Launa-Omni Salesforce voice model vs Agentforce Voice

Verdict first: what Launa-Omni is and what to do this month

Launa-Omni is Salesforce AI Research's real-time voice model, published on October 19, 2026, according to Salesforce (2026), and it pairs a text model adapted for audio with a separate in-house speech generator. This month it is a model to evaluate, not a product to buy, because the launch post publishes no price, no general release date, and no API access.

For a small B2B team the real choice is which route to take: a voice-agent product that Salesforce already prices, a raw speech API that your engineers would run, or waiting until Launa-Omni has a price and access. For most teams already on Salesforce, the sensible first step is to price the shipped option, write a test plan for Launa-Omni, and leave live call flows alone until access is published.

TL;DR: Launa-Omni is promising, but its benchmarks are self-reported and it has no public buying path yet. Agentforce Voice is the route with public unit rates today, and Gemini Live is the route for teams that want to own the build.

Key Takeaways

• Launa-Omni is built for conversations in Salesforce and Slack, and the launch post publishes no price, release date, or API access.

• Its main design claim is a text reply formed before any speech, which lets a team read and correct what the system decided to say.

• On Audio MultiChallenge, its listed score sits less than one point below the top public entry, inside a wide error range.

• The route with public unit rates today is Agentforce Voice, which bills voice actions in Flex Credits.

• A Gemini Live build can look cheap per minute, but it moves turn handling, logging, escalation and access control onto your team.

• Act this month by writing the test you would run and the evidence that would change your mind, then revisit when access is published.

What launched on October 19, 2026, and what has not been published

Salesforce describes Launa-Omni as a model for real-time, two-way conversation, developed and trained for Salesforce and Slack while staying generalizable. Its core is a text model retrofitted for audio, so speech and language share a latent space, and a separate in-house text-to-speech model produces the spoken reply. Salesforce calls the model significantly smaller than other leading voice models, but the vendor has not published a parameter count or a size comparison.

The post names three design features. A real-time harness detects speech and decides when a person's turn is complete. The model can be customized and fine-tuned for named entities, dates, alphanumeric IDs, and industry jargon. It also writes a text response before anything is spoken, so a team can inspect and correct the reply. The post says Launa-Omni can sit behind the experience and hand deeper work to a frontier model, so the voice layer and the reasoning layer can be separate choices.

Slack users can already try a simplified version through Slackbot Voice Dictation, a speech-to-text feature. Salesforce also offers a subset of its models for research on Hugging Face, but the post does not say whether Launa-Omni is in that subset.

The vendor has not published this for pricing, a general release date, API access, or a parameter count. Read any product plan built on Launa-Omni as unconfirmed until those four items appear on a Salesforce page.

Benchmarks: the public boards against the launch claims

On Scale AI's Audio MultiChallenge board, Audio MultiChallenge: 59.69±4.13 for Launa-Omni is listed second, behind gemini-3.8-flash (high) at 60.40±4.49, according to Scale AI (2026).

A GitHub submission opened September 18, 2026 lists VoiceBench: 90.60% for Launa V1 according to GitHub (2026). The same submission describes the model as closed source for now and cites a "new fixed twelve-shot BBH evaluation" for the score. We do not compare the τ³-Voice result, because the launch post runs it on a custom harness and no board we opened carries it.

Signal vs Speculation: what the evidence supports and what it does not

Signal, checked in sources: The launch post's two headline benchmark scores match the independent entries we found, one on Scale's board and one in the GitHub submission. The model's architecture and design features come only from Salesforce's own post, so they are vendor claims until a buyer can test them.

Signal, naming: The same model appears under three names. Scale's board lists it as salesforce-launa-omni-1, the GitHub submission calls it Launa V1, and a Launa project page calls it Launa Omni v1. The sources we opened do not map these names publicly, so confirm which entry is which before quoting a score in a deck.

Speculation, ours: A 0.71-point gap inside a ±4.13 range is a tie on that one board, not evidence that Launa-Omni beats the top entry, according to Scale AI (2026). A single leaderboard does not establish how the model performs on your call types.

Speculation, ours: The post makes responsiveness claims but gives no latency figure, so any speed advantage over a cascaded setup or a raw speech API is unverified.

Speculation, ours: Whether Launa-Omni becomes a priced Salesforce product, and at what unit, is unknown, because the post names no path.

Not used: An earlier listing in search results gave a different VoiceBench score. We did not open that page, so this post does not use it.

Our view: For a buyer who must audit what a voice agent says, the text-before-speech design matters more than the leaderboard placement. That is an opinion about the design, not a measured result.

How we evaluated these options

We compared three routes: Launa-Omni as Salesforce's new model, Agentforce Voice as the Salesforce voice-agent product with public unit rates, and the Gemini Live API as the build route. The weights below are our judgment for a small B2B buyer, not vendor-published figures, and they sum to 100%.

Service Cloud Voice is a separate telephony and seat layer on the Salesforce side. Its seats are priced per user per month and tiered by included telephony minutes, so it adds to total cost even if you never use an AI voice action.

Feature matrix: what each route exposes

"Not described" in a cell means the pages we opened do not describe that feature. It does not mean the feature is missing.

Pricing and a worked example: what each route costs to try

Flex Credits are listed at $500 per 100,000 credits, and Agentforce Voice: 30 Flex Credits per voice action according to Salesforce (2026). At that list rate, one voice action costs $0.15, computed as 30 × $500 ÷ 100,000. The same page says voice usage types and multipliers differ and refers to a separate rate card, and it describes its figures as information only and subject to change.

Gemini 3.8 Live audio: $3.00 input and $12.00 output per 1M tokens according to Google (2026). The same Google page also lists per-minute rates for the same audio, which appear in the table below.

Pricing checked October 20, 2026.

Here is an illustrative case, not a customer: assume a team runs 1,000 calls a month, with 4 voice actions and 6 minutes per call. On Agentforce Voice that is 4,000 actions × 30 credits = 120,000 credits, or $600 a month at the $500-per-100,000 list rate. On the Gemini route, 6,000 minutes × ($0.005 + $0.018) = $138 a month in audio charges according to Google (2026), before text, telephony, hosting or engineering. The two totals are not like-for-like, because one is a platform action rate and the other is model audio only, so the $462 difference is not a saving until you price the engineering. On the build path, your code must read serverContent.turnComplete, which Google's Live API reference describes as the flag that marks the end of a model turn.

Vendor profiles: who each route fits

Launa-Omni: wait, evaluate, and ask for access

Best fit: teams already committed to Salesforce or Slack that want a voice model whose text reply is visible before speech, and that can wait for access. Limitations: no published price, release date or API, benchmarks that are self-reported or from a single board, and a GitHub submission that calls the model closed source for now. Implementation: there is no public build path yet, so prepare test cases now with named entities, dates, IDs and jargon, since the post says fine-tuning targets those. Evidence: Salesforce's launch post, plus the Scale board and GitHub entry above.

Agentforce Voice: the priced option on Salesforce

Best fit: Salesforce orgs that want voice actions with a published unit rate today. Limitations: the rate depends on usage types and multipliers that differ for voice actions, and the final rate depends on your contract. Implementation: budget in Flex Credits, which the pricing page says are shared across Actions, Prompts, Translations and Voice Actions, then confirm the contract rate with a sales representative. Service Cloud Voice seats are a separate line item if you need the telephony layer as well.

Gemini Live API: the build route

Best fit: teams with engineers who want to own the voice flow, logs and model choice. Limitations: the pricing page lists rates but not telephony, escalation or access control, and the Live API uses a stateful WebSocket connection that your team must run. Implementation: wire the audio stream, handle end-of-turn events, log each reply, and add an escalation path before any customer-facing use. Evidence: Google's pricing page and the Live API reference.

Where a proposed US Tech Automations workflow sits

A proposed, configurable US Tech Automations workflow would start on a trigger such as a completed-call event from whichever voice platform you use. The action pulls the transcript and the captured fields, then writes a note or a case update to Salesforce through its API. The output is a Salesforce record with the transcript and the detected name, date or ID, plus a task for a person when a value does not match the account. Prerequisites are API access to the voice platform or a transcript export, Salesforce API credentials with field-level permissions, and an agreed field map. Human review points come before any account, billing or order field changes, and whenever a captured identifier fails a lookup.

A second proposed workflow covers escalation. The trigger is a reply the voice agent flags as uncertain or outside its scope. The action creates a queue item that carries the text reply formed before speech, which is the step Launa-Omni's text-first design makes inspectable. The output is a routed task with the transcript attached. This step needs the voice platform to expose its text reply and a scope flag; if it does not, the step cannot be built as described. A supervisor should check the first batch of routed items before the rule is widened.

Build it yourself or configure it: the no-code contrast

Many teams would stitch this together in Zapier, Make or n8n, or build it in-house. Those tools can keep run histories, retries, error branches and audit evidence when configured, and a team that wires a Salesforce write-back in them can get much of that. What the buyer still owns is design and upkeep: idempotency so a retried call does not write a duplicate record, escalation rules, access controls on who can read transcripts, monitoring across the voice platform and Salesforce, and maintenance when either API changes. A proposed US Tech Automations design could configure those pieces as one workflow with run logs, retry rules, a human review queue and role-scoped access, but it still needs the same prerequisites: API credentials, a sandbox to test in, and a named owner who signs off on field mappings. For a similar Salesforce write-back pattern, see connecting ChurnZero to Salesforce and automating Gainsight to Salesforce.

Who this is for and a decision checklist

Who this is for: a small B2B software team that already runs Salesforce or Slack, needs voice to update records or route work, has one person who can own a pilot, and can wait for published access before committing money. Teams in a specific vertical can compare a Salesforce-side tool in Wealthbox vs Salesforce for financial advisors.

Red flags: you need a voice flow live this month; nobody can own transcripts, access rules and field mappings; or your procurement process needs a published unit price before any pilot starts.

Decision checklist:

• Write down the three questions to put to Salesforce: release date, unit price, and data handling for transcripts.

• Pick one call type and one record update to test, not the whole queue.

• Confirm where the text reply is logged and who can read it.

• Check the Agentforce unit rate against your contract, not only the list page.

• Set a date to revisit once access is published.

Questions buyers ask before acting

Is Launa-Omni available to buy now?

No public buying path is stated, because the launch post gives no price, general release date or API access, so treat it as not yet available to buy.

What does Launa-Omni cost?

The vendor has not published a price, so the pricing table lists it as Quote-based.

Is Launa-Omni better than Agentforce Voice?

This post does not establish that, because the two have no shared public test, and Agentforce's list rate measures cost, not answer quality.

Can a small team build on Gemini Live instead?

Yes, if engineers can own turn handling, logs and escalation; the rates are public, but the build and run cost is not.

Should we switch voice tools this month?

Not on the evidence in this post; set up a test you can run when access opens and keep the current flow in place.

When not to use US Tech Automations: a simpler tool wins when your voice agent already sits inside Agentforce and you only need its native actions, when a single FAQ bot with no record writes will do, or when one person can copy call notes into Salesforce each week with no volume pressure. In those cases, the built-in product or a plain template is easier to run than a configured workflow.

Bottom line for this month

Launa-Omni is worth a test plan now and a purchase decision later. Write down the call type, the record update and the evidence you need before access opens, and keep the current flow running until then. If you want a configured design to compare against, see how US Tech Automations configures this for a Salesforce write-back, including the access and review prerequisites listed above.