Build or buy a phone agent in 2026: Twilio + Pipecat + Grok Voice or Eleven Realtime, versus Vapi, Bland and Synthflow

October 2026 made the build-your-own phone agent cheaper and more credible than it has ever been. Grok Voice 3.0 and ElevenLabs's new Eleven Realtime models both landed at $0.08 a minute for speech-to-speech. Pipecat shipped evals, classifiers and expressive conversation. Twilio has ConversationRelay with Deepgram Flux for turn-taking and an open-source Agent Connect bridge.

So the question UK operators keep asking us has a real answer now: should we assemble the stack ourselves, or pay a platform such as Vapi, Bland or Synthflow to do it? This is the buyer guide.

The two options, honestly

Build: a carrier (Twilio), a framework (Pipecat or Agent Connect), a speech-to-speech model (Grok Voice 3.0, Eleven Realtime, OpenAI's realtime line) or a separate STT, LLM and TTS trio, plus your own hosting, evals and monitoring.

Buy: a platform that bundles telephony, orchestration, model choice, a dashboard and, usually, compliance paperwork, billed per minute. We have reviewed Bland and Synthflow; Vapi and Retell sit in the same bracket.

What changed in 2026 that shifts the answer

• Speech-to-speech models at a flat per-minute price. Grok Voice 3.0 and Eleven Realtime (v1 Mini and Max, released 20 October, 90-plus languages, OpenAI SDK compatible) both list $0.08 a minute. The model layer is now a commodity you can swap; one review already found Grok and Gemini finishing more tasks in blind tests than Eleven Realtime, and another found Grok 3.0 a beat slower than 2.0. Test, do not assume
• Turn-taking got solved at the carrier. Twilio's ConversationRelay with Flux handles end-of-turn and backchannel filtering, which used to be the thing that made DIY agents feel robotic
• Frameworks grew up. Pipecat 1.x gives you evals, classifiers and a multi-agent bus. Agent Connect gives you a model-agnostic Twilio bridge out of the box
• Platforms raised money and prices. Vapi's $50M Series B in May signals a long runway; it also signals that per-minute platform pricing is where their margin lives

Cost per minute, roughly

Vendor-listed, before volume discounts, excluding your engineering time:

The build stack typically lands in the low-teens of pence per minute before people. Platforms are often in the same range or above once concurrency and premium voices are added, but they include the people. At low volumes buy wins. At high volumes, or with unusual requirements, build wins. The crossover depends entirely on whether you have the engineer.

Decision checklist

• Do we have an owner? One engineer who will still be here in a year and is comfortable with Python and WebSockets. No owner, no build
• Is the use case standard? Answer, qualify, book, escalate is a solved platform problem. Read Voice AI for the front door first
• Do we need a specific vendor combination? UK-accent STT (Speechmatics), a particular voice, a particular model for compliance reasons. Platforms limit this
• Do we need evals? Regulated or high-volume calls should be regression-tested. Pipecat has it built in; few platforms do
• What is our monthly minute volume? Under about 5,000 minutes a month, platform pricing is rarely the problem. Over 50,000, the per-minute margin is
• Can we exit? With build, you own the code and the numbers. With buy, check number porting and transcript export before signing
• Compliance paperwork. HIPAA eligibility and PCI scope on Twilio directly, or the platform's attestation. Either is fine; a gap is not

Three recommended patterns

Front-desk agent, SME, no engineer. Buy. Synthflow or Bland class, on your existing Twilio or the platform's numbers. Measure booked appointments, not call volume.

Contact centre pilot, 20,000-plus minutes a month, one engineer. Build on Twilio ConversationRelay with Flux, Pipecat for the pipeline, Grok Voice 3.0 or Eleven Realtime as the model, evals from day one. Keep a platform in the back pocket for overflow.

Product company adding voice to an app. Build with Pipecat on WebRTC or Media over QUIC, no carrier needed until you add phone, model chosen by your own blind tests.

Bottom line

In 2026 build is a genuine option, not a hobby. The model is $0.08 a minute from more than one vendor, the carrier solves turn-taking, and the framework ships evals. What has not changed is that build needs an owner and buy needs a budget. Pick on that, run a two-week bake-off on your real calls, and keep the exit route open whichever way you go.

Related reading

• Grok Voice 3.0 buyer note
• Pipecat in 2026
• Twilio in 2026: ConversationRelay and Agent Connect
• Bland AI review · Synthflow review
• Which AI voice sounds most human? Vapi's Humanness Index

Sources: xAI release notes · Eleven Realtime coverage (explainx) · Twilio SIGNAL 2026 announcements · Pipecat releases · Vapi GPT-Realtime-2 and Series B note

Erhan Timur, Founder, Digital by Default