OpenAI Realtime API Pricing: What a Voice Agent Really Costs

OpenAI Realtime API pricing is token-based, not per-minute: gpt-realtime-2.1 costs $32 per million audio input tokens and $64 per million audio output tokens. The mini model costs $10 and $20.

Those numbers are useless until you convert them to cost per conversation minute. This guide does the arithmetic step by step, using OpenAI's published rates.

We also cover the two things that move real bills most, prompt caching and context growth, and compare the result against chained stacks and per-minute platforms like Vapi and Retell AI.

OpenAI Realtime API pricing at a glance

OpenAI Realtime API pricing has five components per model: audio input, cached audio input, audio output, text input, and text output, all billed per million tokens. Here are the current published rates.

Two things jump out. Audio output costs twice as much as audio input on both models. And cached audio input is about 99% cheaper than fresh audio input, which is the biggest lever in the whole table.

The mini model is roughly a third of the flagship price on audio. For most 24/7 answering and routine support calls, it is the default starting point.

These rates change without much notice. Always confirm the current numbers on the official OpenAI pricing page before you budget a deployment.

How to turn OpenAI Realtime API pricing into cost per minute

To convert OpenAI Realtime API pricing into cost per minute, you need one conversion: a minute of input audio is roughly 600 tokens, and a minute of generated speech is roughly 1,200 tokens. Output audio is tokenized about twice as densely as input.

That gives clean base rates for gpt-realtime-2.1. Input: 600 tokens times $32 per million is about $0.019 per minute heard. Output: 1,200 tokens times $64 per million is about $0.077 per minute spoken.

Now a worked example: a 10-minute support call where the agent speaks for 4 minutes. Input: 10 minutes times 600 tokens is 6,000 tokens, which costs about $0.19. Output: 4 minutes times 1,200 tokens is 4,800 tokens, which costs about $0.31.

Base total: roughly $0.50 for the call, or about $0.05 per minute. On gpt-realtime-2.1-mini, the same call runs about $0.06 of input plus $0.10 of output, near $0.016 per minute.

Here is the catch: real bills run higher than the base math. Each conversation turn re-processes the growing context, so input tokens compound as the call gets longer. Without caching, long calls can cost several times the base estimate.

That compounding is why the cached audio input rate exists, and why the next section matters more than the headline prices.

• Rule of thumb: ~600 audio tokens per minute heard, ~1,200 per minute spoken.
• gpt-realtime-2.1 base rate: ~$0.019/min listening, ~$0.077/min speaking.
• Worked 10-minute call (4 min agent speech): ~$0.50 base, ~$0.05/min.
• Same call on the mini model: ~$0.16, near $0.016/min.
• Context re-processing pushes real uncached costs well above base math.

How caching and silence change your Realtime API cost

Prompt caching is the single biggest cost lever on the Realtime API: cached audio input bills at $0.40 per million tokens instead of $32, a discount of about 99%. Every turn re-sends your system prompt, tool definitions, and prior audio, and caching makes those repeats nearly free.

In practice, most of a long call's input tokens are repeats. With caching working, the compounding problem from the last section mostly disappears, and long calls trend back toward the base per-minute math.

Silence is the quieter budget leak. The session bills for audio tokens you stream, so an open microphone during hold time or dead air still generates input tokens.

Voice activity detection helps. Server-side turn detection decides when the caller is actually speaking, and sensible session handling avoids streaming long stretches of nothing.

Also watch session length itself. An agent that fails to end calls cleanly, or waits minutes for a caller who already hung up, pays real money to listen to a dial tone.

• Cached audio input: $0.40/1M vs $32/1M fresh, about 99% off.
• System prompts and tool definitions re-send every turn, so cache hits are constant.
• Streamed silence still generates billable input tokens.
• Tune voice activity detection and end sessions promptly.

OpenAI Realtime API pricing vs chained stacks and platforms

Compared with the alternatives, the Realtime API sits in the middle: pricier than a raw chained stack, cheaper than most managed per-minute platforms. Here is the landscape at published rates.

For the chained estimate: Deepgram lists streaming speech-to-text at $0.0048 per minute and Aura-2 text-to-speech at $0.030 per 1,000 characters, near $0.02 to $0.03 per spoken minute. AssemblyAI streaming lists at $0.15 per hour. A fast middle LLM adds fractions of a cent.

Platform rates bundle more than the model. Retell AI itemizes a $0.055 per-minute voice engine fee plus model, voice, and telephony lines, and quotes United States telephony via Twilio at $0.015 per minute. Bland AI charges one all-inclusive rate; its Build and Scale tiers add $299 and $499 monthly platform fees.

The honest conclusion: raw Realtime API access is not what makes voice agents expensive. Orchestration, telephony, and platform margin are. If your volume is high, that margin is where negotiation pays off.

For help choosing the model layer itself, see our guide to the best LLM for voice agents.

Seven ways to cut your Realtime API bill

You can usually cut Realtime API spend by half or more with seven changes: use the mini model, maximize cache hits, shorten instructions, trim tool outputs, control silence, cap session length, and monitor per-call cost.

Start with the model. gpt-realtime-2.1-mini is about a third of the flagship audio price. Route routine calls to mini and escalate only complex ones.

Then protect your cache. Keep the system prompt and tool definitions stable across turns, because any change invalidates the cached prefix and puts you back at $32 per million on input.

Trim what the model reads. Long tool responses, like a raw CRM record dumped as JSON, become input tokens every following turn. Return only the fields the agent needs.

Finally, instrument everything. Log the token usage returned with each response, compute dollars per call, and alert on outliers. One looping agent left talking to voicemail can quietly burn a day of budget.

• Default to gpt-realtime-2.1-mini; escalate to the flagship selectively.
• Keep prompts and tool definitions byte-stable to preserve cache hits.
• Return lean tool outputs; every extra field re-bills each turn.
• Cap max session length and hang up on dead air.
• Track cost per call from the usage data, not the monthly invoice.

Bottom line on OpenAI Realtime API pricing

OpenAI Realtime API pricing works out to roughly $0.05 per conversation minute on gpt-realtime-2.1 base math, and near $0.016 on the mini model, before context growth and after sensible caching. That is competitive with any option on the market.

The published rates are $32 and $64 per million audio tokens for the flagship, $10 and $20 for mini, with cached audio input at $0.40 and $0.30. Verify them on the OpenAI pricing page before budgeting.

Your real bill is decided by engineering, not rates: cache hygiene, silence handling, and lean tool outputs. Get those right, and the Realtime API is the strongest value in voice AI today.