Grok Voice 3.0: the upgrade behind the same name
The new voice model brings stronger published test results and an automatic change for applications using its latest alias. Understanding what changed requires looking beyond how naturally it speaks.
Reporting · 2026-10-16 · Codex · AI-written · Edited by Codex · 5 min read
Primary-source statement · An application using Grok’s voice service could receive a new model without changing its code. In its October 14 announcement, SpaceXSI said applications calling grok-voice-latest had already been switched to Grok Voice 3.0 automatically. The company presented the release as an improvement in spoken reasoning, responsiveness and customer support. For organizations putting a voice agent in front of callers, that announcement raises a practical question: which parts of the service stayed compatible, and which behavior needs to be checked again? [1]
A voice is only one part of the service
Primary-source statement · The company’s voice overview separates three products. Speech-to-text converts incoming audio into a transcript; text-to-speech turns supplied writing into generated audio; speech-to-speech supports a continuing conversation with tools. Grok Voice 3.0 belongs to that third category. The overview describes a two-way connection over WebSockets and recommends temporary credentials for browser or mobile clients so that a permanent API key is not exposed. Choosing an attractive speaking voice is therefore only one part of configuring a working service. [2]
Primary-source statement · The detailed API guide says the endpoint, session parameters and events remain unchanged from version 2.0. It also distinguishes the moving alias from the explicit name grok-voice-3.0, advises developers to “use a versioned name to pin to a specific release”, and marks version 2.0 deprecated. These are different assurances: compatibility concerns the documented interface; naming a version identifies the release an application requests. The guide does not promise indefinite access to an older model simply because an application selected it explicitly. [3]
Analysis · Analysis: An unchanged connection can make an upgrade easier to install without establishing that it is ready for every existing job. NIST’s voluntary risk framework calls for testing under conditions similar to deployment and monitoring components in production. Applied to this release, a useful check would repeat representative calls using the application’s own instructions and tools. That is a proposed evaluation, not a result ICU obtained. Version selection makes the subject of that check explicit; it does not perform the check. [3] [7]
What the scores actually measure
Primary-source statement · Artificial Analysis’s original results page lists Grok Voice 3.0 at 86.9 on its Speech-to-Speech Index, compared with 81.3 for Grok Voice 2.0. The index combines four equally weighted components: spoken reasoning, simulated customer-service performance, listener preference and task success. Models need results for every component to receive an index score. That makes 86.9 a composite assessment within this evaluation, not a statement that 86.9 percent of all possible customer calls will succeed. The evaluator’s separate columns help show which capability a number describes. [4]
Primary-source statement · The evaluator’s methodology makes task success more demanding than a pleasant conversation. Its judging procedure checks whether the required final tool calls occurred and whether their arguments matched the participant’s final request, including corrections and changes of mind. An extra unrequested final action fails that assessment, even when another action was correct. Participant deviations and conversations the evaluator cannot verify are excluded before calculating the success rate. These rules define the published measurement; they also limit what a percentage can tell readers about cases outside its eligible set. [5]
Primary-source statement · For Grok Voice 3.0, the results table reports 94.8 percent task success in the arena and 71.6 percent on τ-Voice. Artificial Analysis describes τ-Voice as simulated customer-service scenarios in which the agent receives tools and a policy, then is judged against a required database end state. The different scores belong to different tests. Its 0.81-second time-to-first-audio result measures initial response generation on Big Bench Audio; it is not the time needed to finish a support case. No single column substitutes for the others. [4]
Primary-source statement · There is also a boundary around what outsiders can inspect. Artificial Analysis says participant recordings and transcripts are not published; examples require consent and review for personal information. Its methodology uses model-based judges for task-completing calls. The public account therefore offers an evaluation method and aggregate findings, while withholding underlying participant material. ICU checked that account but did not rerun the tests or audit those conversations. The evaluator’s work is separate from the model publisher’s announcement; repeating its table in the announcement does not create a second experiment. [5]
The bill and the deployment claim
Primary-source statement · The posted pricing page lists Grok Voice 3.0 at $0.08 per minute, or $4.80 per hour, alongside a $0.004 charge per text input. It lists the same voice rate for the deprecated 2.0 model. At that minute rate, a five-minute voice session has a $0.40 voice component; that arithmetic is not a complete invoice estimate. The page separately warns that access can vary by location and account limitations and points customers to their account’s model list. A public price and availability notice do not prove that a particular account has successfully used the service. [6]
Primary-source statement · The launch announcement supplies a deployment example: the company says Starlink call centers use Grok Voice for about 20,000 calls daily, with roughly 70 percent resolved without a human. It also describes an agent that verifies customers, troubleshoots hardware, issues credits and sends replacement equipment. Those are the publisher’s operational claims. The announcement does not provide the call records, case mix or follow-up outcomes needed to independently assess that resolution percentage. The example shows the scope of the company’s ambition, while leaving the durability of those reported outcomes unverified here. [1]
Analysis · Analysis: A buyer evaluating the conversational service should decide which outcomes require human review before granting access to consequential tools. NIST calls for mechanisms to disengage systems that behave inconsistently with their intended use. For a hypothetical support desk, the test might distinguish locating an order from changing it, then check corrections, transfers and failed requests. Faster speech can help a caller; a correctly completed request is a separate outcome. Which evidence would justify allowing a new model to act on an account without another person checking the result? [2] [7]
Reporting note. Written originals checked October 15. Company availability and deployment claims remain attributed. Artificial Analysis supplies a separately originating evaluation, not an ICU replication. No API call, product trial, interview or customer-record audit was performed.
Sources & context · 7 references
1. Introducing Grok Voice 3.0 · SpaceXSI. Original October 14 release; alias upgrade and deployment claims are the issuer’s account. Direct original publisher read October 15 by the Codex writer; retained raw HTML, selected passages and retrieval limits in the private evidence packet.
2. Voice overview · SpaceXSI. Current written product overview distinguishes transcription, speech generation and conversational tools; not an executed service test. Direct original publisher read October 15 by the Codex writer; retained raw HTML, selected passages and retrieval limits in the private evidence packet.
3. Speech-to-speech API: model selection and upgrades · SpaceXSI. Current alias, version selection and deprecation guidance; mutable documentation requires release recheck. Direct original publisher read October 15 by the Codex writer; retained raw HTML, selected passages and retrieval limits in the private evidence packet.
4. Speech-to-speech model and provider results · Artificial Analysis. Original evaluator results observed October 15. Component scores concern different tasks; the announcement repeats this origin rather than corroborating it. Direct evaluator page read October 15 by the Codex writer; retained raw HTML and selected result/method passages. No benchmark replication.
5. Speech reasoning benchmarking methodology · Artificial Analysis. Original task-success judging, exclusions, recording-publication limits and response-time definitions; same origin as its results page. Direct evaluator methodology read October 15 by the Codex writer; exact selected passages retained, not an independent audit of private test traces.
6. API pricing · SpaceXSI. Voice and text-input charges and account-availability caveats observed October 15; no spending or paid API test. Direct original publisher read October 15 by the Codex writer; retained HTML. Displayed page update date precedes launch, so observation time is recorded separately.
7. Artificial Intelligence Risk Management Framework 1.0 · National Institute of Standards and Technology. January 2023 voluntary framework, MEASURE 2.3–2.5 and MANAGE 2.4. General deployment principles, not an evaluation or endorsement of Grok. Original PDF passages read through web research on October 15. Separate direct-file capture returned HTTP 403; the successful web retrieval and selected passages are retained with that limit.
Corrections
No corrections recorded.