OpenAI's gpt-realtime-2.1-mini Brings Reasoning and Tool Use at 6x Lower Cost

OpenAI just pushed two updates to its Realtime API that matter for anyone building voice agents: a new model and a meaningful latency improvement that applies to everything in the lineup.

A smarter mini, at the same price

GPT-Realtime mini is capable of responding to audio and text inputs in realtime over WebRTC, WebSocket, or SIP connections. The original mini was fast and cheap, but it lacked two things that production voice agents almost always need: the ability to reason through a problem before speaking, and reliable tool calling. gpt-realtime-2.1-mini closes that gap.

The new model brings reasoning and tool use to the mini tier, matching the capabilities that GPT-Realtime-2, the most capable realtime voice model, already had , supporting speech-to-speech interactions with configurable reasoning effort, stronger instruction following, and more reliable tool use for complex voice-agent workflows. The key difference: you now get those features at the mini price point.

Pricing stays identical to the original gpt-realtime-mini. Text tokens are priced at $0.60 per 1M input and $2.40 per 1M output. That is a fraction of the flagship gpt-realtime-2 cost, which runs $4.00 per 1M input and $24.00 per 1M output for text.

Why reasoning in a voice model is harder than it sounds

GPT-Realtime-2 was a generational upgrade to OpenAI's speech-to-speech model, bringing internal reasoning to real-time voice. Where previous models responded immediately, it can work through a problem before speaking , making it well suited for voice applications that need to handle complex, multi-step queries entirely in the audio layer without routing to a separate text pipeline.

The challenge with reasoning in voice is the silence problem. The silence problem kills voice agents in production. You ask the model to pull CRM data, it fires off a tool call, and then nothing , three seconds of dead air while the user wonders if the call dropped. The recommended fix is the preamble pattern: prompting the model to verbally acknowledge what it is doing before it calls a tool. Reasoning effort is explicitly adjustable via reasoning.effort , you can request minimal, low, medium, or high , to save on cost and latency.

For most production voice agents, setting reasoning effort to low instead of the default is recommended, increasing only for workflows that require deeper planning.

25% latency cut, across the board

The second announcement is arguably the more impactful one for teams already in production. OpenAI reduced p95 latency by at least 25% across all Realtime voice models through improved caching. P95 latency is the response time at the 95th percentile , meaning 95% of requests now complete faster than before. This is the tail latency number that determines whether a live voice conversation feels natural or broken.

This improvement comes from better prompt caching. In a Realtime session, the system prompt and conversation history are re-sent with every turn. Caching those repeated tokens means the model spends less time processing context it has already seen, which directly cuts the time before the first audio byte starts playing.

Where to use it

The mini tier now covers a much wider set of real use cases. Here is how to think about which model to reach for:

• gpt-realtime-2.1-mini: Customer support bots that need to look up order status, check inventory, or update records. Any voice agent that calls tools but doesn't need heavy multi-step reasoning. Cost-sensitive, high-volume deployments.
• gpt-realtime-2: Complex troubleshooting flows, multi-step agentic tasks, scenarios where the model needs to reason through ambiguous inputs before acting. Use reasoning.effort: high for complex issue resolution within a continuous audio pipeline.

The model connects over WebRTC, WebSocket, or SIP, so your existing Realtime integration works without transport changes. You just swap the model name in your session config.

Here is a minimal session setup using the Python SDK:

import openai
client = openai.OpenAI()
# Create a realtime session with the new mini model
session = client.realtime.sessions.create(
    model="gpt-realtime-2.1-mini",
    modalities=["audio", "text"],
    instructions="You are a helpful support agent. When calling tools, always acknowledge what you're doing first.",
    tools=[
        {
            "type": "function",
            "name": "lookup_order",
            "description": "Look up an order by ID",
            "parameters": {
                "type": "object",
                "properties": {"order_id": {"type": "string"}},
                "required": ["order_id"]
            }
        }
    ],
    # Start low, increase only if needed
    reasoning={"effort": "low"}
)

The bigger picture

Voice can be one of the most direct and productive interfaces for AI , enabling customer support agents that may resolve issues without a single keystroke, live multilingual communication, and voice assistants capable of reasoning through complex requests in real time. Developers building these experiences need models that can keep pace with increasingly demanding latency, accuracy, and language coverage requirements.

The Realtime API lineup now spans a clear spectrum: GPT-Realtime-2 for configurable reasoning in speech-to-speech agents, GPT-Realtime-Translate for streaming speech translation, and GPT-Realtime-Whisper for streaming speech-to-text. The addition of gpt-realtime-2.1-mini fills in the cost-efficient slot with the capabilities that were previously only available at the top of the stack. For teams that shelved voice agent plans because tool-calling mini models were too unreliable, this is the update worth revisiting.