[PaulAsjes]

Summary

The CLI still pointed people and agents at older models. say now defaults to eleven_v4, and agents created from the CLI's templates use eleven_v4_turbo. The skills the CLI generates for text-to-speech and text-to-dialogue now tell agents to pass --model-id eleven_v4, so they no longer end up on the API's older defaults.

Details

• say: DEFAULT_MODEL_ID is now eleven_v4. onboard's test request uses the same constant, so it moves too. The say skill and the README now cover v4 and v3 for audio tags, and suggest eleven_v4_turbo as the faster option, since it still reads tags. I dropped the old "~0.8s slower than flash" claim because there's no matching v4 figure.
• Agent templates: tts.model_id changes from eleven_flash_v2 to eleven_v4_turbo, which matches the API's own default for agents (TTSConversationalModel) in the bundled spec. The CLI never used eleven_v3_conversational anywhere.
• Plain text-to-speech and text-to-dialogue commands send no model_id unless one is passed, so the API picks the model (eleven_multilingual_v2 and eleven_v3). That default comes from the API, so the CLI doesn't hard-code a model. Instead, a new MODEL_HINTS list in skills.rs adds a short "Model" section to those two generated skills. Like the existing additions there, generate-skills fails with an error if a hinted skill disappears after a regeneration, rather than quietly dropping the note.
• Not changed: generated SDK/types code, the bundled spec, speech-to-speech and speech-to-text (separate model families with no v4), and the version number.

Testing

• cargo test: all suites pass (1921 + 839 + smaller suites, 0 failures), including a new model_hints_recommend_v4 unit test.
• Ran elevenlabs generate-skills into a scratch directory: 34 files written, and eleven_v4 appears only in say, text-to-speech and text-to-dialogue.
• text-to-speech convert --dry-run without --model-id sends only text, which confirms the API picks the default.
• Not verified: an actual say call or agent creation against the live API with the new models.

🤖 Generated with Claude Code