OpenAI unveiled three new real‑time audio models for developers: **GPT Realtime 2** (a conversational voice model with GPT‑5‑level reasoning, 128 k token context, parallel tool calling, and adjustable reasoning intensity), **GPT Realtime Translate** (live speech‑to‑speech translation supporting >70 input and 13 output languages), and **GPT Realtime Whisper** (streaming transcription for live captions and note‑taking). The models aim to make voice AI feel like a competent assistant that can understand complex requests, correct itself, and act on multiple tools simultaneously while keeping latency low. Pricing is set per‑million audio tokens or per minute of use, and the models are accessible via OpenAI’s real‑time API with EU data‑residency options and built‑in safety guards.
Underpinning these capabilities is a new networking protocol called **MRC (Multi‑path Reliable Connection)**, which spreads GPU‑to‑GPU traffic across hundreds of paths, recovers from link failures in microseconds, and reduces the number of switch layers needed to interconnect >100 k GPUs. MRC is already running on OpenAI’s largest NVIDIA‑based supercomputers, enabling stable, large‑scale training of frontier models.
The video also touches on the **AI jobs debate**: while some firms cite AI for layoffs that may stem from other pressures, studies so far show little macro‑level employment impact, though experts warn that AI‑driven displacement—especially of entry‑level digital work—could grow and that productivity gains from AI are beginning to appear in the data.
1. OpenAI launched voice AI that can talk, translate, transcribe, and take action in real time.
2. OpenAI introduced three new realtime audio models for developers: GPT Realtime 2, GPT Realtime Translate, and GPT Realtime Whisper.
3. GPT Realtime 2 is built for live spoken conversations with GPT‑5 class reasoning.
4. GPT Realtime 2 can keep track of context, respond to corrections, call tools, and handle multiple actions simultaneously.
5. GPT Realtime 2’s context window increased from 32,000 to 128,000 tokens.
6. OpenAI provides five reasoning intensity levels for developers: minimal, low, medium, high, and XH high; the default is low.
7. On the Big Bench Audio benchmark, GPT Realtime 2 at the high setting achieved 96.6% accuracy versus 81.4% for the previous version.
8. On the audio multi‑challenge benchmark, the X‑high version reached a 48.5% average pass rate versus 34.7% before.
9. GPT Realtime Translate can understand more than 70 input languages and speak back in 13 output languages.
10. GPT Realtime Whisper provides streaming transcription for live captions, notes, summaries, and action items.
11. OpenAI described three voice‑AI patterns: voice‑to‑action, systems‑to‑voice, and voice‑to‑voice.
12. Pricing: GPT Realtime 2 costs $32 per million audio input tokens, cached input $0.40 per million tokens, and $64 per million audio output tokens.
13. GPT Realtime Translate costs $0.034 per minute.
14. GPT Realtime Whisper costs $0.017 per minute.
15. All three models are available via the real‑time API and can be tested in the playground.
16. The API supports EU data residency and is covered by OpenAI’s enterprise privacy commitments.
17. OpenAI built guardrails against spam, fraud, and harmful uses; the system can halt conversations that violate guidelines.
18. MRC (Multi‑ath Reliable Connection) is a new networking protocol for large‑scale AI supercomputer training clusters, developed over two years with AMD, Broadcom, Intel, Microsoft, and Nvidia and published via the Open Compute Project.
19. MRC uses Rossi and RDMA technology plus SRV6 routing to move data directly between machines.
20. MRC spreads data across hundreds of paths to avoid bottlenecks and improve traffic flow.
21. MRC can detect failures and reroute traffic in microseconds, with decision‑making at the network‑card level.
22. If one port fails on an 8‑port network interface, MRC avoids the failed path, reducing capacity by 1/8 and often recovering within about a minute.
23. MRC allows splitting an 800 Gbit/s connection into smaller links, enabling connection of about 131,000 GPUs with only two layers of switches instead of three or four.
24. This design uses two‑thirds of the optics and three‑fifths of the switches compared with a traditional three‑tier network, reducing delay.
25. MRC is already running on OpenAI’s largest NVIDIA GB200 supercomputers, including Oracle Cloud Infrastructure in Abilene, Texas, and Microsoft’s Fairwater supercomputers in Atlanta and Wisconsin.
26. MRC works with 4 and 800 gigabit RDMA network cards from Nvidia, AMD, and Broadcom, with switch support from NVIDIA Spectrum and Broadcom Tomahawk.
27. During a recent training run, OpenAI rebooted four major switches without stopping training.
28. Over 900 million people use ChatGPT every week.
29. A National Bureau of Economic Research study (February) surveyed thousands of executives in the US, UK, Germany, and Australia; nearly 90% said AI had no impact on workplace employment over the past three years after ChatGPT’s launch.
30. Yale Budget Lab found no major evidence yet of AI changing the occupation mix or unemployment length for AI‑exposed jobs through March 2026.
31. Anthropic CEO Dario Amodai warned that AI could wipe out 50% of entry‑level office jobs.
32. Snap CEO Evan Spiegel announced layoffs of about 1,000 people (~16% of the workforce) citing AI.
33. The World Economic Forum’s 2025 Future of Jobs report said around 40% of employers expect to reduce staff because of AI.
34. Apollo Global Management chief economist Torstston Sllock compared the current moment to the old computer era, saying AI is everywhere except in incoming macroeconomic data.
35. Stanford’s Eric Brinolson noted revised job gains of 181,000 while Q4 GDP tracked at 3.7%; his analysis showed a 2.7% year‑over‑year productivity jump linked partly to AI.
36. He also published research showing a 13% relative decline in employment for early‑career workers in jobs highly exposed to AI, while more experienced workers remained stable or grew.