OpenAI adds GPT-Realtime-2.1-mini to the API with reasoning and tool use

GPT-Realtime-2.1-mini is available in the OpenAI API as gpt-realtime-2.1-mini, a distilled reasoning model for realtime voice that supports function calling and reasoning tokens at the same published price as GPT-Realtime-mini. OpenAI announced it on July 6, 2026, before its own docs listed it; the model, its specifications, and its pricing are now documented on OpenAI's developer platform.

This matters if you build lower-cost voice or streaming agents: reasoning and tool use are reachable at the mini price point rather than only on larger Realtime models.

Key Takeaways

• Model string: gpt-realtime-2.1-mini, listed on OpenAI's models page alongside gpt-realtime-2.1.
• What it does: OpenAI documents it as a distilled reasoning model optimized for "faster, lower-cost realtime voice interactions", connecting over WebRTC, WebSocket, or SIP.
• Capabilities: reasoning token support and function calling (tool use) are both listed on the model page.
• Cost: identical published rates to GPT-Realtime-mini -- $0.60/$2.40 per 1M text input/output tokens and $10.00/$20.00 per 1M audio input/output tokens (pricing).

GPT-Realtime-2.1-mini Specifications

Source: OpenAI's GPT-Realtime-2.1 mini model page.

Pricing: The Cost-Parity Claim, Verified

OpenAI's announcement said the model arrived "at the same cost as GPT-Realtime-mini." OpenAI's pricing page now lists both models, and the rates match:

Source: OpenAI API pricing, checked 2026-08-03. The practical read is that moving a mini-tier voice agent from gpt-realtime-mini to gpt-realtime-2.1-mini adds reasoning and tool use without changing the published per-token cost. Reasoning tokens are billed as output tokens, so a model that reasons more can still cost more per interaction even at an identical rate card -- measure your own workload rather than assuming parity in the bill.

What OpenAI Announced

The OpenAI Developers account stated that "GPT-Realtime-2.1-mini is now available in the API, bringing reasoning and tool use to our Realtime mini lineup at the same cost as GPT-Realtime-mini" (OpenAI Developers on X, July 6, 2026).

OpenAI's API changelog carries the release entry for both models, describing gpt-realtime-2.1 as "an updated realtime reasoning model with improved alphanumeric recognition, silence and noise handling, and interruption behavior", and gpt-realtime-2.1-mini as "a faster, lower-cost distilled reasoning model for realtime voice applications."

Note the difference in emphasis: the announcement leads with reasoning and tool use, while the changelog entry for the full gpt-realtime-2.1 model leads with speech-handling improvements -- alphanumeric recognition, silence and noise handling, and interruption behavior. OpenAI has not published benchmark figures for either claim, so treat the speech-quality improvements as vendor-stated and unquantified.

Why It Matters for Voice and Streaming Agents

For real-time voice and low-latency streaming, the value of a mini Realtime model is cost per interaction. With reasoning and function calling documented at the mini rate card, a voice agent can plan and call tools without moving up to a larger Realtime model.

Two things worth validating before you ship it. First, reasoning latency: a model that emits reasoning tokens before speaking can feel slower in a live conversation than one that does not, and OpenAI publishes no latency figures for this model. Second, the reasoning-token billing noted above. Both are workload-specific, and neither is answered by the rate card.

Sources

• OpenAI models page -- lists gpt-realtime-2.1 and gpt-realtime-2.1-mini.
• GPT-Realtime-2.1 mini model page -- description, capabilities, context window, modalities, knowledge cutoff, pricing.
• OpenAI API pricing -- per-1M-token rates for both mini models.
• OpenAI API changelog -- release entry for gpt-realtime-2.1 and gpt-realtime-2.1-mini.
• OpenAI Developers on X -- the original July 6, 2026 announcement. X gates automated reads, so this post is cited as published and every claim above is independently grounded in the OpenAI-owned pages listed here.

Updated 2026-08-03: when this article was published on July 7, 2026, OpenAI's platform docs had not yet listed the model, so it recorded only what the announcement stated. OpenAI now documents the model, its specifications, and its pricing; the specifications, pricing table, and cost-parity verification above were added from those sources.