If you launched a Deepgram Voice Agent pilot this summer, your September invoice is about to jump without a single deploy. I've been helping three teams audit their agent-minute math this week, and the pattern is the same: they picked Standard or Advanced for the bundled stack, assumed Flux TTS would stay subsidized, and forgot that Deepgram's promo pricing ends September 12, 2026. Standard PAYG goes from $0.056/min to $0.075/min on September 13, a 33.9% reset on the tier most pilots run.
You have roughly twelve days to re-tier, negotiate, or unbundle STT before billing catches up. In this guide I'll walk through the cliff, the dollar impact at 10k and 50k agent-minutes, three migration paths, and how to swap only the STT layer with Pipecat and Privocio while keeping your LLM and TTS.
The Sept 12–13 promo cliff: what changes and why
Through September 12, Deepgram's Voice Agent API bundles Flux TTS free inside Standard, Custom BYO-LLM, and Advanced tiers. You're billed at the BYO-TTS equivalent rates. On September 13, those tiers revert to full list price, and Flux TTS moves to $0.045 per 1,000 characters as a standalone product.
Two details bite pilots hardest. First, Voice Agent bills on connection time, not just spoken audio. Silence, hold music, and barge-in windows all count toward agent-minutes. Second, teams on Standard or Advanced with bundled TTS take a double hit: higher per-minute agent rates plus TTS no longer subsidized in-stack.
I verified the post-cliff numbers against Deepgram's pricing page and the breakdown at DIY AI's 2026 pricing guide. The cliff is real, and it lands mid-month for anyone who started a September pilot on promo rates.
Cost table — 10,000 and 50,000 agent-minutes (PAYG)
Here's the math I send to finance before any architecture meeting. BYO tiers avoid the cliff on Deepgram's line item, but you still pay external LLM and TTS costs.
| Tier | Rate through 9/12 | Rate from 9/13 | 10k min (promo → post) | 50k min (promo → post) | % increase |
|---|---|---|---|---|---|
| Standard | $0.056/min | $0.075/min | $560 → $750 (+$190) | $2,800 → $3,750 (+$950) | +33.9% |
| Standard BYO TTS | $0.065/min | $0.065/min | $650 (flat) | $3,250 (flat) | 0% |
| Custom BYO LLM | $0.050/min | $0.065/min | $500 → $650 (+$150) | $2,500 → $3,250 (+$750) | +30.0% |
| Custom BYO LLM + TTS | $0.050/min | $0.050/min | $500 (flat) | $2,500 (flat) | 0% |
| Advanced | $0.122/min | $0.163/min | $1,220 → $1,630 (+$410) | $6,100 → $8,150 (+$2,050) | +33.6% |
| Advanced BYO TTS | $0.122/min | $0.122/min | $1,220 (flat) | $6,100 (flat) | 0% |
At 50,000 agent-minutes on Standard, you're looking at $950 more per month with zero code changes. That's why I tell teams to project September usage this week, not after the first invoice lands.
Three migration paths before September 13
Path 1 is re-tiering to BYO on Deepgram. If you already run an external LLM and TTS, switch to Custom BYO LLM + TTS at $0.050/min flat. You keep Deepgram's orchestration and dodge the Standard/Advanced cliff. Custom BYO LLM alone still jumps 30% on September 13, so read the tier row carefully.
Path 2 is unbundling STT entirely. Drop the Voice Agent API and run LiveKit Agents or Pipecat with separate STT, LLM, and TTS providers. More engineering work upfront, but you control each line item. For a component-level cost model, see Cogniq AI's 2026 voice agent breakdown.
Path 3 is negotiating Growth prepaid credits. Deepgram's Growth plan ($4k+/year prepaid) drops Standard to $0.068/min post-cliff. That's still +33% from the promo rate of roughly $0.051/min equivalent, but it beats raw PAYG if you want to stay fully bundled and can commit spend.
For a broader STT comparison outside the Voice Agent bundle, our Privocio vs Deepgram guide covers streaming STT on its own.
Unbundling STT with Pipecat + Privocio
Full platform swaps aren't always necessary. I've migrated two pilots by keeping their LLM and TTS and swapping only transcription.
Pipecat treats STT as a pluggable service. Point it at Privocio's OpenAI-compatible endpoint (https://api.privocio.com/v1) instead of Deepgram STT. Our speech-to-text for AI agents guide walks through the wiring, and the voice pipeline architecture post shows where Agent output mode trims LLM tokens by 35–40% versus raw transcripts.
Healthcare and fintech pilots often chose Deepgram for speed, not data residency. Self-hosted Privocio keeps audio inside your VPC, which matters when HIPAA or GDPR review boards ask where recordings live.
At 400 hours per month, Privocio's Go plan runs $19 per four weeks on our pricing page. Deepgram Nova-3 streaming STT alone often clears $115+ at similar volume before you add LLM and TTS. Unbundled stacks I've benchmarked land around $0.04–$0.08/min total versus $0.075–$0.163/min on bundled Voice Agent Standard or Advanced post-cliff.
Trade-off: Privocio streaming runs around 500ms partial latency in my tests. That's slightly behind Deepgram's fastest streaming, but fixed pricing and no training on your audio matter more once you're past ~5k agent-minutes per month.
from pipecat.services.openai import OpenAILLMService
from pipecat.services.openai.stt import OpenAISTTService
stt = OpenAISTTService(
api_key=os.environ["PRIVOCIO_API_KEY"],
base_url="https://api.privocio.com/v1",
model="privocio-agent", # token-optimized output for downstream LLM
)
Test with the free tier (3 hours audio per four weeks) before you cut over production traffic. Full API details sit in our docs.
Decision checklist: stay bundled vs unbundle
Run through this before September 13:
- Volume: Below ~5k agent-minutes per month, re-tiering on Deepgram is usually faster than rebuilding orchestration. Above 10k, unbundled STT often wins on total stack cost.
- Engineering capacity: Pipecat or LiveKit Agents need someone who can own WebSocket lifecycle and provider failover. No spare cycles? Path 1 or 3.
- Compliance: If your pilot handles PHI or EU personal data, verify where audio is processed. Bundled Deepgram may be fine contractually; self-hosted STT removes a shared-infra question.
- Current tier: Already on Custom BYO LLM + TTS? You mostly need to confirm Flux TTS billing separately after September 12. On Standard with bundled TTS? You're in the cliff zone.
Frequently asked questions
Does the BYO TTS tier avoid the September 13 increase?
Yes for Standard BYO TTS and Advanced BYO TTS — those rates stay flat. Custom BYO LLM (without external TTS) still rises from $0.050/min to $0.065/min on September 13.
Is Flux TTS still free after September 12?
No. Flux TTS bills at $0.045 per 1,000 characters standalone and no longer subsidizes bundled Standard or Advanced tiers.
Can I keep Deepgram orchestration but swap only STT?
Yes. Custom BYO LLM + TTS keeps Deepgram's agent orchestration with your own LLM and TTS at $0.050/min flat. For full STT control, unbundle via Pipecat and point STT at Privocio while leaving your TTS provider unchanged.
How does Privocio compare on latency?
In my streaming tests, partial transcripts arrive in roughly 500ms. Deepgram's fastest paths can beat that by a few hundred milliseconds. For most agents, LLM inference is the real bottleneck, not STT.
What should I do this week?
Pull your August agent-minute logs, multiply by the post-cliff rate for your tier, and compare against BYO and unbundled scenarios. If the delta exceeds a sprint of engineering time, start the migration now.
Conclusion: audit before September 13
The Deepgram Voice Agent promo cliff is a billing boundary, not a code deploy. Standard and Advanced tiers jump roughly 34% on September 13, and bundled Flux TTS stops being free. If you're past ~10k agent-minutes on Standard, unbundling STT with Pipecat and Privocio often lands a cheaper, more predictable stack.
Run the September projection this week. If unbundling fits, start on our free tier and read the full AI agent voice pipeline guide. For STT-only comparisons, see Privocio vs Deepgram and our pricing page.