevenLabs agent's deprecated LLM being swapped for a replacement model
8 min readBy Carlos Aragon

ElevenLabs Agent LLM Deprecated? I Measured the Fallback

When the LLM behind an ElevenLabs agent is deprecated, the agent doesn't break. ElevenLabs quietly sends a growing share of new conversations to a replacement model, reaching 100% a few days before the provider shuts the old one down, and your agent settings keep showing the old model the whole time. Right now that covers GPT-5, GPT-5 mini, GPT-5 nano, Gemini 2.5 Flash and Claude Sonnet 4.5. I ran 75 test conversations this morning to see how many calls already moved, what it does to cost, and where you can actually see it.

Which ElevenLabs Agent LLMs Are Deprecated Right Now?

The source of truth isn't the changelog, it's the List LLMs endpoint (GET /v1/convai/llm/list). It returned 115 models for my workspace, and 63 of them carry is_deprecated: true (most are dated snapshots like gpt-4o-2024-08-06). Here are the ones that matter for live agents as of November 22, 2026:

Agent LLMReplaced byProvider cutoffFallback rampMoved today
gemini-3.6-flashgemini-3.8-flashNov 19, 2026Oct 20 → Nov 17100%
claude-sonnet-4-5claude-sonnet-4-6Nov 30, 2026Oct 31 → Nov 2589%
gemini-2.5-flash / -litegemini-3.5-flash / -liteNov 30, 2026Nov 15 (25%) → Nov 3060%
gpt-5 / gpt-5-mini / gpt-5-nanogpt-6.1-sol / gpt-6.1-terra / gpt-6-lunaDec 11, 2026Nov 11 → Dec 649%
gpt-4o, gpt-4o-mini, gpt-4.1, gpt-4.1-minigpt-6-lunaJan 14, 2027Dec 15 → Jan 90% (warning)
gemini-3.7-flashgemini-3.8-flashJan 28, 2027not published0% (warning)

If your agent runs GPT-5 mini, about half of today's new calls are already answered by a different model. Two details surprised me. First, the GPT-5 replacements in the live API are gpt-6.1-sol, gpt-6.1-terra and gpt-6-luna, while the November 2 changelog still says traffic moves to the GPT-5.6 family. My test calls ran on gpt-6.1-terra, so the API is right and the changelog is stale. Second, the changelog says the Gemini 2.5 ramp starts November 8, but the API config and the models page both say November 15. Read the API, not the release notes.

How Does the Fallback Percentage Work?

Every deprecated model ships a deprecation_info block. This is GPT-5 mini today:

{
  "llm": "gpt-5-mini",
  "is_deprecated": true,
  "is_in_warning_period": true,
  "is_in_fallback_period": true,
  "fallback_percentage": 49,
  "provider_deprecation_date": "2026-12-11T00:00:00Z",
  "replacement_model": "gpt-6.1-terra",
  "deprecation_config": {
    "warning_start_days": 90,
    "fallback_start_days": 30,
    "fallback_complete_days": 5,
    "fallback_start_percentage": 10,
    "fallback_complete_percentage": 100
  }
}

The percentage is a straight line. It starts at fallback_start_percentage on (cutoff minus fallback_start_days) and hits fallback_complete_percentage on (cutoff minus fallback_complete_days). For GPT-5 mini that's 10% on November 11 climbing to 100% on December 6. November 22 is 11 days into a 25-day window: 10 + 90 x 11/25 = 49.6, and the API says 49. The same math gives 89 for Sonnet 4.5 and 60 for Gemini 2.5 Flash (which uses a 15-day window starting at 25%), matching the API both times.

So you can work out the exact day your agent passes 50% or 100% without waiting for an email. The warning_start_days: 90field is the 90-day heads-up before the cutoff. That's where GPT-4o and GPT-4.1 sit now, at 0% until December 15.

What Happened When I Called Deprecated Agents

I created six throwaway agents through the API, each pinned to a deprecated model, set to text-only with temperature 0, and opened conversations over the agent WebSocket with a one-word question. Then I pulled each conversation and read which model it was actually billed for.

Agent set toAPI saidCalls movedLLM price per turn (old → new)Time to first sentence (median)
claude-sonnet-4-589%16 / 20 → Sonnet 4.6$0.00114 → $0.00116 (same)0.62 s → 0.57 s
gemini-2.5-flash60%9 / 20 → Gemini 3.5 Flash$0.00037 → $0.00054 (1.4x)1.07 s → 0.64 s
gpt-5-mini49%9 / 26 → GPT-6.1 Terra$0.00011 → $0.00040 (3.5x)0.65 s → 0.95 s
gemini-3.6-flash100%3 / 3 → Gemini 3.8 Flashn/an/a
gpt-4o-2024-05-13 (shut down Oct 23)gone3 / 3 → GPT-5.6 Soln/an/a
gemini-1.5-flash (long retired)gone3 / 3 → Gemini 3.5 Flash-Liten/an/a

The share that moved was a bit under the advertised number

80% vs 89%, 45% vs 60%, 35% vs 49%. With 20 to 26 calls per model that's inside normal noise, but all three came in low, so I wouldn't plan around the exact figure. Plan around the trend: by the complete date it's 100%.

The model is picked once per conversation

I ran six four-turn conversations on the GPT-5 mini agent. Four stayed on GPT-5 mini for all four turns, two ran on GPT-6.1 Terra for all four. In 24 turns the model never changed mid-call. For voice that's the good news: a caller won't hear the agent change personality halfway through. The bad news is that two callers on the same day can get two different models.

Cost and speed can go either way

Sonnet 4.6 cost the same as Sonnet 4.5. Gemini 3.5 Flash cost 1.4x Gemini 2.5 Flash per turn but answered faster. GPT-6.1 Terra was the expensive one: 3.5x the LLM price of GPT-5 mini per turn, and about 0.3 seconds slower to its first sentence. On the free tier that showed up as 34 credits per conversation instead of 31. These are tiny one-word turns on small samples, so treat the latency as a direction, not a benchmark. A long system prompt and real tool calls will widen the cost gap, not shrink it.

Nothing tells you

The agent config still read gpt-5-miniafter the test. The agent object has no deprecation field. The conversation's warnings array was empty on all 75 calls. The only place the swap shows up is the billing block: metadata.charging.llm_usage, where the model key is gpt-6.1-terra instead of the model you picked. The same key is on every agent turn in the transcript.

You can still create agents on dead models

POST /v1/convai/agents/create accepted gemini-1.5-flash, gpt-4o-2024-05-13 (which OpenAI shut down on October 23) and claude-sonnet-4-5 (already removed from the dashboard selector) with a 200 and no warning. Every call on them ran on the replacement. If you clone agents from old JSON templates, you can end up on a model that hasn't existed for months and never notice.

Why You Shouldn't Just Let the Fallback Happen

The fallback is a safety net and a decent one. But a voice agent prompt is tuned to one model's habits: how long its answers run, when it calls a tool, how it handles a caller who goes off script. A newer model in the same family can be smarter and still break the thing you tuned for. I measured that kind of drift in the Sonnet 4.5 migration on the Claude API and in Eleven Realtime Mini vs Max.

During the ramp you also get two agents in production at once, and your analytics mix them together. If conversion or average handle time moves this month, you won't know if it was your prompt change or the model swap unless you split by the billed model.

One more catch: the default backup LLM sequence is Gemini 3.5 Flash, GPT-4o, Gemini 3.5 Flash-Lite, Claude Sonnet 4.6. GPT-4o is itself in the warning period for January 14, 2027. Your backup chain is aging too.

How to Fix It in 30 Minutes

1. Audit every agent

This script lists every agent in the workspace and flags the ones on a deprecated model, with the replacement, today's percentage and the cutoff. I ran it against my own workspace before publishing (both of my agents came back clean: one on gemini-3.5-flash, one on a custom LLM).

import os, requests

API = "https://api.elevenlabs.io/v1"
H = {"xi-api-key": os.environ["ELEVENLABS_API_KEY"]}

llms = requests.get(f"{API}/convai/llm/list", headers=H, timeout=30).json()["llms"]
dep = {m["llm"]: m["deprecation_info"] for m in llms
       if (m.get("deprecation_info") or {}).get("is_deprecated")}

cursor = None
while True:
    params = {"page_size": 100, **({"cursor": cursor} if cursor else {})}
    page = requests.get(f"{API}/convai/agents", headers=H, params=params, timeout=30).json()
    for a in page["agents"]:
        cfg = requests.get(f"{API}/convai/agents/{a['agent_id']}", headers=H, timeout=30).json()
        llm = cfg["conversation_config"]["agent"]["prompt"]["llm"]
        d = dep.get(llm)
        if d:
            print(f"{a['name']}: {llm} -> {d['replacement_model']} "
                  f"({d['fallback_percentage']}% of conversations now, "
                  f"cutoff {d['provider_deprecation_date'] or 'already gone'})")
        else:
            print(f"{a['name']}: {llm} ok")
    if not page.get("has_more"):
        break
    cursor = page["next_cursor"]

It only checks the agent's main LLM. If you use workflows, check the LLM overrides on subagent and update-state nodes too, since the November 2 release lets those set their own model.

2. Test the replacement, then switch on purpose

Take replacement_model as a starting point, not an order. Run your real test calls (or the edge cases that usually break voice agents) on it and on one cheaper option, compare answers and price, then set the agent's LLM yourself. For a GPT-5 mini agent, GPT-6.1 Terra was 3.5x the LLM price in my test, so check whether a smaller model holds up first.

3. Set your own backup order

Set backup_llm_config to override with an orderof current models, so a provider outage doesn't route calls to GPT-4o. Details are on the ElevenLabs models page.

4. Log the model that actually answered

import os, sys, requests

cid = sys.argv[1]
conv = requests.get(f"https://api.elevenlabs.io/v1/convai/conversations/{cid}",
                    headers={"xi-api-key": os.environ["ELEVENLABS_API_KEY"]}, timeout=30).json()
usage = conv["metadata"]["charging"]["llm_usage"]["irreversible_generation"]["model_usage"]
print(cid, "ran on", ", ".join(usage) or "no hosted LLM (custom LLM?)")

Put that in your post-call webhook handler and store the model next to the outcome. It's the only way to compare before and after a switch, deprecation or not.

5. Export conversations before you delete an agent

When I deleted my test agents, their conversations went with them: GET /v1/convai/conversations/{id} returned 404 document_not_found right after. If you retire an old agent instead of editing it, pull its history first.

How I Tested

November 22, 2026, from a Mac mini in Allen, TX, on a free ElevenLabs workspace. Six throwaway agents created with POST /v1/convai/agents/create, text-only, temperature 0, English, deleted afterwards. Conversations ran over wss://api.elevenlabs.io/v1/convai/conversation with Python websockets: 20 single-turn calls each on Sonnet 4.5, Gemini 2.5 Flash and GPT-5 mini, six four-turn calls on GPT-5 mini, and three calls each on Gemini 3.6 Flash, GPT-4o (2024-05-13) and Gemini 1.5 Flash. That's 75 conversations and about 3,300 credits. Model, price and timing come from GET /v1/convai/conversations/{id} (llm_usage and convai_llm_service_ttf_sentence). Text-only calls skip speech, so real phone calls add ASR and TTS time on top. If you're still choosing a platform, my ElevenLabs Agents vs Retell AI cost comparison covers the rest of the bill.

Running Voice Agents in Production?

I build and maintain ElevenLabs and Retell voice agents for real businesses, including the boring part: model migrations, test calls and knowing which model answered which call. If your agents are on GPT-5 or Gemini 2.5 and you want them moved cleanly before December, let's talk.

Related Posts