Grok Voice Think Fast 2.0 takes the top spot on speech-to-speech benchmarks

Voice AI has a new benchmark leader, and it is not OpenAI.

On July 29, 2026, xAI released Grok Voice Think Fast 2.0, and it took the top spot on Artificial Analysis’ speech-to-speech benchmark with 82.9%, ahead of GPT-Realtime-2.1 at 79.1% and Gemini 3.1 Flash at 69.5%. On agentic voice tasks the gap is wider still: 56.5% against 45.7% for OpenAI’s model.

The speed numbers are the practical story

Benchmark scores get headlines, but latency decides whether talking to an AI feels like conversation or like leaving voicemails. Here the release delivers real movement:

• Time to first audio dropped from 1.25 to 0.70 seconds
• Transcription errors fell 1.4 to 2x across 24 languages, up to 10x better in noisy and telephony conditions
• Reasoning token use dropped about 60%, which is what makes real-time conversation feel instant

The telephony improvement is worth pausing on. Noisy, compressed phone audio is where voice agents actually get deployed, in support lines and call flows, and it is historically where they fall apart.

Pricing and rollout

Pricing sits at $0.08 per minute of audio. Anyone using the grok-voice-latest alias gets migrated automatically: on August 5 the alias switches to 2.0, so builders who prefer to test first should pin a version before that date.

Voice is becoming the next model war

The competitive rhythm around voice has picked up noticeably. OpenAI shipped voice into ChatGPT Work and Codex last week, and now xAI claims the benchmark crown days later.

Voice is quietly becoming the next front in the model war, and for a reason: it is the interface where speed, cost and quality all have to work at once, and where improvement is immediately felt by the user. The benchmark lead will change hands again. The interesting question is whether talking to an AI becomes something people actually do this year, or stays a demo that impresses and then gets closed.