On May 7, 2026, OpenAI shipped GPT-Realtime-2, GPT-Realtime-Translate, and GPT-Realtime-Whisper. TechCrunch covered the launch as a product release. It is not. It is a pricing event that resets the unit economics of every AI voice agency on the planet, and most operators have not finished the math yet.
Five days in, the agency community is still digesting it. The pieces in the wild are developer-focused or generic SMB framing. No one has done the agency-margin breakdown. This is that breakdown.
The math, before the narrative
GPT-Realtime-2 lists at $32 per 1M tokens for audio input and $64 per 1M tokens for audio output. Cached audio input drops to $0.40 per 1M tokens. Translate is $0.034 per minute. Whisper-streaming transcription is $0.017 per minute. AIPricing.guru's pricing impact analysis has the full sheet.
Run one client through the math. A retainer that does 1,000 minutes of outbound AI voice per month on GPT-Realtime-2 at full input pricing burns roughly $180 in raw model inference. The same 1,000 minutes with a properly cached system prompt, product catalog, and compliance script burns closer to $20. That is not a 10% efficiency gain. That is the difference between an 85% margin business and a 30% margin business on the same revenue line.
"GPT-Realtime-2 cached audio input is listed at $0.40 per 1M tokens, a huge discount from $32 input. If your voice agent repeatedly uses the same instructions, product catalog context, compliance script, or workflow setup, caching strategy can materially change margins." AIPricing.guru, May 2026
If you are an operator and you cannot tell me your current cache hit rate this afternoon, your P&L has a hole in it that you cannot see yet. The hole gets bigger every minute someone routes traffic to you.
Three economic shifts that hit your agency in the next 60 days
Shift one. ElevenLabs and Deepgram get squeezed. OpenAI just ate the part of the stack that those two have been charging premium rates for. ElevenLabs' Pro pricing on conversational voice and Deepgram's premium streaming transcription tiers both sit above where the OpenAI Realtime line now starts. Expect a Q3 repricing on both, expect at least one of them to consolidate upmarket the way Deepgram already started doing with its $130M Series C and the OfOne QSR acquisition. When infra players go enterprise, the agency layer gets cheaper upstream costs but also fewer hands on the wheel for SMB-grade support.
Shift two. Cache strategy is the new moat. Six months ago, agencies competed on prompts and voices. Today they compete on what fraction of their token volume hits cache. An agency that runs 70% cache-hit rate beats an agency that runs 20% cache-hit rate by 3 to 4x on raw inference cost per call, every call, forever. This is now a platform-engineering problem, not a marketing problem. If your stack does not expose cache controls, your stack is leaking money.
Shift three. Wrapper agencies inherit a decision they did not make. If you build on Voicerr, Vapify, VoiceAIWrapper, Stammer, or Assistable, the cache strategy lives at the upstream layer. You pay whatever the upstream chooses, and you find out about price changes the same week your clients do. Trillet's documentation of Voicerr's 7 to 10x price hike from $28 per month to $199 to $299 per month earlier this year is the proof point everyone already paid for. The Realtime-2 launch makes the same dynamic ten times more important, because the cache decision is now where the margin lives.
Why margin opacity is now a balance-sheet problem
For two years, "I do not know my exact per-call cost" was an annoyance. As of May 7, it is a P&L hazard. If you bill a client $1,500 per month for 1,000 minutes and you cannot tell me whether your inference cost on that contract is $20 or $180, you are not running an agency. You are running a coin toss.
The agencies that win the next two quarters are the ones who treat every active client as a margin question, not a relationship question. Pull your last 30 days of call logs. Bucket them by client. Compute landed cost per minute per client. If you cannot do that in your current stack in under an hour, your stack is the problem. Not the model. Not the upstream. The stack.
What we are doing at Hermes about it
Hermes runs cache strategy, model selection, and per-minute optimization at the platform level. Agencies do not engineer cache hits. They do not pick models per agent. They do not chase the upstream pricing page every Friday. The landed cost is $0.18 per minute. The price to the agency's client is $0.24 per minute. The 25% spread is locked, regardless of how OpenAI prices the next two model generations.
That is the entire pitch. One platform. Your brand. Your margins. From $149 per month for Starter (300 included minutes), $399 for Business (1,000 minutes), $699 for Agency (2,000 minutes). Overage at $0.24 per minute. No upstream passthrough markups because we run the upstream relationship. By builders, for builders.
Action steps for agencies affected (this week, not Q3)
- Pull your inference cost per client for the last 30 days. If your platform cannot show this, that is the answer. Move.
- Identify your three largest token consumers in the system prompt. Compliance script, brand voice rules, product catalog. These are your cache candidates. If they change less than once a week, they should hit cache 100% of the time.
- Renegotiate your client contracts to per-minute pricing, not per-call. Per-call pricing breaks when cache hit rate changes. Per-minute with a fixed monthly minimum protects margin in both directions.
- Stop hiding cost line items from your own P&L. If you do not track ElevenLabs, Deepgram, OpenAI, Twilio, and CRM costs as four separate lines per client, you cannot tell which infrastructure shift just helped you and which one just hurt you.
- Pressure-test your platform on the cache question by Friday. Ask your provider, in writing, what their default cache hit rate is, who manages it, and what happens when OpenAI changes Realtime pricing again. If the answer is vague, you have your answer.
Frequently asked questions
How much does GPT-Realtime-2 actually cost per minute for an AI voice agency?
List price for GPT-Realtime-2 audio input is $32 per 1M tokens and audio output is $64 per 1M tokens. Cached audio input drops to $0.40 per 1M tokens, an 80x reduction. In practice, a well-tuned agent that caches its system prompt, product catalog, and compliance script can land somewhere between $0.06 and $0.12 per minute on raw inference. An agent with no cache strategy will land between $0.18 and $0.28 per minute. The gap is the entire margin of a typical agency contract.
Why do wrapper platforms like Voicerr, Vapify, and VoiceAIWrapper inherit this risk?
Wrappers do not control model selection, cache configuration, or per-minute optimization on the upstream provider. They pass through whatever Vapi or Retell charges, plus their own markup. When the upstream changes pricing or cache behavior, the wrapper has no negotiating position and the agency on top of the wrapper has even less. Voicerr's 7x to 10x price hike in early 2026 is the textbook example. The economics flow downhill.
Does Hermes pass GPT-Realtime-2 cache savings through to agencies?
Yes. Hermes operates at a $0.18 per minute landed cost and charges $0.24 per minute to agency clients, with the cache strategy, model selection, and routing managed at the platform level. Agencies do not have to engineer cache hits or tune cost. The 25% spread is locked. Starter is $149 per month with 300 included minutes, Business is $399 with 1,000 included, Agency is $699 with 2,000 included.
Where this leaves you
Pricing events like this happen twice a year now. Every time, the agency layer rediscovers that wrappers do not own the decisions that matter, and that platform-level cache strategy is the difference between running a real business and running a wholesale-margin business with extra steps.
The right move is not panic. The right move is to look at your last 30 days of margins, decide what you actually control, and move the rest to a stack that controls the parts you do not.
Further reading: OpenAI's own announcement, The Next Web's launch coverage, and Trillet's wrapper risk breakdown.
Originally published at buildwithhermes.com/blog/openai-realtime-2-voice-ai-agency-margins.
