01
The Realtime API left beta
Generally available for production voice agents for the first time. That is a procurement and support fact rather than a capability one, and for most teams shipping voice it was the actual blocker.
★ FLAGSHIP · REALTIME API, NOW GAOpenAIModel brief
OPENAI · SPEECH TO SPEECH · REALTIME APIGPT-Realtime-2.1. A voice agent that can think mid-sentence.
The Realtime API left beta with this generation, which is the part that matters commercially: OpenAI now supports it for production voice agents. Where the previous realtime models were built to sound convincing, 2.1 carries GPT-5-class reasoning and function calling inside the audio loop, so the agent can hold a conversation and do the work at the same time. The cost model is unusual enough to design around before you build.
Realtime APIgpt-realtime-2.1Function callingVoice agents
Released
May 2026
Model ID
gpt-realtime-2.1
Audio in
$32 / 1M tokens
Audio out
$64 / 1M tokens
Context
128K tokens
Max output
32K tokens
Modalities
Text, audio, image → text, audio
Knowledge
Sep 2024 cutoff
[TL;DR FOR CEO + CTO]
What to know.
- 01
The Realtime API is generally available, which is the release.
It shipped as a beta and stayed there through several model generations. Production voice agents were a support-contract conversation rather than a documented path. That changed with this generation, and for anyone who has been holding a voice product behind a beta label, the GA line matters more than the model card.
- 02
Reasoning is now inside the audio loop.
GPT-Realtime-1.5 was built to sound like a person. 2.1 is built with GPT-5-class reasoning and supports function calling and configurable effort, so it can handle a request that needs a lookup, a decision, and a follow-up question without handing off to a separate text model and back.
- 03
Audio tokens cost eight times text tokens.
Audio input runs $32 per million against $4 for text, and audio output $64 against $24. Audio is also token-dense, so a conversation of a few minutes is not a cheap request. Design the transcript-versus-audio split deliberately rather than sending everything as audio because the API allows it.
- 04
Cached input at $0.40 is the lever that makes it viable.
Cached input is $0.40 per million for both text and audio, which is 1.25% of the uncached audio rate. A voice agent with a large stable system prompt, a persona, and a tool catalogue re-reads that prefix on every turn. If you do not wire caching properly here, you are paying eighty times more than you need to on the largest part of the bill.
- 05
The knowledge cutoff is September 2024, and that is older than it looks.
The text flagships sit on April to June 2026 cutoffs. A realtime model nearly two years behind will state stale facts confidently in a voice a caller trusts more than a chat window. Ground anything time-sensitive through tools rather than through the model's memory.
[SOFTWARE DEVELOPMENT IMPACT]
What it changes for the team building with it.
What a team actually has to build around. The comparisons that matter are against the previous realtime generation, and against the alternative architecture of stitching a transcription model, a text model, and a synthesis model together yourself.
| Dimension | vs GPT-Realtime-1.5 | vs a stitched STT + LLM + TTS pipeline |
|---|---|---|
What the agent can do | 1.5 was a premium audio model without reasoning in the loop, so anything that needed a decision meant handing off. 2.1 carries GPT-5-class reasoning and function calling, so a lookup, a decision, and a clarifying question happen inside one session. | A stitched pipeline gives you a stronger text model and free choice of transcription vendor, at the cost of latency at every hop and the loss of prosody, interruption, and overlap. Route by whether the conversation is the product or the transcript is. |
Latency you can ship | Speech to speech in one session removes the transcribe, think, synthesise round trip entirely. This is the reason to use a realtime model at all, and no stitched pipeline gets close on turn latency. | Three network hops and three model warmups per turn is the ceiling on a stitched design. Deepgram Flux Multilingual narrows the gap on the front half by putting end-of-turn detection in the model, at under 300ms median. |
Cost control | Same $32 / $64 audio pricing as gpt-realtime-2. The mini tier at $10 / $20 is the new lever, and it is the right default for scripted or narrow flows that never needed deep reasoning. | A stitched pipeline is usually cheaper on paper: GPT-Transcribe at $0.0045 per minute plus a text model plus TTS. It is also three failure surfaces and three vendors. We price both before choosing, per flow rather than per product. |
Operational surface | The API is GA, so this is now a supported production path rather than a beta you explain to your CTO. Rate limits still start low: 200 RPM and 40,000 TPM at tier 1, which is a real constraint on a launch. | One vendor and one session against three integrations. Against that, a stitched pipeline lets you swap the weakest component without touching the others, which is worth a lot while this category moves monthly. |
We build voice the same way we build everything else: the transport and the model sit behind one interface, so a flow can move from a stitched pipeline to a realtime session, or from the full tier to mini, as a config change with an eval behind it.
[WHAT IS NEW]
The features that ship with it.
02
GPT-5-class reasoning in the audio loop
Where gpt-realtime-1.5 was a premium audio model, 2.1 handles more complicated requests because it reasons before it answers, with configurable effort. Reasoning tokens bill as output at the audio or text rate depending on the modality.
03
Function calling during a live conversation
The agent can look something up, book something, or check a record while the caller is still on the line, without breaking the session or dropping to a text model and back. This is what separates a voice agent from a voice demo.
04
A mini tier at roughly a third of the price
gpt-realtime-2.1-mini runs $10 audio in and $20 audio out, against $32 and $64. Text is $0.60 and $2.40 against $4 and $24. Most production voice traffic is narrower than it looks, and a lot of it belongs here.
05
Image input alongside audio
The session accepts text, audio, and image input. A support agent that can look at the photo a customer just sent while talking to them is a different product from one that cannot.
06
Prompt caching on the realtime session
Cached input at $0.40 per million on both audio and text. On a voice agent, the system prompt, persona, and tool catalogue are re-read every turn, so this is the difference between a viable unit cost and an alarming one.
[WHAT CHANGES IN YOUR CODE]
The API surface, exactly.
The realtime surface is not the Chat Completions surface, and most of the integration pain comes from assuming it is. These are the constraints worth knowing before the first spike.
Watch out
One endpoint, and it is not the one you already use
gpt-realtime-2.1 serves v1/realtime and nothing else. Chat Completions, Responses, Assistants, Batch, and Fine-tuning all reject it. A realtime session is a persistent bidirectional connection rather than a request and response, so your transport, your error handling, and your retry logic are all different code from the rest of your integration.
v1/realtimeWatch out
Audio tokens are billed on a separate scale from text
Audio input is $32 per million against $4 for text, and audio output $64 against $24. Any cost model that counts tokens without splitting them by modality will be wrong by roughly an order of magnitude. Meter audio and text separately from day one, because retrofitting that split after launch is painful.
audio: $32 in / $64 out · text: $4 in / $24 outWatch out
A September 2024 knowledge cutoff
Almost two years behind the current text flagships. In a voice interface a confident wrong answer carries more weight than it does in a chat window, because callers do not scroll back and check. Ground anything time-sensitive with a tool call rather than trusting model memory.
knowledge cutoff: 2024-09-30New
Cached input at $0.40 per million
The same rate for cached audio and cached text, against $32 uncached audio. Structure the session so the stable prefix (system prompt, persona, tool definitions) is cacheable and the variable part comes after it. This is the single highest-leverage thing you can do to the bill.
cached input: $0.40 / 1MNew
Function calling with configurable effort
Tools work inside the live session, and effort is tunable, so you can keep routine turns cheap and raise effort only for the turn that needs a decision. Treat effort as a per-turn setting rather than a session-wide constant.
function_calling · prompt_caching · reasoning effortNew
Rate limits start low
Tier 1 allows 200 RPM, 1,000 RPD, and 40,000 TPM. For a concurrent voice product those numbers arrive faster than they look, because a session holds capacity for its whole duration rather than for the length of a request. Check your tier before a load test, not after.
tier 1: 200 RPM · 1,000 RPD · 40,000 TPM[THE SPEC]
Everything an integration depends on.
The limits, the endpoints, and the lines that decide whether this model fits your transport and your budget. Kept here so nobody has to reconstruct it from three vendor pages.
| Model ID | gpt-realtime-2.1 |
|---|---|
| Compact tier | gpt-realtime-2.1-mini, roughly a third of the audio price |
| Context window | 128,000 tokens |
| Max output | 32,000 tokens |
| Knowledge cutoff | 30 September 2024. Older than the text flagships, so ground anything time-sensitive with tools |
| Input | Text, audio, and image |
| Output | Text and audio |
| Endpoint | v1/realtime only. No Chat Completions, no Responses, no Batch |
| Features | Function calling, prompt caching, reasoning tokens with configurable effort |
| Audio pricing | $32 / MTok in, $0.40 / MTok cached in, $64 / MTok out |
| Text pricing | $4 / MTok in, $0.40 / MTok cached in, $24 / MTok out |
| Rate limits, tier 1 | 200 RPM, 1,000 RPD, 40,000 TPM |
| Predecessors | gpt-realtime-2 (same pricing), gpt-realtime-1.5 (prior generation, no reasoning) |
[WHICH MODEL FOR WHICH JOB]
Routing, not picking a favourite.
Voice is several different products wearing one word. These are the splits we actually route on.
GPT-Realtime-2.1
gpt-realtime-2.1
- Inbound support that has to resolve, not just deflect
- Booking, scheduling, and account changes over the phone
- Any flow where the caller can interrupt and change direction
- Conversations that need a tool call mid-sentence
$32 / $64 audio. Worth it when the conversation itself is the product and the agent has to decide things. Wire caching before you launch.
GPT-Realtime-2.1 Mini
gpt-realtime-2.1-mini
- Scripted qualification and intake
- Appointment reminders and confirmations
- Narrow flows with a small tool surface
$10 / $20 audio. Our default until an eval shows the flow genuinely needs the reasoning of the full tier.
GPT-Transcribe
gpt-transcribe
- Meeting, call, and media transcription after the fact
- Anything feeding search, summarisation, or extraction
- High-volume archives where cost per hour dominates
$0.0045 per minute, about $0.27 an hour. If nobody is waiting on the words, this is the right model and a realtime session is waste.
GPT-Realtime-Translate
gpt-realtime-translate
- Live cross-language conversation
- Multilingual support desks
- Events and briefings with mixed-language audiences
$0.034 per minute, more than 70 input languages and 13 output languages. Billed by the minute, so the unit economics are legible before you ship.
[PRICING]
What it costs to run.
Realtime billing splits by modality, which is the part that catches teams out. These are list rates per million tokens.
| Audio input | $32.00 | Eight times the text input rate. Audio is also token-dense. |
|---|---|---|
| Audio output | $64.00 | The most expensive line on the bill for a talkative agent. |
| Cached input | $0.40 | Same rate for audio and text. 1.25% of uncached audio input. |
| Text input | $4.00 | Send text as text wherever the flow allows it. |
| Text output | $24.00 | Reasoning tokens bill here when the turn produces text. |
| Mini tier, audio | $10 / $20 | gpt-realtime-2.1-mini, in and out. Text at $0.60 / $2.40. |
The arithmetic that decides a voice product: a caller speaking for a minute is not a small number of tokens, and both sides of the conversation bill. Before committing, model a representative call end to end at these rates, split by modality, with and without caching. We have seen the cached-versus-uncached difference change a per-call cost by more than the choice of model did, which is why we treat prompt structure as a cost decision rather than a prompt-engineering one.
[RISK + GOVERNANCE]
What to watch.
Voice carries more authority than text, and the model knows less.
A September 2024 cutoff behind a confident synthetic voice is a specific risk shape. Callers do not scroll back, do not see citations, and generally cannot tell a grounded answer from a remembered one. Anything time-sensitive, priced, or contractual should come from a tool call, and the prompt should say so explicitly.
Disclosure is not optional in most of the places you will deploy this.
Several jurisdictions require callers to be told they are speaking with an automated system, and consent rules for recording vary by state and country. This is a product requirement rather than a legal footnote: it changes the opening line of the call, the consent flow, and what you are allowed to retain.
Plan the handoff before you plan the happy path.
The failure mode that damages a brand is an agent that will not let a frustrated caller reach a person. Build the escalation path, the trigger conditions, and the context handover first, then build the automation. It is also the cheapest insurance available against a bad launch week.
[OUR TAKE]
How we are using it.
01
General availability is the news, not the benchmark.
We have had voice work sitting behind a beta API label for a while, and that label is a real blocker for regulated customers regardless of how good the model is. A supported production path changes what we can responsibly recommend.
02
Most voice flows do not need the flagship.
Once you actually map a call flow, a large share of the turns are intake, confirmation, and routing. Those belong on the mini tier at a third of the price. We route within a single call rather than picking one model for the whole product.
03
Caching is the difference between a demo and a business.
At $32 per million for uncached audio input and $0.40 cached, prompt structure is a unit-economics decision. We design the stable prefix first, measure the cache-hit ratio in staging, and treat a low ratio as a bug rather than a tuning opportunity.
04
We still stitch pipelines, and we say so.
When the transcript is the product rather than the conversation, GPT-Transcribe at $0.0045 a minute plus a strong text model beats a realtime session on both cost and quality. Realtime is a latency tool. Reaching for it when nobody is waiting is an expensive habit.
05
Nobody wins every axis, so do not buy one vendor.
OpenAI leads the agent surface, ElevenLabs currently leads word error rate at 2.2%, and Deepgram leads end-of-turn detection at under 300ms median. We build behind an abstraction and route by what the flow is actually sensitive to, because this ranking has changed twice in a year.
[METHODOLOGY · K-FRAMEWORK]Integrated through the
Integrated through the
K-Framework.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-Framework01
Foundations
Direct API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
02
Amplification
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
03
Judgment
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
[OBSERVABILITY]
Observability your team can read.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Instrumented
Cost per call
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Instrumented
Latency p50 / p95 / p99
Real distributions, not averages. We know which routes are slow, and why.
Instrumented
Eval pass rates
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
Instrumented
Prompt + completion logs
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
[COMMON QUESTIONS]
Questions we are getting asked.
- Should we use a realtime model or stitch transcription, an LLM, and TTS together?
- Ask whether anyone is waiting. If a human is on the line and can interrupt, use a realtime session: no stitched pipeline matches it on turn latency, prosody, or overlap handling. If the output is a transcript, a summary, or a record, stitch it, because GPT-Transcribe at $0.0045 per minute plus a strong text model is cheaper and gives you a better text model.
- Why is audio so much more expensive than text?
- Audio input is $32 per million tokens against $4 for text, and audio output $64 against $24. Audio is also token-dense per second of speech, so the two effects compound. The practical answer is to send text as text wherever the flow allows, cache aggressively, and route the routine turns to the mini tier.
- How much does a typical call cost?
- It depends entirely on call length, how much of the conversation is audio in both directions, and your cache-hit ratio, so any single number would be marketing. What we do is model a representative call end to end at list rates, split by modality, with and without caching, before quoting anything. The cached-versus-uncached gap is usually larger than the gap between models.
- Is the September 2024 knowledge cutoff a problem?
- It is if you rely on model memory. It is not if you ground the answers. Prices, availability, policies, and anything dated should come from a tool call, with the prompt explicitly telling the model not to answer those from memory. In a voice interface this matters more than in chat, because callers cannot see a citation and will not check.
- What about accuracy on accents and noisy audio?
- Word error rate on clean English has largely plateaued, with the top providers within a point or two of each other, so the axes that actually separate them now are streaming latency, end-of-turn detection, multilingual depth, and cost at scale. For accented or noisy audio the only honest answer is to test on your own recordings. We build an eval set from real customer audio before choosing a vendor.
- Can we self-host any of this?
- Not the realtime models. Whisper's open weights remain self-hostable and are still a reasonable choice where the audio cannot leave your infrastructure, though you should expect to trade accuracy and latency for that control. If self-hosting is a hard requirement, say so at the start, because it removes most of this table.
DIRECT INTEGRATION · NO FRAMEWORKWant GPT-Realtime-2.1
Want GPT-Realtime-2.1
in your product?
Eval suite at handoff, full source ownership. We integrate against the model API the same way we integrate against Postgres, and route by task at runtime. Sized to your scope.