#Sometime in the last 2 weeks - without prompt or tool changes - our agent has gone from returning a

1 messages · Page 1 of 1 (latest)

tacit lilyBOT
#
AI-generated Answer for mikegreen7585

Short answer: There’s no documented ElevenLabs change that replaces ASCII spaces with U+202F in chat responses. In Chat Mode, your app receives the model’s raw text via the agent_response event; ElevenLabs does not document any server-side whitespace normalization for chat. (elevenlabs.io)

What to check next (grounded in docs):

  • Verify which LLM your agent is configured to use. LLM selection is configurable in Agents, and output formatting can differ by model; if that setting changed, text style can change. (elevenlabs.io)
  • Normalize before rendering: handle agent_response and replace any U+202F (narrow no‑break space) with a regular space in your UI layer; this is supported by the documented pattern of intercepting agent_response and updating your chat UI. (elevenlabs.io)
  • If you also feed that text into Text-to-Speech, note that “text normalization” is a TTS-only feature and is configurable (apply_text_normalization on/off/auto). It doesn’t apply to Chat Mode text delivery. (help.elevenlabs.io)

If you want, share which LLM the agent is set to and how you’re rendering agent_response; I can point to the exact place to normalize it per the docs. (elevenlabs.io)

Sources:
Chat Mode | ElevenLabs Documentation
Large Language Models (LLMs) | ElevenLabs Documentation
Chat Mode | ElevenLabs Documentation
Why are numbers, dates, symbols and acronyms not properly pronounced or spoken in the correct language? – ElevenLabs
Chat Mode | ElevenLabs Documentation

pale zodiac
#

happening with the gpt-oss-20b model. which I notice is no longer available...
switched to gpt-oss-120b - same thing.
appears glm-4.5-air works now.

but begs the question - should we expect model changes underneath us like this?!

tacit lilyBOT
#
AI-generated Answer for mikegreen7585

Short answer: yes—if you select ElevenLabs’ experimental LLMs or upstream preview models, availability and behavior can change.

What the docs say

  • ElevenLabs-hosted LLMs labeled “experimental” (including GPT-OSS-20B and GPT-OSS-120B) explicitly note that “pricing and availability may change as these models are being evaluated and optimized.” In other words, they can change or disappear without notice. (elevenlabs.io)
  • Upstream model deprecations can also trigger changes. For example, on July 14, 2025, ElevenLabs announced that Google Gemini preview models were deprecated and that “all agents using these models will automatically be transferred to gemini-2.5-flash the next time they are used.” (elevenlabs.io)

How to avoid surprises

  • Prefer non-experimental, generally available models from the “Supported LLMs” list in Agents, or bring your own provider key to control the exact model/version yourself. (elevenlabs.io)
  • If you need strict control, use “Custom LLM” (OpenAI-compatible) so your agent points at your own endpoint or your own OpenAI/Azure OpenAI key, rather than ElevenLabs’ internal credentials. (elevenlabs.io)

Notes on “GPT-4.5 Air”

  • I can’t find “GPT‑4.5 Air” referenced in ElevenLabs’ public docs or model lists, so I can’t confirm its availability within Agents. The current documented OpenAI options include GPT‑4.1/4o/4o‑mini, etc. (elevenlabs.io)

If you want, I can point you to the exact steps to pin a non-experimental model or wire up a Custom LLM so model behavior doesn’t change underneath you. (elevenlabs.io)

Sources:
Large Language Models (LLMs) | ElevenLabs Documentation
July 14, 2025 | ElevenLabs Documentation
Large Language Models (LLMs) | ElevenLabs Documentation
Integrate your own model | ElevenLabs Documentation
Large Language Models (LLMs) | ElevenLabs Documentation
Integrate your own model | ElevenLabs Documentation

pale zodiac
#

glm-4.5-air - i meant