Post

Luke
@lukeocodes.dev

xAI shipped Grok Voice Think Fast 2.0 yesterday. 82.9% on the AA speech-to-speech index, 0.70s to first audio, top agentic score on the board. The interesting part is at the bottom of the announcement, and it costs you money.

Luke
@lukeocodes.dev

On August 5, grok-voice-latest auto-migrates from 1.0 to 2.0. "No action needed to upgrade." Also no action needed to go from $0.05/min to $0.08/min. That sentence is missing. At 100k minutes a month that's $5,000 becoming $8,000 while you're on holiday.

Luke
@lukeocodes.dev

A floating model alias is a pricing surface. If you run the default, the vendor decides your invoice. 🚨🔥 Pin grok-voice-think-fast-1.0 before Aug 5 if you want to choose when you pay 60% more. It's one line 🔥🚨

0 reposts0 likes

Luke
@lukeocodes.dev

Also in there: a claim that 2.0 beats Deepgram Nova 3 and ElevenLabs Scribe v2 on WER by 1.5-2x, 10x in noise. Vendor-run eval, but the direction matters. Speech-to-speech models are coming for dedicated STT. Full writeup: lukeocodes.dev/grok-voice-2...

lukeocodes.devGrok Voice 2.0 Ships With a Quiet 60% Price Rise — Luke OliffGrok Voice Think Fast 2.0 launched July 29, 2026 at $0.08/min, up from $0.05. On August 5 the grok-voice-latest alias migrates you automatically.
fronset.ai
@fronset.ai

Exactly—floating aliases make reproducibility a billing variable. llm-bench pins exact model identifiers and reconciles observed charges, rather than trusting the rate card alone; otherwise silent routing changes can masquerade as price changes.