27 October 2026 · Site Methodology & Pricing Notes · 5 min read
I went to bump the "last verified" date on our Echtzeit-Sprach-API-Kostenrechner -- it hadn't been touched since mid-August, back when "the Realtime API" meant one model. Pulling up OpenAI's current pricing page to grab a fresh number, I counted six separate Realtime model strings: gpt-realtime, gpt-realtime-1.5, gpt-realtime-2, gpt-realtime-2.1, gpt-realtime-mini, Und gpt-realtime-2.1-mini. My first assumption was that each version bump came with its own price, the way GPT generations usually do. It didn't. Lined up side by side, those six names collapse into exactly two prices.
| Model string | Audio input | Audio output |
|---|---|---|
| gpt-realtime | $32/1M tok | $64/1M tok |
| gpt-realtime-1.5 | $32/1M tok | $64/1M tok |
| gpt-realtime-2 | $32/1M tok | $64/1M tok |
| gpt-realtime-2.1 | $32/1M tok | $64/1M tok |
| gpt-realtime-mini | $10/1M tok | $20/1M tok |
| gpt-realtime-2.1-mini | $10/1M tok | $20/1M tok |
Checked directly against OpenAI's own pricing documentation, not a secondary source. Four full-size version strings, one price. Two mini version strings, one (different) price. The "2.1" in the name tells you nothing about cost -- it's the same $32 input / $64 output as the original gpt-realtime string from over a year ago. The only variable that moves your bill is whether the string ends in -mini.
Every other OpenAI model family we track -- GPT-5, GPT-5 mini, GPT-5 nano -- prices each tier and each version separately, so "check the model string before you budget" is already a reflex for text models. Realtime looks the same on the surface: four version-numbered names sitting in the same pricing table, inviting the same assumption that newer costs more (or less). It doesn't, here. If you're comparing gpt-realtime-1.5 against gpt-realtime-2.1 for cost reasons, you're comparing a number against itself. The comparison that actually matters -- full-size versus mini -- is a 3.2x gap on both audio input and audio output, applied as cleanly as a flat multiplier: $10 is $32 divided by 3.2, $20 is $64 divided by 3.2, exactly.
That symmetry matters in practice. Because the ratio is identical on input and output, switching from any full-size Realtime model to the matching mini cuts your Realtime bill by the same 3.2x regardless of how input-heavy or output-heavy your conversation traffic is. You don't need to re-model your input/output split to estimate the savings -- divide the current bill by 3.2 and that's the mini-tier number, give or take whatever behavioral differences the smaller model introduces on your actual calls.
If your code or infra pins an exact Realtime model string -- common practice for reproducibility -- a teammate deploying with gpt-realtime-2 while you're still on gpt-realtime-1.5 will see an identical bill at identical usage. No cost audit needed there; that diff is a non-issue. The audit that's actually worth running is a grep for -mini across your codebase and infra configs, because that's the one substitution that silently changes a bill by over 3x in either direction depending on which way someone switched it.
Refreshed the "last verified" date on the Echtzeit-Sprach-API-Kostenrechner and added a note pointing out that the per-minute default rates on that page correspond to full-size Realtime pricing, with a pointer to divide by roughly 3.2x for the mini tier instead of treating it as a separate manual entry. If you're running or speccing a voice agent and the per-minute math is the part you care about, that calculator still does the multiplication -- this post is just the reminder to check which model string you're actually passing in the API call before you trust the number.
New to API cost tracking in general? Start with the kostenlose API-Kostenleitfäden.
Stellen Sie es selbst bereit: DigitalOcean – 200 $ kostenloses Guthaben ↗ - Hostinger VPS ↗
Checked directly against OpenAI's official API pricing documentation, 2026-10-27. Reference figures, not live quotes -- confirm current terms on OpenAI's pricing page before budgeting.