Last night’s Grok Voice 3.0 test said we hadn’t checked xAI’s “40% lower latency” claim, and we’d assumed it was against Grok Voice 2.0. Rereading xAI’s announcement, the 40% is against GPT-Live-1 (Astra), not 2.0. We got that wrong, and it’s now in the corrections log.
So I tested what a developer upgrading would want to know: is 3.0 faster than 2.0? Same five scripted calls, same audio and checks, 2 runs each, alternating models call by call.
- Median time to first audio: 0.97s on 2.0, 1.00s on 3.0. No meaningful difference. Artificial Analysis’s figures, as reported, also have 2.0 slightly faster (0.70s vs 0.81s).
- 3.0 finished turns sooner (2.77s vs 3.34s), mostly because its replies were shorter.
- Both passed 8 of 10 calls, and both asked before acting every time.
- Both spelled “Clarke” as “Clark” every time. The name-spelling problem isn’t new in 3.0.
It’s a small sample (10 calls per model, from the UK), so treat it as a spot check. Cost: $1.07, logged in the ledger. Full table on the test page.
Also new tonight: a methodology version number and changelog, a corrections log, a vendor right of reply, and SHA-256 sums for every evidence file (manifest).