On October 14, 2026, SpaceXSI released Grok Voice 3.0, an AI model that listens to speech and replies in speech. It scored 86.9 on the overall speech-conversation index from the independent evaluator Artificial Analysis, ahead of the OpenAI and Google comparison models. The Voice Agent API is priced at $0.08 per minute, the same as the previous version. However, the big improvement over the previous generation is in completing customer-service tasks, while the time to start a response actually got longer. Separating the gains from the trade-offs shows what this top ranking means for running phone support desks and building voice apps.
Customer service drove the gain in the overall score
Grok Voice 3.0's overall index rose from 81.3 for the previous generation to 86.9. In the comparison table published by Artificial Analysis, GPT-Live-1 (Astra, medium setting) scores 83.2 and Gemini 3.8 Live (extended reasoning, High setting) scores 82.6. Even for the same model name, different settings produce different results, so rankings cannot be discussed without these conditions.
The overall index is an equally weighted average of four components. It evaluates the ability to solve problems posed by voice and the ability to complete customer-service tasks. It also adds the degree to which people prefer the conversation, and the share of target conversations in which the final tool operation was completed correctly. Only models with results for every component receive a score. Full Duplex Bench, which handles conversational pauses and interruptions, is presented as a separate metric.
| Metric | Grok Voice 2.0 | Grok Voice 3.0 | GPT-Live-1 (Astra, medium) | Gemini 3.8 Live (extended reasoning, High) |
|---|---|---|---|---|
| Overall index | 81.3 | 86.9 | 83.2 | 82.6 |
| Customer service (τ-Voice) | 56.5% | 71.6% | 74.5% | 68.6% |
| Time to start of response | 0.70 sec | 0.81 sec | 1.34 sec | 1.35 sec |
Source: Artificial Analysis public comparison table (checked October 15, 2026). τ-Voice is the score on simulated customer service, and time to start of response is measured on Big Bench Audio; the two come from separate tests. τ-Voice averages three trials where possible, while the GPT-Live-1 Astra setting is reported as a single-trial result.
Of the 5.6-point rise in the overall index, the τ-Voice improvement alone accounts for about 3.8 points. The difference between the new and old τ-Voice scores is 71.6 − 56.5 = 15.1 points. Because the four components are averaged with equal weight, the contribution to the overall score is 15.1 ÷ 4 = 3.775, or about 3.8 points when rounded. This is a score-based contribution calculated from published figures; it does not show how much the training method or real-world productivity improved.
In τ-Voice, a simulated voice customer raises issues such as changing a flight, a retail billing problem, or a telecom service problem. The AI handles them using business tools and rules, and is judged on whether the final database state matches the correct answer. Smooth conversation alone earns no points; the requested task has to be completed.
That said, in this test GPT-Live-1's 74.5% beats Grok's 71.6%. Taking first place overall is not the same as resolving each individual inquiry most reliably. For the preference-based component, the overall index uses a value fixed at publication, so working backward from the continually updated conversation-preference ranking will not match the overall score.
How to compare sub-second responses and pricing
Grok Voice 3.0 starts responding in 0.81 seconds, 0.11 seconds longer than the old version's 0.70 seconds. GPT-Live-1, the comparison model, takes 1.34 seconds, so dividing the gap by the GPT time shows Grok about 40% shorter. It raised its customer-service score while keeping a shorter wait than its competitors, but not everything got faster compared with the previous generation.
This time is the average until the first audio token is generated on Big Bench Audio. It is neither the total wait over an entire call including the phone line, nor the time to finish resolving a request in τ-Voice. To judge how it feels at a support desk, you need to check separately how quickly it starts speaking and how long users wait until processing is complete.
There is also a difference in denominators in the pricing comparison. The official pricing page lists Grok Voice 3.0 at $0.08 per minute, or $4.80 per hour, and also lists $0.004 per text input. It confirms that the hourly rate was held steady, but in implementations that also use text, the hourly charge alone is not the whole bill.
By contrast, the cost column at Artificial Analysis converts the cost of processing a fixed set of 40 questions into an hourly figure based on the length of input audio. It includes audio and text input and output, plus reasoning tokens shown separately, but excludes tool-usage fees and cache discounts. Under those conditions, Grok costs $4.80, GPT-Live-1 with the Astra setting costs $5.83, and Gemini with the extended reasoning High setting costs $3.50.
In other words, Grok is cheaper than this OpenAI comparison model but more expensive than the Google model. Even when the same $4.80 figure appears, the API's hourly rate and the cost converted from evaluation audio are different things. When estimating adoption costs, it is best to compare under conditions close to your own calls, including how long users speak and how long the AI replies.
Starlink's track record, and what business evaluations measure
According to SpaceXSI, Starlink's call center uses Grok Voice to handle 20,000 inquiries a day in multiple languages and resolves about 70% without human intervention. Agents using Grok Voice 3.0 reportedly move from identity verification to equipment troubleshooting, and even issue credits and ship replacement equipment. It is a deployment example that goes beyond answering to carry out processing tied to contracts and equipment.
However, the 70% is the company's own reported result, not an improvement rate achieved solely by updating to the new version. The announcement does not give the aggregation period, the mix of inquiries, or the repeat-inquiry rate, so another company's desk may not achieve the same share.
On the enterprise evaluation EVA-Bench as well, the company announced that Grok Voice 3.0 ranked first with 74% accuracy. ServiceNow AI Research's EVA has simulated bots hold multi-turn voice conversations and evaluates separately whether the task was completed correctly and whether the conversation experience was good. It consists of 50 customer-service scenarios such as flight changes, 80 internal IT support scenarios, and 83 HR-related ones.
EVA's design shows why conversational impression and business outcome are measured separately. If a response is easy to hear but the processing is wrong, the support job is not done; and even with correct operations, repeated requests to say it again leave users waiting. Starlink's report shows scale in real operations, while simulated tests help with comparisons under matched conditions. Each answers a different question.
Existing APIs have already switched over
Existing sessions that use grok-voice-latest have already switched to Grok Voice 3.0. According to the developer documentation, the endpoint, session parameters, and event specifications are the same as in the old version. Grok Voice 2.0 is deprecated, and those who specify the version name directly are advised to change to grok-voice-3.0.
Even if API compatibility is maintained, the model's responses are not necessarily identical. The documentation recommends "pinning to a version-specific model name" for production stability. When a name that points to the latest version is updated, as happened here, there is reason to recheck the content and timing of replies even if the app's code has not changed.
Meanwhile, guidance for those migrating from OpenAI Realtime recommends shortening the system prompt and removing instructions added solely to work around bugs or limitations in GPT models. Reasoning defaults to reasoning.effort set to high, and it can be disabled by setting it to none. The relationship between this setting and response time should be measured on your own use cases, and the speed differences in the public table cannot be assumed to result from the reasoning setting alone.
Supported languages include Japanese. If you adopt it for a Japanese-language phone desk, you would use real inquiries, fix the model version and settings, and compare task completion rate and response wait time at the same time. If both meet the required level, the improved customer-service ability this time could potentially be used to reduce the number of calls handed off to humans.