Models
- Claude is the primary model in my customer-facing chat assistants. The page includes the $50 billing limit that took every tenant offline in August 2026.
- Groq serves open models faster than anything else I use. It ran my phone agent, and then it retired the model I built on.
- OpenAI is third in line in my stack. Its request format is in half my code. Its models bill me almost nothing.
Voice and telephony
- Twilio carried the calls. The page covers Media Streams wiring, numbers that keep billing after a customer leaves, and a texting registration that never cleared.
- Deepgram did live speech-to-text. Includes the Safari bug that made a voice demo deaf.
- Cartesia was the voice on the phone line and in my pitch videos, at about 1.8 cents per call minute.
- Kokoro TTS is a free model on my Mac that now voices my narration. Five failure modes and the checks that catch them.
Data
- Supabase holds logins and data for six of my apps. The page covers free-tier pausing and a login hang.
Comparisons
- ChatGPT vs Claude for business. API list prices now match on three of four tiers, so price no longer settles it. Here is what does.
- AI voice agent platforms compared. Published per-minute prices for seven platforms next to the 3.6 cents my self-built agent cost.
How one call uses five of these
The pages make more sense as a chain. On my AI receptionist, a single inbound call went through this path:
| Step | Platform | Share of a call minute |
|---|---|---|
| Phone number and live audio stream | Twilio | $0.0129 |
| Caller's speech to text | Deepgram | $0.0048 |
| Deciding what to say | A small open model on Groq | about $0.0006 |
| Text to voice | Cartesia | about $0.018 |
| Accounts and call records | A database on my own server | $0 marginal |
That is about 3.6 cents a minute at November 14, 2026 list prices. The chat products swap the first four rows for a single Claude call and keep the last.
The rules for a page
- My code has to call it. I searched my own repositories and servers for each platform before writing. A job posting that mentions a tool does not count, and neither does an account I opened once.
- Prices come from the vendor's own page. Each page links the source and gives the date I read it. If I could not read a number, I left it out and said so.
- Failures stay in. The outage, the registration that never cleared, the key I pasted wrong. Those are the useful parts.
- Clients stay anonymous. I describe the business. I do not name it.
Platforms without a page, and why
n8n, Make and Zapier. I have not built a client system on any of them. My one contact with Make is a landing-page form I wired to post into a client's existing Make scenario. I write code on a server I control, for the reasons on the custom vs no-code page. A three-way comparison from me would be a rewrite of other people's reviews.
ElevenLabs. I designed one game announcer voice in its app by hand. No code of mine calls its API. Its published agent price appears on the voice comparison page, labeled as not run.
Google Gemini. It has been a fallback model in three of my apps and never a primary. That is too thin for a page.
When one of these moves into a system I run, it gets a page with dates on it.