Meta is quietly building custom voice generation and a fresh round of model upgrades into Muse, its AI agent, according to a September 10, 2026 report from TestingCatalog. The leak lands two days after Muse’s public debut and just two weeks before Meta Connect, giving the company a narrow window to decide how much of this work it shows on stage. None of it is live yet, but the pieces on the table (on-demand voice cloning, hidden reasoning controls, and a new model codenamed “Watermelone”) point to a company trying to close the gap with OpenAI, Google, and Microsoft on voice-first AI.
Meta Prepares Custom Voices and Model Upgrades for Muse: What Leaked
The core of the leak is a voice feature that lets a user simply ask Muse for a specific sound and get one back. Per TestingCatalog’s testing, “asking Muse to create a robotic voice generated a new voice and added it directly to the voice library.” That library sits behind a dedicated voice button and a selector, mirroring the voice mode already shipped in the standalone Meta AI app but built specifically for Muse’s agent interface.
None of this is available to the public. Meta has not confirmed a ship date, a rollout region, or even whether the feature survives internal review intact. That matters because Muse itself only became a public, downloadable app on September 8, 2026, according to TechCrunch. Meta is iterating on the next feature set before the current one has finished its first week in the wild, which says something about how much pressure the company feels to keep pace with rivals shipping voice and agent features on a near-monthly cadence.
Inside the Avocado 5.14 and 5.16 Codenames
TestingCatalog’s report also surfaces two internal build references, Avocado 5.14 and Avocado 5.16 v0, that appear to map onto Meta’s public model timeline. The outlet’s own read is that “Avocado 14 corresponds to the existing Muse Spark 1.3 setup, while Avocado 16 could be a placeholder for the upcoming Muse Spark model update.” In plain terms, Meta numbers its shipped models one way externally and tracks them under fruit-themed codenames internally, and the gap between Avocado 14 and Avocado 16 hints at how close the next Muse Spark revision already is.
A third name, Watermelone, shows up as a separate track. TestingCatalog describes it as expected to arrive soon but says it will “likely no longer share the ‘Avocado’ codename,” suggesting Meta may be spinning up a distinct model family rather than just another point release. Nothing in the leak nails down what Watermelone actually does differently, only that it exists and that Meta appears to be retiring the Avocado branding for it.
Why Meta May Be Borrowing Microsoft’s Voice Tech
The most surprising detail in the leak has nothing to do with Meta’s own research stack. TestingCatalog found internal references pairing “Muse Spark with a Microsoft voice model,” specifically noting that “Microsoft currently offers MAI-Voice-2-Flash for low-latency speech and voice cloning.” If that holds up, Meta would be running its own Muse Spark language model for reasoning while leaning on a competitor’s audio stack for the voice layer, at least during testing.
That’s an unusual arrangement for a company that has spent three years building an in-house model family under Meta Superintelligence Labs, and one that already ships its own speech-to-text model for Muse Code and its AI glasses roadmap. It could mean Meta’s own voice generation research isn’t ready for a consumer-facing rollout on Muse’s timeline, or it could just be an engineering shortcut for an early internal build that gets swapped out before launch. Either way, it undercuts the idea that Meta’s AI stack is fully self-contained, and it puts a Microsoft product inside a Meta app if the pairing ships as tested.
Muse Spark 1.3: The Model Powering the Voice Push
Any voice or reasoning upgrade to Muse rides on top of Muse Spark 1.3, the model Meta shipped on September 2, 2026, six days before Muse’s own public launch. According to Artificial Analysis, the model comes in two configurations: an “xhigh” tier that scores 61 on the Intelligence Index, tying GPT-5.6 Sol at its max setting and Grok 4.6 at high, and a “max” tier that scores 62, trailing only Claude Opus 5’s 63.
Pricing is where Muse Spark 1.3 pulls ahead. Artificial Analysis reports the xhigh variant “costs $0.55 per Intelligence Index task,” well under GPT-5.6 Sol’s $0.95 and Grok 4.6’s $0.94 at comparable tiers, a gap the firm characterizes as a “70%+ premium” for the rival models. Token pricing held steady at $1.25 per million input tokens and $4.25 per million output tokens, with cached inputs priced at $0.15. The model keeps its 1 million token context window from the prior release and posted a 52% score on Tau3-Bench Banking at the max tier, the top result on that agentic benchmark as of publication.
| Model (tier) | Intelligence Index | Price per task | Context window |
|---|---|---|---|
| Claude Opus 5 (max) | 63 | Not disclosed by Artificial Analysis | Not disclosed by Artificial Analysis |
| Muse Spark 1.3 (max) | 62 | Not disclosed by Artificial Analysis | 1M tokens |
| Muse Spark 1.3 (xhigh) | 61 | $0.55 | 1M tokens |
| GPT-5.6 Sol (max) | 61 | $0.95 | Not disclosed by Artificial Analysis |
| Grok 4.6 (high) | 61 | $0.94 | Not disclosed by Artificial Analysis |
Muse Spark 1.3 also cut its own operating overhead. The company’s own materials, summarized by explainx.ai and Emergent, put the model at roughly 20% fewer tool calls and 25% fewer tokens burned per agentic task compared with Muse Spark 1.2, the release from earlier this year. Fewer tool calls per task is the kind of efficiency gain that makes voice features economically viable at scale, since an always-listening assistant multiplies the number of model calls a user triggers in a normal day.
Muse’s Rocky App Store Debut: 83,000 Downloads and a No. 2 Ranking
The custom voice leak arrives at an awkward moment for Muse’s numbers. TechCrunch, citing Sensor Tower data, reported that Muse pulled in more than 83,000 iOS downloads in the U.S. as of September 10, two days after launch, enough to reach No. 2 on the App Store’s Top Charts. That sounds strong until it’s placed next to Meta’s own launch history. The standalone Meta AI app drew 108,000 U.S. downloads on its debut day back in April, and Threads pulled in 4.3 million U.S. downloads on day one when it launched in 2023.
TechCrunch’s comparison to ChatGPT is the sharpest data point. OpenAI’s app reportedly hit “half a million installs” in the U.S. within its first week, an average of about 83,300 downloads a day. Muse needed roughly two days to match what ChatGPT did in a single day of its first week. Android adoption looks weaker still: Muse ranks No. 338 in the Productivity category on Google Play, well outside the charts iOS shows it climbing.
| App | Launch window | U.S. downloads | Source |
|---|---|---|---|
| Threads | Day 1 (2023 launch) | 4.3 million | TechCrunch / Sensor Tower |
| ChatGPT | First week (average per day) | ~83,300/day (~500,000 total) | TechCrunch / Sensor Tower |
| Meta AI app | Day 1 (April 2026) | 108,000 | TechCrunch / Sensor Tower |
| Muse | First 2 days (Sept 8-10, 2026) | 83,000+ | TechCrunch / Sensor Tower |
TechCrunch’s own reporting floats a plausible explanation beyond product quality: Meta is launching an agent that reads email, manages calendars, and touches payments just months after an $18 billion settlement tied to social media harms, and the company carries a long public record of privacy fines from the FTC and European regulators. Users handing a Meta product access to their inbox and bank-linked apps may simply be more hesitant than they were with a chat-only assistant.
Reasoning Controls: The Other Hidden Feature in the Leak
Custom voices aren’t the only thing sitting behind Meta’s internal flag system. TestingCatalog also found evidence of “additional controls over how Muse reasons, though these settings aren’t exposed in the public interface.” The framing suggests something closer to a reasoning-effort dial, similar to the tier system OpenAI and Anthropic already expose to developers through API parameters, rather than a consumer-facing toggle.
If Meta ships this as a user setting, it would let people trade response speed for depth on a per-query basis, useful for an agent that’s supposed to book restaurant reservations one minute and debug a spreadsheet formula the next. If it stays server-side, it more likely becomes a cost-control lever Meta uses to route cheaper, faster reasoning to simple requests and reserve the expensive tier for genuinely hard agentic tasks, similar to how Muse Spark 1.3’s xhigh and max tiers already split on price and score.
Historical Context: Meta’s Long Road From Chatbot to Agent
Muse didn’t appear out of nowhere. Meta’s AI assistant work traces back through the original Meta AI chatbot, then the Meta AI standalone app that climbed to No. 5 on the App Store after the first Muse Spark launch in April 2026, according to earlier TechCrunch coverage. Muse Spark 1.3, shipped September 2, is the fourth Muse Spark release in five months, a release cadence that outpaces most rivals and shows Meta treating model iteration as a monthly, not quarterly, exercise. That pace has also meant turnover among the researchers building it, following high-profile departures from Meta’s AI research ranks earlier this year.
The jump from a chat app to Muse, a full agent that operates inside a secure virtual machine and touches email, payments, health data, and smart home controls, is the bigger leap. Meta built that sandboxing specifically so the agent can click through third-party apps and websites on a user’s behalf without directly exposing credentials, a design choice that puts Meta in the same architectural camp as OpenAI’s and Anthropic’s own computer-use agents, both of which also run inside isolated environments for safety reasons.
Competitive Landscape: The Voice AI Race Across Big Tech
Meta is a late entrant to consumer voice AI. Amazon has shipped voice assistants through Alexa for over a decade. Google’s Gemini has voice built into Android and its own assistant surfaces. Microsoft, the company Meta appears to be borrowing voice technology from in this leak, already sells MAI-Voice-2-Flash as a developer-facing product for low-latency speech and voice cloning. OpenAI’s Advanced Voice Mode has been a flagship ChatGPT feature since 2024.
What makes Meta’s position interesting isn’t that it’s behind on voice tech generally. It’s that Meta is trying to bolt voice onto an agent product that’s still finding its footing, based on the download numbers TechCrunch reported. Rivals largely shipped voice features into assistants that already had an established user base. Meta is doing both at once: building the audience for Muse and building out its feature set in parallel, with Meta Connect on September 23-24 as the likely venue where some of this gets formally unveiled or quietly shelved.
Market Impact: What a Voice Push Means for Meta’s AI Strategy
Voice matters commercially because it changes how often people open an app. A text-only assistant competes for deliberate taps. A voice assistant that can be summoned hands-free competes for ambient attention, the kind of usage pattern that drives daily active use rather than occasional checks. If Meta can get custom voice generation right, and pair it with an agent that already reasons across a user’s connected apps, it has a shot at making Muse sticky in a way a chatbot alone rarely achieves.
There’s a cost side too. Muse Spark 1.3’s efficiency gains, the 20% cut in tool calls and 25% cut in tokens per task cited by explainx.ai and Emergent, aren’t cosmetic. An agent that fields voice queries throughout the day generates far more model calls than one used for occasional text chats, so every percentage point Meta shaves off per-task cost compounds fast at scale. Meta’s pricing on the xhigh tier, at $0.55 per Intelligence Index task versus $0.94 to $0.95 for GPT-5.6 Sol and Grok 4.6, gives it room to run a voice-heavy product without the margin pressure a pricier model would create.
Privacy and Trust Questions Raised by Custom Voice Generation
On-demand voice cloning inside a mainstream consumer app raises the obvious question of misuse. A feature that lets a user type a description and get back a synthetic voice sits close to technology that’s already been used for scam calls and impersonation elsewhere in the industry. Meta hasn’t published any policy details on how it plans to gate custom voice creation, verify who’s generating a voice, or prevent someone from cloning a real person’s voice without consent, because the feature hasn’t shipped and no public documentation exists yet.
That gap matters more for Meta than it would for a smaller company, given the $18 billion settlement and the FTC history TechCrunch flagged in its coverage of Muse’s download numbers. Regulators and privacy researchers will likely scrutinize any voice-cloning feature Meta ships far more closely than a similar feature from a company without that track record, and Meta’s early download numbers suggest at least some of the public is already applying that same scrutiny before the feature even exists.
What Meta Connect Could Confirm on September 23-24
Meta Connect, scheduled for September 23-24, 2026, falls less than two weeks after the TestingCatalog leak, and it’s the company’s usual venue for major AI and hardware announcements. Given that Muse only became public on September 8, and Muse Spark 1.3 shipped September 2, the timing lines up for Meta to use Connect as the stage where custom voices, the Watermelone model, or the reasoning controls get an official reveal, assuming any of them clear internal testing in time.
It’s just as plausible Meta says nothing about these specific features at Connect and instead focuses on AI glasses hardware, where voice interaction is arguably more central to the product than it is on a phone screen. A custom voice library built for a phone-based agent app and a voice system built for always-on smart glasses aren’t necessarily the same engineering effort, even if they share underlying models.
Predictions: Where Muse’s Voice Features Go From Here
- Meta ships a limited custom voice beta to a small U.S. cohort within the next two to three months, gated behind the same invite-code system TestingCatalog previously reported for Muse’s own closed alpha.
- The Microsoft voice model pairing gets replaced by an in-house Meta Superintelligence Labs voice model before general availability, since shipping a rival’s technology inside a flagship consumer product is an unusual long-term bet for a company Meta’s size.
- Reasoning controls launch as a developer-facing API parameter on the Meta Model API before they ever reach a consumer toggle in the Muse app, following the same pattern Muse Spark 1.3’s tiered pricing already established.
- Muse’s download numbers improve meaningfully once voice features land, since ambient, hands-free interaction tends to drive habitual use more than text-only agent features do.
- Watermelone arrives as a distinct model announcement, likely timed to Meta Connect or the following quarter, rather than folded quietly into a Muse Spark point release.
Frequently Asked Questions
Is Meta’s custom voice feature for Muse available yet?
No. TestingCatalog’s September 10, 2026 report describes the feature as internal and disabled for end users, with no confirmed ship date from Meta.
What are Avocado 5.14 and Avocado 5.16 in the Muse leak?
They are internal Meta build codenames. Per TestingCatalog, Avocado 14 lines up with the existing Muse Spark 1.3 configuration, while Avocado 16 appears to be a placeholder for a future Muse Spark update.
Is Meta using Microsoft’s voice technology for Muse?
TestingCatalog found internal references pairing Muse Spark with a Microsoft voice model, specifically Microsoft’s MAI-Voice-2-Flash. This is unconfirmed by Meta and appears limited to internal testing at this stage.
How many people have downloaded Muse since launch?
Muse topped 83,000 U.S. iOS downloads in its first two days, reaching No. 2 on the App Store’s Top Charts, according to Sensor Tower data reported by TechCrunch on September 10, 2026.
What is Muse Spark 1.3 and when did it launch?
Muse Spark 1.3 is Meta’s latest agentic model, released September 2, 2026. It scores up to 62 on the Intelligence Index according to Artificial Analysis, with roughly 20% fewer tool calls and 25% fewer tokens per task than Muse Spark 1.2.
What is the Watermelone model Meta is reportedly building?
Watermelone is an internal codename TestingCatalog surfaced for a model expected soon that will likely drop the “Avocado” naming scheme. Meta has not confirmed details publicly.
Will Meta announce these Muse features at Meta Connect 2026?
Meta Connect runs September 23-24, 2026, close enough to this leak that some announcement is plausible, but Meta has not confirmed any Muse voice or reasoning-control reveal for the event.
How does Muse Spark 1.3 pricing compare to GPT-5.6 and Grok 4.6?
Muse Spark 1.3’s xhigh tier costs $0.55 per Intelligence Index task, compared with $0.95 for GPT-5.6 Sol and $0.94 for Grok 4.6 at comparable tiers, according to Artificial Analysis.
