SpaceXAI kept pricing at $0.10 per batch hour and $0.20 for streaming while claiming a 2x accuracy gain.
By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
· Published
Primary source: [SpaceXAI](https://x.ai/news/grok-voice-transcribe-2)
Why it matters #
Musk is turning SpaceXAI's costly audio research into a low-priced infrastructure service. Loom gives the model immediate volume and a credible route into agent workflows.
Elon Musk (@elonmusk)'s SpaceXAI released Grok Voice Transcribe 2.0 on September 18th, pairing a claimed jump in speech recognition accuracy with a customer that immediately gives the model meaningful distribution: Atlassian says Loom now uses it to transcribe every video.
The model became available through SpaceXAI's API on September 17th, according to the release notes, before the public launch the following day. Existing speech-to-text integrations can move to version 2.0 without code changes. SpaceXAI plans to make it the default and deprecate version 1.0 in the coming weeks, though developers can temporarily pin the older model.
Musk founded xAI in 2023 with a stated mission to understand "the true nature of the universe." He studied physics and business at the University of Pennsylvania, according to his Tesla biography. The transcription release shows how far SpaceXAI's commercial work has traveled from that founding language: a metered API for call centers, video platforms, voice agents and devices where names, numbers and email addresses need to survive a bad phone line.
That practicality is the point. Founded in 2023, xAI has raised successive funding rounds to develop large models and build compute infrastructure, including a $20 billion Series E announced in January. SpaceXAI is now turning the resulting audio foundation model into cheap, narrowly useful services that other software makers can embed without adopting Grok as a consumer assistant.
The benchmark supports a narrower claim than the launch language
SpaceXAI says Grok Voice Transcribe 2.0 is twice as accurate as version 1.0 at the same price. Its public evidence includes the Artificial Analysis streaming leaderboard, where SpaceXAI says the model ranks first for accuracy among 32 systems. The Artificial Analysis page currently describes its benchmark as covering 27 of 33 models, so the count and ranking should be read as the company's characterization of the comparison.
Artificial Analysis measures word error rate using roughly eight hours of audio. Its methodology gives 50% of the score to AA-AgentTalk, a private dataset focused on speech directed at voice agents, with VoxPopuli and Earnings22 accounting for 25% each. The benchmark includes network delay and separately measures final and partial transcripts after speech ends.
The independent leaderboard provides support for a narrower benchmark result. It does not independently establish SpaceXAI's claim that version 2.0 is twice as accurate as version 1.0 across the company's evaluations. SpaceXAI's broader claims about telephony, spoken credentials and multilingual commands rely on internal evaluations drawn from production traffic.
SpaceXAI tested version 2.0 on four sets: 8 kHz customer-support calls, English conversations with Grok, spoken account codes and contact details, and short commands across 19 languages. SpaceXAI says version 2.0 beat its predecessor on all four and led every model it tested on telephony.
The sharpest disclosed gain came from short multilingual phrases, where reported word error rate fell from 20.6% to 6.8%. That is roughly a 67% reduction in errors. SpaceXAI uses the broader phrase "twice as accurate" to summarize performance across its evaluations, so developers assessing a particular language or audio environment will still need to test their own recordings.
The production focus is unusually specific. Clean podcast audio has become a comfortable benchmark case for transcription vendors. SpaceXAI instead emphasizes overlapping speakers, narrowband phone calls, regional accents, language switching and structured information read aloud. Those are the errors that break customer-support routing and voice-agent tool calls, even when the surrounding transcript looks readable.
SpaceXAI is using price to remove the procurement debate
Grok Voice Transcribe 2.0 costs $0.10 per hour for recorded audio and $0.20 per hour for streaming, matching version 1.0. The current SpaceXAI pricing page confirms those rates.
Speaker diarization, word-level timestamps, confidence scores and key-term biasing are included. The API can transcribe up to eight channels independently and accept as many as 100 domain terms per request. It also formats dates, currencies, phone numbers and email addresses, removes filler words and offers end-of-turn detection for voice agents.
SpaceXAI's pricing makes experimentation inexpensive enough that model switching can be decided with a test corpus rather than a purchasing cycle. Maintaining the same API shape also lowers the cost for existing customers to accept the upgrade. SpaceXAI gets production audio volume and a chance to expand those accounts into its higher-priced speech-to-speech service.
The model supports 25 languages, recorded files and real-time streams, according to the speech-to-text documentation. SpaceXAI says its training material includes noisy, multilingual recordings from varied environments, followed by post-training intended to improve recognition under those conditions.
Loom turns transcription into infrastructure for agents
Atlassian says Grok Voice Transcribe 2.0 was more accurate than Loom's previous system and now handles every Loom video. Sanchan Saxena, Atlassian's senior vice president for its Teamwork Collection, framed the deployment as a bridge between recorded context and work performed elsewhere.
SpaceXAI illustrates that idea with a developer recording an action plan in Loom, passing the transcript to Cursor and having the coding agent make the requested changes. The launch presents this as a workflow, rather than documenting a generally available native integration between Loom and Cursor.
The distinction matters because Loom gives SpaceXAI a high-volume reference customer, while the larger commercial opportunity comes from transcripts becoming input for software agents. A missed product name in a meeting transcript is annoying. A missed repository name, account number or command can send an agent toward the wrong system.
SpaceXAI says the underlying Grok Voice technology already handles tens of thousands of customer-support calls each day, transcribes millions of hours of video narration and runs inside physical products, including the Grok assistant in Tesla vehicles. Those usage figures are SpaceXAI's own and cover the wider voice system, rather than Transcribe 2.0 alone.
Musk is building a full API shelf, one practical service at a time
The transcription launch follows SpaceXAI's April release of standalone speech-to-text and text-to-speech endpoints and its July introduction of Grok Voice Think Fast 2.0. It also fits a broader push to place Grok inside software developers already use. RuntimeWire reported in August that SpaceXAI put Grok 4.6 in Google Cloud's Vertex AI Model Garden, extending distribution beyond SpaceXAI's own console.
SpaceXAI raised a $20 billion Series E in January from investors including Valor Equity Partners, StepStone Group, Fidelity, Qatar Investment Authority, MGX and Baron Capital Group, with Nvidia and Cisco Investments participating as strategic investors. SpaceX then acquired xAI on February 2nd, creating the SpaceXAI identity now attached to the Grok product line.
For Musk, Grok Voice Transcribe 2.0 is a small API with a direct route into a large market. Its importance rests on whether SpaceXAI can repeat the Loom deployment across call centers, developer tools and device makers. At $0.10 an hour, SpaceXAI has made the first test cheap.