OpenAI Realtime 2.1 Costs: Audio, Text and Cache Are Different Meters

Three modalities, one session

For gpt-realtime-2.1, OpenAI’s published list prices differentiate text, audio and image input. Standard text is $4 per million input tokens and $24 output; audio is $32 input and $64 output; image input is $5. The rate card also lists cached-input tariffs for each eligible modality. A single session can touch several of these meters, so a single token count is insufficient to reconcile spend.

A talking assistant may receive streaming microphone audio, transcribe or reason about text, produce speech and occasionally inspect an image. Record modalities separately, not as one abstract “API call.” Session transport, such as WebSocket or WebRTC, does not make these model tariffs equal to a phone-minute charge.

Why duration cannot replace usage measurements

Two calls of the same wall-clock duration may have different token output and different turn structures. Silence, barge-in interruptions, repeated instructions and long responses can change the amount of metered work. It would be unsound to convert speech minutes into audio tokens with a universal, unstated constant. The price CSV stores token rates as quoted; duration belongs in operational measurement, not as a fabricated rate conversion.

Do not confuse the separately listed per-minute GPT-Live session product with gpt-realtime-2.1’s modality-based token billing. They are distinct line items and should not be combined without the actual product and metering details.

An illustrative mixed-modality estimate

Suppose an observed usage record contained 200,000 uncached audio-input tokens, 70,000 audio-output tokens, 20,000 ordinary text-input tokens and 12,000 text-output tokens. At the listed rates, those components contribute $6.40, $4.48, $0.08 and $0.288, or $11.248 before any cached reads, image tokens, tools or local tax. This is a made-up workload for explaining arithmetic, not a claim about expected speech length.

The exact computation uses usage counters returned for the session. If the provider reports reasoning tokens as output, those belong in the output meter; do not charge them a second time after incorporating reported totals.

Cache hit economics and stateful dialogue

Cache reads may be far cheaper than uncached input, but only when the relevant prefix genuinely qualifies and is hit. A changing instruction set, different voice configuration or accumulating conversation history can reduce reuse. Maintain distinct counters for uncached and cached audio/text inputs. Logging only the latest prompt fragment can also miss earlier context that the model received again.

For a production voice agent, store enough telemetry to trace a costly session while respecting privacy and retention requirements. Avoid retaining raw customer voice merely to prove a billing number when aggregate usage metadata would suffice.

Build a cost guardrail before launch

Start with a maximum expected number of concurrent sessions, estimated active duration, possible tools and an output budget. Set separate alerts for unusually high generated-audio volume, long multi-turn loops and tools called repeatedly. Test cost with diverse conversations only when the organisation has actual consent and a suitable evaluation set; this article reports no hands-on audio experiments.

Comparison with a transcription-plus-text-to-speech pipeline should include latency and error handling, but keep its transcription rate units separate from Realtime token metering. Pay attention to account tiers and regional availability as well.

Version and quotation caveats

The product reference names gpt-realtime-2.1 specifically. Other Realtime versions or the mini offering have different tariffs. Model release dates and the original effective date of the mutable price page are not verified here. The numbers are a 25 October 2026 reference, denominated in USD, not a cross-region guarantee. If a provider adds session charges or a tool fee, quote those only from their own documented schedule.