ElevenLabs adds a fast LLM tier for agents, clone strength for v4 voices, and expiring API keys
ElevenLabs shipped several developer-facing changes on October 26: a faster, pricier LLM service tier for ElevenAgents, a new control over how closely v4 Professional Voice Clones follow their training samples, website crawl exclusions for agent knowledge bases, and expiring service account API keys.
Details
- Fast LLM service tier: ElevenAgents can run supported LLMs on a
fasttier that trades a per-token price premium for higher throughput and lower latency. Setservice_tiertofastin the agent prompt configuration via Update agent.available_service_tiersfrom List LLMs shows which tiers a model supports. The default remainsstandard, and the LLM usage calculator acceptsservice_tierto estimate the cost difference. The premium amount is not stated in the changelog. - Clone strength for v4 voices: Text to Speech and Text to Dialogue accept
clone_strength(0 to 1) for a Professional Voice Clone fine-tuned for a v4 model, set invoice_settingsorsettings. The recommended value for most use cases is 0. On Text to Dialogue, it now decides whether a v4 PVC runs as a PVC or an IVC, anduse_pvc_as_ivcis deprecated. - Expiring service account API keys: Pass
expires_at_unixwhen creating or updating a key, or choose Expire after in the dashboard. Expiry must be between 15 minutes and 90 days ahead; omit it for a key that never expires. - Knowledge base crawl exclusions: Up to 50
exclude_patternscan be passed to Create crawl job; matching URLs are skipped even if they match the crawl pattern. - Other changes: Slack, Telegram, Custom Channel and Intercom integrations leave alpha; Slack agents can keep replying in a thread;
claude-sonnet-5-5is selectable as an analysis LLM; Voice Changer moves to/v1/voice-changer/{voice_id}with the old speech-to-speech paths deprecated; Realtime Speech to Text accepts silence thresholds down to about 0.19 seconds; legacy/v1/dubbingendpoints are deprecated; Business plans on current pricing pay about 17% fewer credits per image and per video second.
What happened next
The fast tier follows a pattern seen across LLM providers: latency is sold as a priced option, which matters for voice agents where response delay is felt directly. Expiring keys and the cleaner PVC/IVC control are practical hardening and quality tools for production deployments. Several deprecations (Voice Changer paths, legacy dubbing endpoints, use_pvc_as_ivc) mean API users should check their integrations; removal dates are undisclosed.