The Routing Decision Changes On A Render, Not A Launch

Every model on the routing table is scored 1 to 5 on 7 axes and the axes are fidelity, consistency, control, cost, speed, availability and uncensored, and the 7 axes are the only thing the eye reads. Fidelity is realism of output and the source of truth is a Grok-vision or Gemini gate plus the operator eye. Consistency is identity and continuity across shots and the source of truth is the H3 consistency test plus the Sal y Miel identity lock. Control is image-conditioning and framing and style lock and seed and the source of truth is the MiniMax and Wiro probe results. Cost is euros per output or per token. Speed is render and iteration time. Availability is on-plan and balance and key status from IDENTITY.md. Uncensored is NSFW capability and the source of truth is the Wiro face-guard tests. The weighting is per task type and the task weighting is what flips the winner. Video is fidelity 30, consistency 25, control 20, cost 10, speed 10, uncensored 5. Coding is cost 30, speed 20, control 15, fidelity 15, consistency 10, availability 10. Speech is fidelity 50, cost 25, speed 25. The same ranked list picks a different model per task type and the picked model is what the cron writes into the table. The 7 axes plus the 3 weightings is the whole engine. A vendor launch that arrives on a Tuesday morning has to beat the incumbent on the axes that matter for the task or it stays off the table. The probability that a launch beats the incumbent on every axis that matters is not zero, and most people are pricing that probability at zero.

The video routing list is 8 rows and the 8 rows cover every render path the studio runs. Uncensored body and dance stills plus video at 4 to 30 seconds is Wiro seedance-2-5-uncensored and there is no fallback because the explicit passes plus the face-guard rejecting readable faces only is the gate. The still render identity anchor is Wiro seedream-v5-pro-uncensored at 2K 16:9 and the fallback is Gemini 2.5-flash-image. Face plus lip-sync is OpenArt Vibe Direct and there is no fallback because Vibe is the only renderer that holds face identity and syncs audio. The 10 second image-conditioned cinematic shot on-plan no extra euro is MiniMax Hailuo-2.3 I2V with first_frame_image at 768P and the fallback is Wiro seedance, and the verification was 10 seconds at 1364 by 768 at 24 fps on the coding-plan key. The 15 second subject-to-video named character is MiniMax S2V-01 and the fallback is none because the 15 second was accepted and needs a clear face subject. Dialogue beat with synced speech is Fast-H3 T2AV at about 0.175 dollars a run and the fallback is MiniMax H3 on PAYG and the output is 15 seconds native at 14.375 seconds with scene lock and no character lock. Long high-res at 15 seconds plus at 2K is MiniMax H3 with PAYG credits needed and the fallback is Wiro seedance 30 second and the H3 is plan-gated with the gate labelled 2013 on the coding key. Cheap fully-automated pipeline is Seedance API at about 10 euros an output and there is no fallback because no LoRAs and the output is flat and the grade is in post. The do-not-use list is Grok imagine plus Wan 2.7 plus Wan 3.0 and the verdict is Grok is retired and Wan is ethnicity drift plus super AI look. A routing list that has 8 rows and no fallback on 4 of them is a routing list that names the truth. The probability that the fallback shows up is real and the operator picks the fallback when the primary is down and the picked fallback is what the table writes down.

The coding routing list is 3 rows and the 3 rows cover the 1M context bulk and the fast iteration and the hard reasoning. Bulk agent and automation coding at 1M context is MiniMax M3 on the coding plan and the fallback is MiniMax M2.7 and the verdict is Anthropic-compatible and cheapest at volume plus 1M window. Fast iteration plus high-speed is MiniMax M2.7-highspeed and the fallback is DeepSeek v4-flash and the verdict is speed over depth. Hard reasoning plus architecture is DeepSeek v4-pro and the fallback is Claude and the verdict is text-only no vision plus pair with Gemini for images. The dead row is Nous Hermes-4-70B plus 405B and the verdict is 401 invalid blocked or out of funds. A dead row that stays in the table is a routing bug and the routing bug is what a cron run finds and removes at 6 am. The coding plan carries about 5.1B M3 tokens per month and the plan renews on 2026-09-28 and the included quota is what makes M3 the volume winner. A routing list that names the dead row out loud is a routing list the operator can trust. The probability that a retired model sneaks back into a cron run is not zero and the cron run is what catches the sneak.

The speech routing list is 3 rows and the rows cover cinematic narration plus bulk cheap TTS plus the female voice roster. Cinematic narration in the Anna voice is ElevenLabs eleven_flash_v2_5 with the voice Elise and there is no fallback because the warm unhurried read is the locked Anna voice. Bulk plus cheap TTS is MiniMax speech-* T2A on plan and the fallback is none because the plan shares the coding-plan quota and the marginal cost is zero. Other female voices is ElevenLabs Cassidy plus Maria plus Matilda and the verdict is the office voice roster. The image routing list is 2 rows and the rows cover uncensored identity stills plus brand social safe imagery. Uncensored identity stills is Wiro seedream-v5-pro-uncensored and the fallback is Gemini 2.5-flash-image and the verdict is reference-guided plus 100 percent face-accuracy lock. Brand social safe imagery is Gemini 2.5-flash-image and the fallback is MiniMax image-01 and the verdict is fast clean on key. The do-not-use list is Grok imagine and the verdict is retired. The music routing list is 2 rows and the rows cover song generation plus lyrics-first co-write. Song generation is Suno via HeartMuLa and the fallback is none. Lyrics-first co-write is MiniMax M3 draft then Suno render and the fallback is none. The vision gating routing list is 2 rows and the rows cover render A/B gating plus image understanding caption. Render A/B gating is Grok vision and the fallback is Gemini 2.5-pro. Image understanding caption is Gemini 2.5-pro and the fallback is Grok vision. A locked pick on narration is a locked pick on the Anna voice and the locked voice is what holds the channel brand across 100 episodes.

A vendor announcement is not a routing signal. A render is.

The pricing snapshot is 8 rows and the rows carry the verified cost plus the live status. MiniMax M3 plus M2.7 on the coding plan is about 5.1B M3 tokens per month included and the plan renews 2026-09-28 and the status is live. MiniMax Hailuo-2.3 video is included on plan at 10 seconds max at 768P and the status is live. MiniMax H3 plus S2V-01 is PAYG credits and H3 is plan-gated and the status is H3 blocked on plan. Wiro seedream plus seedance is pay-as-you-go and the balance was 6.52 dollars and 1 concurrent and the status is low balance. Fast-H3 is about 0.175 dollars a run and the status is live. Seedance API is about 10 euros an output and the status is live. ElevenLabs is per-char eleven_flash tier and the status is live. Suno plus HeartMuLa is per-song and the status is live. Nous Hermes-4 is 401 out of funds and the status is dead. Prices refresh every morning via the Model Watch cron and stale price is flagged and never trusted. A price that is not refreshed is a price that the cron run fails the routing table on. The Wiro balance check is what catches the low balance at 6 am and the low balance is the signal to top up or route to a different model. A stale price that lands in the table is a routing bug. The probability that a vendor changes the price without an email is not zero and most people are pricing that probability at zero and the cron is the only thing that catches the silent change.

The routing table is rebuilt every morning by the Model Watch cron and the rebuild scores every model on the 7 axes and the rebuild writes the new row plus the new fallback plus the new verdict and the rebuild checks the live balance and the rebuild checks the live API status. A routing table that does not rebuild is a routing table that drifted the day after the launch. The change log holds the dated delta and the dated delta is the audit trail. The 2026-08-31 row is the system created and routing derived from all vault render and API tests. Every morning after that is a new row with a new delta and the new delta is what the operator reads at 7 am. The operating rule is one sentence and the sentence holds the table. Score every model on 7 axes, weight per task type, write the run plus the fallback plus the verdict, refresh prices every morning, retire on a render that fails the gate, retire on a launch that does not beat the gate. The held table is what makes 8 video paths plus 3 coding paths plus 3 speech paths read as one studio.