Why Gong, Observe.AI and CallMiner are still paying yesterday's STT prices
The three biggest conversation intelligence platforms run on STT pricing that hasn't moved with the market. A modeled estimate of what their bills look like at scale — and what changes for the next platform that doesn't carry their incumbency cost.
Conversation intelligence is one of the few SaaS categories where the incumbents and the prices of their inputs moved in opposite directions over the last 24 months. The category got more valuable, more competitive, more customer-funded. The STT layer underneath it got an order of magnitude cheaper. Almost none of that delta has shown up in the public pricing of Gong, Observe.AI or CallMiner — and the structural reason why is worth understanding for any team building in the same space.
The volumes, modeled
Public coverage and earnings color from CI's largest names suggests rough seat counts that, multiplied by reasonable per-seat usage, produce per-month minute volumes in recognizable ranges:
- Gong — frequently cited around 4,000+ customers with an average that puts aggregate platform throughput comfortably above 50M minutes a month at the low end of any reasonable model.
- Observe.AI — a contact-center-heavy customer base where seat counts run much higher per account; even conservatively modeled, throughput sits in the high-tens-of-millions of minutes per month.
- CallMiner — enterprise-anchored, with contract structures designed around large minute commits. Their internal volume is a multi-billion-minute annual number.
None of these volumes are exceptional in the modern STT market. They're operational. What's exceptional is how much each minute still costs once it lands inside the platform's infrastructure.
The bills, illustratively
Assume the modeled volume of 50M minutes/month for a CI-class platform. Run that volume through the major public STT options:
- AWS Transcribe (the legacy go-to for conversation analytics): $0.024/min × 50M = $1.2M / month.
- AssemblyAI Universal + intelligence: ~$0.0162/min × 50M = $810K / month.
- Deepgram Nova-3 + diarization: ~$0.0083/min × 50M = $415K / month.
- OpenAI Whisper API: $0.006/min × 50M = $300K / month.
Even an aggressive enterprise discount — assume 40% off list — leaves the AWS bill at $720K/month, and the Deepgram bill at $250K/month. Annualized: between $3M and $8.6M per year for the same transcripts a platform priced at infrastructure rates would produce for under $300K.
The incumbents in CI are paying 2021 STT prices for 2026 revenue. The structural reason is incumbency, not engineering.
Why none of them have moved
It's tempting to read CI-incumbent inertia as engineering slowness. It's not. Three structural reasons keep them locked in:
- Vendor commit contracts. Three-year minute commits with the existing STT vendor are signed when the platform raises a major round. Walking before the commit ends costs hard cash.
- Customer-facing SLAs. The contracts the CI platform signed with its enterprise customers reference uptime, redaction, residency and accuracy guarantees that ride on the incumbent STT vendor's certifications. A swap requires re-certifying that downstream.
- Internal tooling debt. Five years of observability, QA harnesses, custom-vocabulary management and prompt-biasing tooling wired to one vendor's API shape. Even when the destination API is wire-compatible, the soft layer around it isn't.
None of these are unsolvable. All of them are expensive compared to "do nothing." The longer the platform exists, the more expensive the swap gets — which is the entire reason next-generation challengers have a structural advantage on the same workload.
What this means for the next CI platform
If you're building a conversation intelligence platform in 2026, you don't carry any of the incumbents' baggage. No legacy commit. No five years of vendor-shaped internal tools. No customer contracts that lock you to a vendor's certifications. You can pick the STT layer that prices like infrastructure from day one and never pay the incumbency tax the category currently runs on.
At Orchard's per-minute rate — $0.00042, bundled diarization, no real-time premium, OpenAI Whisper-compatible SDK — the same 50M-minute workload costs $21,000 / month. That's not a competitive advantage you'd announce in a press release; it's a unit- economics floor that lets you out-price the incumbents on every deal without sacrificing margin.
If you're the customer
If you're a CI customer at one of the incumbents shopping for alternatives, the math at your own seat count is the only argument that matters. The platforms born after 2026 will price their seats against a $0.0004 floor, not a $0.024 one. It's only a matter of when that re-prices the entire category.