Real AI. Real models. Named.

Some "AI" products have turned out to be people behind a script. Others use real AI but won't tell you which one, so you can't tell when the model is wrong, drifted, or deprecated. We name every model that touches your data.

Every model ID below is imported at build time from the same catalog our production worker runs (apps/worker/src/lib/modelIds.js), so this page cannot drift from what actually serves your account. The text models are the exact IDs our router (llm.js) calls; a drift test verifies the image, video, and voice IDs against the running code on every build.

Text generation

Reviews, GBP posts, ad copy, blog posts, brand voice, lead scoring, voice-call grading, and the Maya assistant.

Capability

Model

Vendor

Review responses, GBP posts, lead scoring, ad-copy quick variants, Maya voice mode (low-latency)

Claude Sonnet 5.5

claude-sonnet-5-5

Anthropic

Long-form blog posts, landing-page copy, ad campaign briefs, brand strategy, multi-step reasoning, Maya chat

Claude Sonnet 5.5

claude-sonnet-5-5

Anthropic

Highest-effort copy: reel / story-ad scripts (the five-beat ad arc)

Claude Opus 5.5

claude-opus-5-5

Anthropic

Image generation

Ad creatives, landing-page imagery, GBP photos, brand asset generation. Called direct from each vendor, no reseller in the middle.

Capability

Model

Vendor

Photorealistic product and lifestyle imagery, brand-styled hero shots (primary)

Nano Banana (Gemini 3.1 Flash Image)

gemini-3.1-flash-image-preview

Google

High-fidelity fallback when the primary is busy or declines

Imagen 4

imagen-4

Google

Alternate variants, strong text-in-image

GPT Image 1

gpt-image-1

OpenAI

Additional fallback tier

Imagen 3

imagen-3

Google

Video generation

Ad video creative, talking-avatar reels with native audio, UGC-style testimonials. Called direct from Google, no reseller.

Capability

Model

Vendor

Cinematic ad video + talking-avatar reels (native audio) from prompt + reference images

Veo 3.1

veo-3.1-generate-preview

Google

Fallbacks, in order, when the primary is busy

Veo 3.1 Lite, then Veo 3.1 (Vertex)

veo-3.1-lite-generate-preview / veo-3.1-generate-001 / veo-3.1-lite-generate-001

Google

Voice

The phone receptionist, the Maya assistant voice, tutorial narration, and reel voiceovers.

Capability

Model

Vendor

Maya assistant voice + tutorial narration (low-latency TTS)

Cloudflare Aura-2

@cf/deepgram/aura-2-en

Cloudflare Workers AI

AI phone receptionist: hears the caller and speaks the reply (speech to speech)

GPT Realtime

gpt-realtime-2025-08-28

OpenAI

Receptionist fallback voice when the realtime engine is unavailable

Amazon Polly

Polly (Twilio <Say>)

Amazon, through Twilio

Reel / short-form video voiceover

Gemini 3.1 Flash TTS

gemini-3.1-flash-tts-preview

Google

Reel voiceover fallback

GPT-4o mini TTS

gpt-4o-mini-tts

OpenAI

Spoken-response generation behind the voice (what Maya says)

Claude Sonnet 5.5

claude-sonnet-5-5

Anthropic

Quality + safety gates

Written AI output is graded before it ships; below the threshold it waits for a person. Live phone replies are not pre-graded: calls are graded afterwards.

Capability

Model

Vendor

Output quality scoring on written AI output (1-5 scale)

Claude Sonnet 5.5

claude-sonnet-5-5

Anthropic

Brand-voice fit scoring (does the output sound like the customer)

Claude Sonnet 5.5 + brand_dna anchor

claude-sonnet-5-5

Anthropic

Three more commitments around AI

If a human writes it, we say so.

You'll see a "Drafted by [name]" tag on any output that wasn't AI-generated. We don't pass off contractor work as AI, or AI work as a senior strategist's, ever.

Per-account cost ledger.

Every model call we make on your behalf is logged with token count, model name, and dollar cost. You can see this in your Plan and Usage panel inside the platform. Daily spend caps prevent runaway costs from a buggy agent.

Quality gate on every output.

Before written AI text goes to your customers (review responses, Google posts, ad copy), we score it 1-5 against your brand voice. Live phone replies cannot be scored before they are spoken, so calls are graded afterwards. Below threshold = human review required, never auto-shipped.

Last updated 2026-11-13.