Primary interaction · agent orchestration · code refinement. Fable 5 (Mythos-class) is the session default since July — survived a 18-day US export-control freeze; Opus 4.8 stays as the flat-rate fallback.
A visual atlas of every ONETEN automation pipeline. See at a glance which LLMs and tools cooperate inside each workflow.
Each pedestal is one LLM or tool inside the ONETEN stack — a real 3D exhibit you can spin. Beneath it: the pipelines it currently powers.
Primary interaction · agent orchestration · code refinement. Fable 5 (Mythos-class) is the session default since July — survived a 18-day US export-control freeze; Opus 4.8 stays as the flat-rate fallback.
OpenAI's coding agent — and the CC storyboard engine's realtime image generator: view_image anchors + built-in image_gen produce face-locked cinematic shots per beat, zero API fees on subscription.
Google's agentic coding IDE (latest 2.0). Another parallel-agent surface in the dev workflow alongside Claude Code and Codex.
Heavy-lift outsourcing: long-context analysis, dialogue writing, boilerplate, refactor planning. Cheap.
OpenAI's new three-tier family (Sol = frontier reasoning · Terra = 2x cheaper than 5.5 · Luna = fast/cheap). Legacy GPT-5.5 handled Haifan v1 storyboards; 5.6 tiers are the new option pool for ORACLE votes and dialogue rotation.
Full-duplex voice model — listens while it speaks, real-time interpretation, delegates complex asks to GPT-5.x. API pending; earmarked as the future voice layer of the ONE TEN living room.
Local resident: OCR, translation, polish, default LLM for the dali bot. Free batch runs.
TTS — 4-character bilingual voice synthesis (speech-02-hd). Since July routed through the Fal relay (same model + voice ids, pay-per-use pennies) after MiniMax raised its direct recharge minimum to $25. All calls pass tts_normalize first.
Lip sync + image gen (gpt_image_2 / Seedream / Flux). Lip sync for character portraits, gpt_image_2 for hero infographics, Seedream for ref-based avatar pool.
HTML → MP4 render layer: composes storyboards + dialogue + voice into 1080p output.
Frontier-tier reasoning at half-price. Drop-in upgrade for fast LLM calls when DeepSeek is too slow.
Multimodal video gen — text/image/audio/video in, physics-aware video out. Used for PA1 motion graphic hero scenes (SPY chart, McKinsey infographic).
Aggregator: 9-model rotation for MCV3 dialogue + the ORACLE voting panel (GLM-5.2 / Grok-4.3 / Gemini-3.5 / GPT-5.4-mini and friends). Model ids validated against /v1/models before every run.
One card per pipeline. Click a card to expand the flow + description.