Grok is now a first-class citizen across Gravity Rail. Grok 4.5 is available as a chat model everywhere agents think, and Grok Voice Think Fast 2 is our recommended voice model for phone calls and web voice chats. Both run under Zero Data Retention with a signed BAA, which is what makes them usable for real patient communication rather than just demos.
Here's what changed and where you'll see it.
Grok 4.5 for reasoning
Grok 4.5 is selectable anywhere you pick a Chat Model — on an Agent, on a Workflow, or as a per-Chat override. That model drives web chat, SMS, and email.
Two things make it worth reaching for:
- A 500K-token context window. Long Assignment histories, large Member records, and hefty knowledge-base excerpts fit in a single prompt instead of being summarised down and losing detail.
- Configurable reasoning effort. Grok 4.5 takes a low / medium / high reasoning setting, so the same model covers a quick intake triage and a careful clinical-protocol read without swapping models. You'll find it behind the gear icon next to the Chat Model picker.
Pricing follows xAI's published rate card, metered in AI Units like every other model. Note the long-context tier: prompts over 200K tokens bill at double the short-context rate, so an agent that routinely stuffs its whole context window costs meaningfully more than one that doesn't. The cost estimator in the model picker reflects this.
Grok Voice Think Fast 2 for every call
Voice Think Fast 2 is a speech-to-speech model. That distinction matters more than it sounds.
The conventional way to build a voice agent is a pipeline: speech-to-text, then a language model, then text-to-speech. Every hop adds latency, and every hop throws away information — tone, hesitation, the sound of someone getting upset — because the middle of the pipeline only ever sees text. A speech-to-speech model hears the audio and answers in audio. Nothing gets flattened to a transcript in between.
In practice that means shorter gaps before the agent starts talking, and interruptions that behave the way they do in a real conversation. Think Fast 2 also improved telephony-band transcription specifically, which is the hard case: 8kHz phone audio is much harder to work with than a laptop microphone, and it's what almost all of our voice traffic actually is.
Think Fast 2 bills at a flat rate per minute of audio regardless of how much is said — which makes voice cost a straightforward function of call minutes rather than something you have to model per conversation.
Grok Voice Think Fast 1 is now retired from the picker. Agents already pinned to it keep working and keep billing at the v1 rate — nothing breaks and nothing needs migrating. It just won't be offered for new agents. When you're ready, switching to Think Fast 2 is a single change on the agent.
Voice Presets: pick a voice, not a stack
Choosing a voice used to mean choosing an engine, then a voice, then a set of settings. Most people wanted a voice.
The Voice Model picker now leads with Voice Presets: a list of voices by name, each one bundling a model and the settings we recommend for it. Each has a preview button, so you can hear a voice before committing to it. Pick "Ara" and you're done.
xAI recently expanded its catalogue to 26 voices, and all 26 are available — each one multilingual. Previously we shipped five of them.
Nothing is hidden. More Models and Voices opens the full picker: every voice model we support, including composable pipeline stacks that let you mix a listening model, a thinking model, and a speech model from different vendors. The gear icon opens the same dialog, and it can now change the voice model itself — so switching an agent from Grok Voice to a custom pipeline stack no longer means backing out to a different screen. Adjust anything a preset bundles and the picker simply says Custom.
Which model does what
The two model settings on an agent do genuinely different jobs, and the labels now say so:
| Setting | Drives |
|---|---|
| Chat Model | Web chat, SMS, and email |
| Voice Model | Phone calls and web voice chats |
An agent can run Grok 4.5 for its text channels and Grok Voice Think Fast 2 on the phone, and that's the configuration we'd suggest starting from. They're independent settings — mixing vendors across the two is fine and common.
The compliance part
None of the above would matter for healthcare if the data posture were wrong.
Gravity Rail has Zero Data Retention enabled on its xAI account and a signed BAA. That's what makes these endpoints approved for HIPAA workspaces, and it's why we're comfortable putting Grok on live patient calls rather than restricting it to internal tooling.
As always, PHI authorization comes from your workspace's explicit PHI posture and contractual coverage — adding a model never changes that on its own. If you're unsure what your workspace is provisioned for, ask us and we'll confirm.
Getting it
Nothing to install. Grok 4.5 is in the Chat Model list and the Grok voices are in the Voice Model picker for every workspace.
If you're running a voice agent on an older model, the fastest thing you can do is open the Voice Model picker, preview two or three of the new voices, and pick one. It takes about a minute, and it's the change your callers will actually notice.