Murf AI Review 2026

Updated August 2026 · Every fact below was read from Murf’s own published documentation, terms, help centre and production pricing endpoints on August 11, 2026, with the source linked beside the claim.

Rows a review of ours normally carries — Tested, Model tested, Voice tested, Benchmark cost — are absent rather than filled with placeholders, because our audio benchmark is designed but not yet running.

Quick verdict

Recommended, with three named reservations — a qualitative, documentation-based verdict as of August 11, 2026. Murf AI is one of the more checkable vendors in this market, which is a real buying signal: a dated changelog, a clean-Markdown twin of every API documentation page, per-model character prices, published concurrency limits, and a commercial-use clause you can quote in a sentence. For a buyer justifying a tool to procurement, that beats a marketing adjective. The three reservations are equally concrete, and each is documented by Murf itself:

1. The control-versus-streaming split is a genuine product gap. Murf’s recommended migration target, Falcon 2 (API identifier falcon-2), is the path on which its own speech-customization features are documented as unavailable: “All the speech customization features will only be available on the Gen2 non-streaming Synthesize Speech endpoint” (Speech Customization, checked August 11, 2026). A buyer who needs studio-grade control and streaming has no single documented model.
2. Names and numbers are unreliable across Murf’s own surfaces. Five different strings circulate for one streaming model; two different annual Studio price sets ship in Murf’s own production assets; post-termination deletion is 30 days in one Murf document and 90 in another. All read August 11, 2026; the full list is in Cons.
3. Three things a self-serve buyer would reasonably assume are not true. Voice cloning is enterprise-only and sales-gated; the Studio free trial produces no downloadable file at all; the “we do not train on your data” commitment is scoped to enterprise customers. Each is documented below with its source.

Who it is best for, and who should skip it

Murf AI fits you if

• You produce voiceover as a document, not as an API call. Murf Studio’s script editor is built from blocks (paragraphs) and sub-blocks (sentences), each with its own voice actor, style, speed and pitch (help.murf.ai, checked August 11, 2026), and re-voicing text you have already generated costs nothing extra.
• You need a commercial-rights answer you can paste into a contract. Terms of Service §5.1 is quoted in full in licensing, and it expressly permits making output available to third parties.
• You are building a real-time voice agent and want a cheap per-character rate, and can evaluate the API before talking to sales — the trial is 100,000 characters with no time limit (API FAQ, checked August 11, 2026).

Skip Murf AI if

• You want self-serve voice cloning. Murf states plainly: “We do not currently provide self-served Voice Cloning solutions” (help.murf.ai, checked August 11, 2026). There is no API path and no self-serve tier.
• You evaluate tools by listening to their free output. The Studio free trial permits no downloads or exports of any kind: “The free trial does not support downloads” (help.murf.ai, checked August 11, 2026). It answers “should I subscribe?”, not “can I ship this?”.
• You stream from outside US-East and need parallelism. On both the free tier and pay-as-you-go, Falcon 2 concurrency outside US-East is documented as 2 (rate limits, checked August 11, 2026).
• You need streaming plus Murf’s expression controls, or word-level timings on a real-time path. Pause tags, audioDuration, variation and wordDurations are all documented as non-streaming Gen2 only (API reference, checked August 11, 2026).
• You need a per-word Emphasis control on a solo plan, a Pronunciation Library, or comment-and-review workflow. Murf documents these as Business/Enterprise or Enterprise-only — see expression and control.
• You send long text in single API requests. “The API supports up to 3,000 characters per request” (API FAQ, checked August 11, 2026).
• Your scripts are confidential and you are buying self-serve. The no-training commitment we could find is in the enterprise Master Service Agreement, not in the self-serve terms — see model training.

Price, free plan and commercial rights, in brief

Three products, three billing units, and they do not convert into each other. Murf Studio sells hours of Voice Generation Time (VGT), $29 to $199 a month across four paid tiers. The Murf API sells characters: “Gen 2 model: $0.03/1000 characters” and “Falcon model: $0.01/1000 characters” (help.murf.ai, checked August 11, 2026). Murf Dub sells credits at $1 each. Murf publishes no characters-to-minutes conversion anywhere we could find, which is why this page never prints a Studio cost per 1,000 characters or an API cost per minute. Full ladder and arithmetic in pricing and normalized cost.

• Free plans: Studio — 10 minutes of VGT, no expiry, no credit card, no downloads. API — 100,000 characters, one key, no time limit. Dub — 200 credits, watermarked downloads, one project. All checked August 11, 2026; conditions differ sharply (detail).
• Commercial rights: granted in Terms of Service §5.1 with no plan qualifier in the clause itself, while the help centre scopes the same grant to “all paid plans”. Both read August 11, 2026; stated, not resolved, in commercial use.
• Output ownership: no Murf document we read assigns ownership or copyright of generated audio to the customer, and the API version of the right is expressly “revocable” (Product Specific Terms §9.1.2(a), checked August 11, 2026). “Commercial rights” and “you own it” are not the same thing here.

Models and versions — the deprecation calendar

Murf’s model line is not one name: on August 11, 2026 its documentation, API enums and marketing pages describe what appears to be a single streaming model under five different strings, and one model carries a deprecation date five days out. Hence the rule for this page — no bare model names: every mention carries the exact string Murf uses, the API identifier where it differs, the surface it was read on, and the date.

The Falcon / Falcon 2 naming split, stated as a discrepancy

The API documentation says “Murf offers two text-to-speech models: Falcon 2 and Gen2” (API overview, checked August 11, 2026). The marketing pages say something else. Fetching the raw HTML of murf.ai, murf.ai/api and murf.ai/falcon on August 11, 2026 and counting strings: “Falcon 2” appears zero times on all three, while “Murf Falcon” appears 6, 9 and 18 times; and the marketing pages print 99.38% pronunciation accuracy (3, 6 and 2 occurrences) where the documentation prints 99.37%, a string appearing zero times on all three. Those are counts of published HTML on a stated date, reproducible by fetching the same URLs — not Murf’s claims and not test results. What they establish is narrow but real: Murf publishes no statement reconciling “Murf Falcon” with “Falcon 2”.

One further piece of evidence, and its limit. The old documentation URL murf.ai/api/docs/text-to-speech-models/falcon returns HTTP 308 and redirects to …/falcon-2 (verified with curl, August 11, 2026), while the November 2025 changelog entry still links to the old path. That establishes a rename of the documentation page. It does not establish whether the underlying model is the same one renamed or a new model that inherited the page, and we assert neither. The same split runs through the numbers: marketing quotes 130 ms time-to-first-audio and a per-minute price for “Murf Falcon”, the documentation ~100 ms and a per-character price for “Falcon 2”, both live the same day. We report both, attributed, and resolve neither.

Every model and parameter, by name and identifier

What breaks, and when

Murf’s changelog entry of July 20, 2026, headlined “Upcoming Gen2 Streaming Deprecation”, reads verbatim:

“Our legacy Gen2 streaming model is set to be deprecated on 16th August 2026. We recommend migrating to our latest Falcon 2 model promptly.

Gen2 non-streaming text-to-speech will remain available through the Synthesize Speech endpoint.”

— murf.ai/api/docs/changelog/2026/7/20, checked August 11, 2026.

That notice appears in eight places in three wordings — “We recommend using our latest Falcon 2 model for streaming instead”, “Please migrate to the latest Falcon 2 model”, and the “promptly” variant above. The date is identical in all eight; only the guidance verb varies. Scope is exactly as documented — the streaming path only; Gen2 on the non-streaming Synthesize Speech endpoint survives.

Three things Murf does not publish about this deprecation, none of which we will guess:

• What happens to a model: gen2 streaming request on 17 August 2026. Murf uses only the word “deprecated” — never “sunset”, “retire”, “end-of-life” or “removal” — and publishes nothing on hard failure, error codes, grace period or fallback.
• What the US-East streaming default becomes. The Stream Speech reference states “Defaults to falcon-2 for all the regions except US-East” (stream reference, checked August 11, 2026) — so a US-East caller omitting model today receives the very model being deprecated, and Murf does not say what that default becomes.
• Whether the older FALCON identifier still works. The January 7, 2026 changelog says List Voices’ model parameter accepts “FALCON and GEN2”; the current reference says “Valid values: falcon-2, gen2”. No changelog announces the identifier change and no page says whether the old value is still accepted. The absence of a migration note is itself the finding.

Which endpoint accepts which model

1. Falcon 2 is not available on the non-streaming Synthesize Speech endpoint. modelVersion accepts only GEN2, and the Falcon 2 launch changelog says it plainly: “Falcon 2 is now available on: Streaming API, WebSocket API” (changelog 2026-06-16, checked August 11, 2026).
2. Gen2’s speech-customization features are not available on streaming. “All the speech customization features will only be available on the Gen2 non-streaming Synthesize Speech endpoint” (Speech Customization, checked August 11, 2026).

Put together: a buyer who needs both studio-grade control and streaming has no single documented Murf model, and Murf’s recommended migration target for the deprecation is the path that loses the controls. A documented trade-off, not a criticism of how anything sounds.

Where Murf does not document what changed

We will not infer an improvement from a version number. As of August 11, 2026, Murf publishes no statement of what changed for:

• Earlier models → Gen2. The only text is “Outputs from GEN2 sound more natural and high-quality compared to earlier models.” Those models are never named, dated or described; no “Gen 1” appears in any Murf document we read; and Gen2 has no dated launch entry, the changelog’s oldest being February 13, 2025.
• Falcon (Beta) → Falcon 2. Four adjectives, no numbers, no benchmark comparing the two.
• The FALCON → falcon-2 identifier change, and the deprecation dates of four response fields and three List Voices fields — none in any changelog entry.

Two figures did move between the two dated Falcon documents, and this is as far as the documentation lets anyone go: the November 6, 2025 changelog claimed “Sub-130ms” time-to-first-audio, the Falcon 2 page claims “~100ms” / “under 100ms”, and pronunciation accuracy is 99.37% in both — unchanged. Murf’s published figures on Murf’s published dates; neither is a measurement, and we make no claim about whether Falcon 2 sounds better than what preceded it.

What this review covers, and what it is based on

This review assesses the two text-to-speech models Murf documents as current on August 11, 2026 — Falcon 2 (API identifier falcon-2, streaming and WebSockets only) and Gen2 (GEN2 on the non-streaming endpoint, gen2 on streaming) — together with Murf Studio, the Murf API, Murf Dub and Murf’s voice cloning offering.

What it is based on, and what it is not. Everything here comes from Murf’s published documentation, terms, help centre and production pricing endpoints, read on August 11, 2026 and listed in full below. Our standardized audio benchmark — the same short prompts generated under identical conditions for every provider — is designed but not yet running, so no audio was generated, no account or API key was held and no voice was selected. Nothing about how Murf sounds is asserted anywhere below; there are no audio players and no placeholders for them; and no benchmark cost figure exists, because nothing was spent. Every dollar figure here is a list price Murf publishes, or arithmetic on those list prices with the formula shown. Where a voice identifier appears — en-US-natalie, for instance — it is quoted from Murf’s documentation as an example of identifier format, never as something we heard.

Assessment by category

A qualitative, documentation-based assessment, category by category, with each judgment’s basis stated so you can disagree with it on the evidence. No numeric score appears, because score bands are only finalized against real measurements and precise-looking numbers without them are false precision (methodology).

Voice quality and naturalness

We have not listened to a Murf voice. Everything below is a vendor claim, attributed and dated — not a finding of ours.

• Of Falcon 2: “an ultra-fast, scalable, and reliable speech synthesis model built for real-time conversational AI… delivers natural, human-like speech with ~100ms time-to-first-audio” (falcon-2 model page, checked August 11, 2026).
• Of Gen2, only that “Outputs from GEN2 sound more natural and high-quality compared to earlier models”, no earlier model named (generate reference, checked August 11, 2026). The Gen2 page also states “native 44.1kHz output”; no equivalent native-sample-rate statement exists for Falcon 2, whose documentation publishes only a request default of 24000 Hz — not a native model rate, and not reported here as one.
• Murf’s homepage groups the family as “models, Falcon and Gen2… built on ethically sourced speech and deep linguistic modeling” (murf.ai, checked August 11, 2026) — again “Falcon”, not “Falcon 2”.
• Murf publishes its own competitive benchmark at murf.ai/falcon/benchmarks, comparing “Murf Falcon” against four named rivals on a “Voice Quality Metric”. It is a vendor-run comparison of the vendor’s own product, carries no publication date, and does not state which Falcon version was measured. We report its existence and will not repeat its rankings as fact.

What we can say without listening: Falcon 2’s documented style vocabulary is far narrower than Gen2’s (5 strings against 30 — counts here), which constrains the delivery available on the streaming path regardless of raw voice quality.

Pronunciation, numbers and difficult text

The accuracy figure, and the three versions of it. The API documentation states “With 99.37% accuracy, Falcon 2 ensures clarity across industries where precision matters” (falcon-2 page, checked August 11, 2026), and the same 99.37% appears on the Gen2 model page and in the November 2025 Falcon (Beta) changelog — one accuracy result attributed to three different model objects, which Murf never explains. Its marketing pages print 99.38%; its Gen 2 marketing page prints a third figure, “over 98.8% word level pronunciation accuracy”, from a different test. We quote all three with their sources and average none. The only method Murf describes anywhere is on that Gen 2 page: a blind evaluation across five languages drawn from 300,000 news sentences, 2,610 and 2,100 words assessed by two native speakers each.

What you can control when a word comes out wrong. Studio documents three mechanisms (help.murf.ai/pronunciation, checked August 11, 2026): Smart Suggestions, IPA-labelled alternatives previewable before applying; Custom IPA, entered directly — “if the IPA shown is /wund/, make sure to exclude the slashes and enter it simply as wund”; and Alternative Spelling, respelling phonetically, Murf’s example being “livv” for the verb “live”. A Pronunciation Library stores these in a root folder and applies them to every project inside it, re-rendering only affected blocks. Its help-centre heading reads “Pronunciation Library- Enterpirse Only” — Murf’s typo, reproduced as published — so standardizing pronunciations across projects is an Enterprise capability. On the API side Murf documents no IPA parameter at all; pronunciation control there is a Studio feature.

Expression, pacing and control

Murf’s control surface is genuinely deep — on one path. The table below is what Murf documents; we name no control Murf does not publish. As of August 11, 2026 there is no stability slider, no similarity setting, no emotion dial, no temperature and no style-exaggeration parameter anywhere in its documentation, and readers arriving from other vendors should not expect their equivalents.

The trade-off in one sentence. Six of the controls above are documented as Gen2-only, and the Gen2 streaming path is the one being deprecated on August 16, 2026 — so Murf’s own recommended destination, Falcon 2, is where those controls do not exist. If your product needs pause tags, duration targeting or per-word emphasis, you are on the non-streaming Synthesize Speech endpoint.

One control that is streaming-specific and genuinely useful: the WebSockets advanced settings document min_buffer_size — “sets the minimum number of characters required before sending input for audio generation”, range 40–160 characters, larger buffers described as improving audio quality and smaller ones reducing time-to-first-byte — plus max_buffer_delay_in_ms, range 0–1000 ms (advanced settings, checked August 11, 2026). An explicit quality-versus-latency knob with published ranges, which is more than most streaming TTS vendors offer.

Voice cloning

If you want self-serve voice cloning, Murf is the wrong product, and it says so itself. Three sources agree, all checked August 11, 2026: “We do not currently provide self-served Voice Cloning solutions” (help.murf.ai); “Currently, Voice Cloning is available only with our Enterprise plan” and “While we do support voice cloning, unfortunately, this is not a self-serve solution” (help.murf.ai); and the API documentation page for cloning contains no endpoint, request schema or code sample — only a “Contact Sales” card (API documentation).

What the process involves, where Murf’s two sources differ: recording volume is “less than 90 minutes” per the API documentation against “~1-2 hrs.” per the help centre, so the honest requirement is between 90 minutes and 2 hours — we state the range rather than splitting the difference. Turnaround is “1 to 4 weeks” in both, in a “professional studio environment”, delivered as 48 kHz 16-bit WAV. Cloning languages: five, enumerated — English, Spanish, Portuguese, French and German (help.murf.ai, checked August 11, 2026) — against a marketing page advertising “Clone voices in 20+ languages, dialects, and accents” without enumerating them. We treat the specific list as operative and report the marketing number as unenumerated.

Consent: a contractual promise, not a verified one

This is the part third-party write-ups get wrong routinely. Terms of Service §4.2.5 states, verbatim:

“The submission of a third party’s unauthorized voice recordings as training audio is prohibited, and you agree that you will not submit any recording to Murf that contains the voice of any speakers that have not explicitly consented in writing to have their Cloning Voice created and to have their voices synthesized.”

— Terms of Service §4.2.5, document dated December 19, 2024 (murf.ai/legal/terms-of-service, checked August 11, 2026).

That obligation is framed as a customer representation and warranty. No Murf document we read describes a consent-verification step performed by Murf — no consent form, no recorded consent statement, no identity check, no upload-time attestation — and its own five-step cloning process (share requirements, sign up and select a voice actor, record a custom script, AI processing, access the voice) contains no consent step. The accurate framing: Murf requires explicit written consent contractually; whether any verification occurs happens inside a sales-gated onboarding process Murf does not publish. We will not claim Murf verifies consent, and readers should discount sources that do. Editorial judgment, labelled as such: the sales-gated, one-to-four-week, studio-recording workflow is itself a substantial practical barrier to casual non-consensual cloning, far more than an instant self-serve clone button — our assessment of the workflow, not something Murf claims.

Where Murf contradicts itself, and it matters

Against §4.2.5’s prohibition, the FAQ block on Murf’s own voice-cloning marketing page reads, verbatim:

“Can I clone a celebrity voice? Yes. However it can be deemed as unethical as the use of a celebrity voice without their permission can lead to intellectual property infringement or even identity theft.”

— murf.ai/voice-cloning, checked August 11, 2026.

The contract governs; the marketing page does not. A reader who acted on that “Yes” would be in breach of §4.2.5, which prohibits submitting recordings of any speaker who has not explicitly consented in writing, and separately of §11.2.11, which forbids using the service to “Stalk, threaten, impersonate, dox, harass or defraud anyone”. Both documents were live the same day. We report the conflict rather than quietly citing whichever side reads better.

One more distinction that is easy to miss: “Consenting Speaker” is defined twice, differently. Terms of Service §4.2 defines one by the giving of explicit written consent; Product Specific Terms §9.2.1, governing Murf Dub, defines one as “a person whose likeness forms part of Your data”. The Dub regime then adds obligations the Studio cloning regime does not have: §9.2.2(b) requires “irrevocable consents… insofar as they allowed for learning by and within Murf’s AI model”, surviving termination, and §9.2.2(c) an irrevocable waiver of moral rights (Product Specific Terms, document dated December 3, 2024, checked August 11, 2026). If you use Murf Dub, the consent burden is materially heavier than the cloning pages describe.

Languages and accents

Murf publishes more different voice and language counts than any vendor we have checked. On August 11, 2026 its own pages claim voice catalogues of 120+, 130+, 150+, 200+ and 300+, language counts of 13+, 20, 20+, 33, 35, 35+, 39 and 40+, and style counts of 10+, 20 and 20+ — three of them on one help article, which opens with “over 200 voices in 40+ languages”, then lists “the following 33 languages and accents”, then says “300+ ultra-realistic voices across 33 languages” (help.murf.ai). Rather than pick one, we counted the voice-library tables Murf designates as authoritative, its own instruction being that voices and languages vary by model and the relevant model section governs (voice library, checked August 11, 2026).

The sharpest single finding those tables support: Murf’s recommended replacement for the streaming model it is deprecating on August 16, 2026 covers fewer base languages (20 against 33) and far fewer documented styles (5 against 30) than the model it replaces, by Murf’s own table on August 11, 2026. The thirteen base languages in the Gen2 table and absent from the Falcon 2 table are Bulgarian, Croatian, Czech, Greek, Indonesian, Malay, Romanian, Slovak, Swedish, Tagalog, Thai, Turkish and Vietnamese; no base language appears in the Falcon 2 table that is absent from Gen2. If your streaming product speaks one of those thirteen, the deprecation is a product problem, not a config change.

Two counts match the help centre exactly — 33 base languages and 39 language-accent sections — suggesting the “33 languages / 300+ voices” figures describe the Gen2 and Studio library while “150+ voices / 35 languages” describes the API surface. Murf never says so, and we offer it only as an observation about which numbers match which table. Two caveats before a multilingual rollout. The Falcon 2 page claims “150+ voices in 20+ expressive styles” while the Falcon 2 tables in that same documentation set contain 117 voice names and 5 style strings; we present claim and count side by side and assert neither that the table is incomplete nor that the claim is false. And Studio’s language list and the API voice library do not match: Studio includes Russian, Arabic, Danish, Norwegian, Finnish and Portuguese (Portugal), none appearing as a base-language section in the API voice library read the same day.

Multilingual behaviour, as documented. Murf markets “MultiNative” voices that “speak multiple languages with native accuracy” across all plans, and says Falcon 2 “voices can seamlessly switch between multiple languages within a single sentence”. The operative constraint is per voice — the locale you send must be supported by the voice you chose — and coverage varies enormously: in the Gen2 table en-US-natalie lists 12 locales plus “+27 more”, en-US-ariana only en-US. All checked August 11, 2026.

Editor, workflow and export

Murf’s Studio page is a JavaScript-only shell serving no content to a plain fetch, so everything here comes from its help centre, checked August 11, 2026. None of it is an impression of using the product.

The document model is the strength. A project is a script of blocks (paragraphs) containing sub-blocks (sentences), with voice actor, style, speed and pitch set per block and unlimited blocks per project — for scripted voiceover that beats a prompt box, because you revise one paragraph without touching the rest. The billing behaviour that follows matters before you buy: changing voice actor, style, pitch, speed, pause, emphasis, pronunciation, punctuation or volume for the same text consumes no voice generation time — re-voicing is free — but “When text is added or modified in a sub-block, speech is generated for the entire sub-block again”, so editing a sentence re-bills the whole sentence (help.murf.ai, checked August 11, 2026). Scripts that churn cost more than their word count implies.

The timeline is not a video editor. Three fixed tracks (Voice Over, Image and Video, Background Music), assets left-aligned with no gaps, and asymmetric trimming: voice blocks cannot be trimmed at all, images trim and extend, video and music trim but do not extend. Expect to finish elsewhere. Limits (help.murf.ai): 15,000 words per script upload, which also applies to pasted text (SRT exempt); audio-to-text capped at 45 minutes; no cap on media uploads. Transcription bills whole-file duration regardless of how much speech it contains, and the Voice Changer round-trip consumes both transcription time and Voice Generation Time. Export is paid-plan-only and broad: MP3, WAV, FLAC, a-LAW and µ-LAW at 8/24/48 kHz, 16-bit, mono or stereo; video as MP4 or MOV with subtitles burned in or as SRT/VTT. The Download Manager keeps only ten days of history.

Collaboration is heavily Enterprise-gated, and Murf contradicts itself about how heavily. Two help articles restrict Comments & Feedback and project sharing to Enterprise, while the free-trial article lists “Share project preview” as a trial feature. The plausible reading is that preview links are broad and workspace collaboration is Enterprise, but Murf does not say so, and we report the contradiction rather than resolving it. Free accounts get one workspace, paid accounts unlimited (checked August 11, 2026).

API, SDK and integrations

The API is the part of the product a reviewer can verify most thoroughly, because Murf publishes a clean-Markdown twin of every documentation page — append .md to any documentation URL — plus an index at llms.txt and an MCP server. That makes automated re-verification cheap: a genuine, if unglamorous, advantage over vendors whose specifications live only in rendered JavaScript.

Surface, checked August 11, 2026. Base URL https://api.murf.ai; authentication by an api-key header on REST and an api_key parameter on WebSockets (Murf’s inconsistency, not our transcription error). Endpoints: Synthesize Speech (POST /v1/speech/generate, Gen2 only), Stream Speech (POST /v1/speech/stream), WebSockets (wss://<region>.api.murf.ai/v1/speech/stream-input), List Voices, Generate Token, Voice Changer, Translate Text and Murf Dub jobs — the last requiring a separately issued API key. Output formats run MP3, WAV, FLAC, ALAW, ULAW, PCM and OGG at 8000–48000 Hz (default WAV 44100 Hz mono); streaming adds 16000 Hz and defaults to 24000 Hz on Falcon 2 against 44100 Hz on Gen2; Voice Changer drops PCM and OGG; ALAW and ULAW are mono-only at 8000 Hz.

Word timings are a real limitation on the real-time path. wordDurations (word, startMs, endMs) is returned by the non-streaming Synthesize Speech endpoint only; the streaming endpoint documents no timing field and WebSocket receive messages carry only audioOutput and finalOutput. Phoneme, viseme and alignment output appear nowhere — an absence in the documentation, not proof the capability does not exist. So captions and avatar lip-sync are a Gen2 non-streaming capability, unavailable on the Falcon 2 path Murf markets for voice agents. Determinism is undocumented too: searching Murf’s raw endpoint definitions for “seed”, “deterministic” and “reproduc” returned zero occurrences, and nothing states that identical input yields identical audio. We infer nothing the other way — variation implies deliberate variability, but behaviour at variation: 0 is undocumented. If reproducibility matters, that is a question for sales.

SDKs: one, for Python. pip install murf, repository murf-ai/murf-python-sdk, MIT licence, sync and async clients; PyPI package murf at version 2.3.0, uploaded March 5, 2026 (GitHub and PyPI, checked August 11, 2026). The API marketing page claims “Production-ready Python and Javascript SDKs”, but we could not locate a JavaScript or TypeScript SDK in the documentation, in the fifteen public repositories of its GitHub organisation, or on npm (no @murf-ai scope; the unscoped murf package was unpublished in 2020). Murf’s own Falcon page says only “Production-ready Python SDK”; the reference’s generated samples in cURL, Python, JavaScript, Go and Ruby are code samples, not SDKs.

Integrations: seven documented for developers — an MCP server, Zapier (text-to-speech only), Make (Synthesize Speech only), n8n (the only one exposing four operations), Pipecat and LiveKit (both defaulting to "falcon-2"), and Agora, whose sample configuration specifies "FALCON" — a fourth spelling of the same family in live circulation. Studio-side, ten are marketed and four documented: the integrations page lists Articulate 360, PowerPoint, Adobe Captivate and Captivate Classic, Adobe Audition, Google Slides, Canva, Notion, Webflow and WordPress with no stated plan requirement, while the help centre documents only the PowerPoint plugin, Windows Voices Installer, Canva and Mobile Web App. Where limits exist they are specific — the Canva app caps a voiceover block at 1,000 characters and 50 audio blocks. No browser extension is documented anywhere (all checked August 11, 2026).

Data residency and on-premise, both Falcon 2 only. Eleven regional WebSocket endpoints plus a Global Router; account data stays in the United States while “text input, audio synthesis, and processing will be confined to the selected regional environment”; and the trap is stated plainly — “do not use the global URL if you require strict residency”, because routing can select any available region. On-premise runs via a “Murf Local License Agent” and “Murf TTS Server”, with no price, licence fee or minimum term published (all checked August 11, 2026).

Latency and generation speed

We have measured nothing, so this section ranks nothing. Every number is Murf’s own, attributed and dated, and it is here because those numbers disagree with each other — which a buyer deserves to know before quoting one in a design document.

• API documentation, Falcon 2: “time-to-first-audio under 100ms” and “~100ms time-to-first-audio” across the overview, model and streaming pages; “Sub-100 ms latency performance” for on-premise.
• Marketing pages, “Murf Falcon”: “55ms model inference. 130ms end-to-end latency”, “130ms time-to-first-audio consistently across geographies”, and “Fastest 130 ms End-to-End Latency” on the homepage.

Two restrained observations. The marketing pages measure two different things — a 55 ms model latency and a ~130 ms time-to-first-audio — and only the second is comparable with the documentation’s figure. And the 130 ms matches the November 2025 Falcon (Beta) changelog’s “Sub-130ms”, consistent with marketing not having been updated for Falcon 2 — but consistency is not proof, and we do not assert it.

Where the latency documentation gives detail it is candid and genuinely useful: some languages “require additional preprocessing… which can add a 10-20 milliseconds to first-byte latency”, Hindi in Latin characters “adds 5–10 ms of overhead”, and “Persistent connections help you avoid 30 - 80 ms of extra overhead per request”. Murf also names the definitive metric: “the key metric to track is Time to First Audio Byte. This is the only metric that reflects true real-time performance” (latency documentation, checked August 11, 2026) — what our benchmark will report against when it runs, measured rather than quoted. One claim needs care: Falcon 2’s page says it “can support 10,000+ concurrent calls in parallel”, which Murf’s own rate-limits table shows is the Enterprise ceiling; on pay-as-you-go outside US-East the documented number is 2.

Pricing and normalized cost

How these prices were read

Murf’s public pricing page is client-rendered and returns no content to a plain fetch — on August 11, 2026 it served an identical 6,230-byte shell to a browser and to a crawler user agent, its only body text “You need to enable JavaScript to run this app.” Studio prices here therefore come from Murf’s own production pricing endpoint (murf.ai/Prod/ping/new-pricing-plans/true), the one that page’s JavaScript calls, plus its help centre; API prices from the help centre, API documentation and API pricing endpoint. Two consequences for a procurement reader: Murf runs pricing cohorts — its API pricing endpoint returns "pricingTestGroup":"B", so every API price here is a dated group-B observation and may not be what you are quoted — and because the page could not be rendered, we cannot say which of two conflicting annual price sets a buyer sees (below).

Normalized pricing, with the formula shown

The per-1,000-word figures use Murf’s own conversion and carry Murf’s own caveat, which we repeat rather than hide: “A 1,000-word English script would consume ~ 6 minutes of VGT without accounting for the text changes. (This is an approximate value and will vary based on the script, voice selected, and speed of the voiceover.)” Order-of-magnitude, not a quote.

The annual price conflict, and why we will not print “33% off”

The annual-saving row above is the sharpest reason to check your invoice: the endpoint the pricing page calls at runtime and that page’s own JavaScript bundle publish different annual prices for the same four plans, a ~20% saving against a ~33% one, with Murf’s help centre corroborating the bundle the endpoint does not serve. We could not render the page to see which a buyer is charged, so we publish the range and the uncertainty rather than picking a side — confirm the annual figure at checkout before committing.

Two further behaviours observed the same day affect what you are quoted rather than what is listed. Murf’s pricing endpoint serves an Indian-rupee plan set at exactly 83.0× the dollar figure on every tier — one flat conversion factor rather than regionally-adjusted pricing (we make no claim about whether that rate is favourable). And its pricing bundle carries a first-time-buyer coupon experiment on Creator Lite whose internal label, rendered badge and promo code disagree about the discount; no checkout was reached, so we publish no Murf discount percentage from it — only that such experiments run.

Overage, carry-forward and seats

There is no metered overage: nothing we read prices an extra minute of Studio Voice Generation Time, and the documented remedies are upgrade, repurchase or support — “If your voice generation time runs out before the end of your billing cycle, don’t worry. You can always upgrade your plan or repurchase it” (help.murf.ai, checked August 11, 2026). Unused time survives exactly one cycle: “Active and paused plan VGT is carried forward to the next billing cycle” but “will only roll over to the immediately following month”, and “If a subscription is canceled, any carried-forward VGT will be lost” — pausing preserves it, cancelling does not (help.murf.ai, checked August 11, 2026). The same rule applies annually, making the 24/48/96/240-hour annual pools use-it-or-lose-it with one period of grace, not a bankable balance. Seats: Creator tiers single-seat, Business up to 3 plus 5 view-only, Enterprise up to 100 plus 100 — but no per-seat price is published anywhere we read, so a growing team cannot model the cost of its fifth editor from Murf’s public materials.

Free plans: four offerings, four different sets of terms

“The Murf free trial” is not one thing, and the differences matter. All checked August 11, 2026.

• Studio trial. 10 minutes of voice generation time and 10 minutes of transcription, no credit card, no expiry date — a one-time credit that expires only when exhausted, and produces nothing you can keep: “The free trial does not support downloads”. Multiple accounts are a documented Terms of Service violation, so the obvious workaround is prohibited.
• API trial. 100,000 characters, one API key, concurrency 5, 1,000 requests per minute, valid until the credits are exhausted. Whether it requires a credit card is not stated on any Murf page we read, and the Studio trial’s “no credit card required” must not be assumed to carry over.
• Murf Dub trial. 200 credits, watermarked downloads, one project — roughly 100 minutes of automated dubbing in one language, and unlike the Studio trial it does let you download something.
• Startup Incubator Program. Application-gated, one-time: “$1500 free credits for a 3-month period”, equivalently 50 million characters, for companies under 10 years old with fewer than 100 employees and a demonstrable use case; resale disqualifies. Overage at $0.03 per 1,000 characters.

Which model the Studio trial may use is not documented. Murf gates Gen2 per voice and per workspace, never per plan, and no page states whether a trial workspace can enable it. Recorded as a gap rather than filled.

Commercial use, licensing and privacy

We quote rather than summarize here, because a paraphrase of a licence is worth nothing. Three documents govern, they differ materially, and every statement names which one it quotes.

One structural point that is easy to miss: all of Murf’s legal documents predate its entire current model line. None names Falcon, Falcon 2, Gen 2 or GEN2. The terms are model-agnostic, so there is no model-specific licence carve-out to find — and none can be inferred.

Commercial use, quoted in full

Terms of Service §5.1, verbatim and complete:

“You can use Murf created voices for commercial purposes. You can also make available to third parties who intend to use Murf created voices for commercial purposes. You however agree that Commercial Rights (defined below) are not resell rights and you cannot resell or offer to resell Murf’s offerings (including the library of voices available on Murf Studio) to any third-party. You further agree not to use Murf created voices for training any AI model or to synthesize Murf created voices in any way or for any purpose.

‘Commercial Rights’ means any use of Murf created voices for commercial or business purposes, except as restricted by this User Agreement.”

— Terms of Service §5.1, document dated December 19, 2024 (murf.ai/legal/terms-of-service, checked August 11, 2026).

One of the plainest commercial grants in this market, expressly permitting output to be made available to third parties — useful for an agency delivering to clients. Three things need care.

First, which plans it covers is unclear from Murf’s own materials. The clause contains no reference to plan, tier, subscription or payment — we read it in full to confirm that — but the help centre scopes the same grant differently: “Yes, all paid plans offer commercial rights over the voiceovers generated in the Studio” (help.murf.ai, checked August 11, 2026). Both Murf sources, both live the same day; we state the disagreement rather than harmonizing it. Second, the last sentence is oddly drafted: read literally, “not to use Murf created voices for training any AI model or to synthesize Murf created voices in any way or for any purpose” would prohibit the product’s core function. We flag the ambiguity and resolve it neither way; if it matters to your use case, get it clarified in writing. Third, the rights survive cancellation: “commercial rights and business license rights remain valid even after your subscription ends… We provide lifetime commercial rights and business usage rights for the voiceovers you create” (help.murf.ai, checked August 11, 2026).

Can you publish free-trial output? The honest answer

No Murf document we read grants commercial rights to free-trial output — the help-centre grant is scoped to “all paid plans” and the trial is not one. Equally, none expressly prohibits it: there is no “no commercial use on the free tier” clause. The closest operative language is Terms of Service §10.7, permitting trial use “only for User’s self-evaluation” — a restriction on the purpose of the trial, not a rule about its output. We extend neither §5.1’s tier-neutral grant nor the help centre’s paid-plan framing to settle it. In practice the question is nearly moot: the Studio trial produces no downloadable or exportable file at all, so there is no artefact to publish. Only a shared project preview and an embed link leave Murf.

Output ownership: what the terms do not say

The finding most competing write-ups get wrong. No Murf document we read assigns ownership or copyright of the generated audio to the customer. Terms of Service §3.1 addresses the input side only:

“You own the content and information that you submit to the Services (‘Content’). For the purpose of obtaining the Services, you are granting to Murf a limited, non-exclusive, non-transferable, worldwide, royalty-free and sub-licensable license to use, process and transmit the Content.”

— Terms of Service §3.1, checked August 11, 2026. “Content” is what you submit, not what Murf generates. We read the Terms end to end: §§1–13 contain no output-ownership grant, and the enterprise Master Service Agreement is the same — §14.1 “Reserved Rights” grants no output ownership and “Customer Data” again covers what the customer submits. On the API the right is not only a use right but an expressly revocable one. Product Specific Terms §9.1.2(a), verbatim:

“Subject to these API Terms, Murf grants you a limited, non-exclusive, non-transferable, non-sublicensable, worldwide, revocable license during the term of the User Agreement to use and make calls to the API to generate, access and distribute the Generated Content for the Use Cases that comply with this User Agreement.”

— Product Specific Terms §9.1.2(a), document dated December 3, 2024 (murf.ai/legal/product-specific-terms, checked August 11, 2026; the emphasis is ours, the word is Murf’s), reinforced by §9.1.3(c): “Failure to comply with these TOS can result in suspension and termination of your API and Generated Content access.”

The defensible statement: Murf grants broad commercial use rights over generated voices, but neither its self-serve Terms nor its Master Service Agreement assigns ownership or copyright in the generated audio to the customer, and the API version of that right is expressly revocable (checked August 11, 2026). Commercial rights are not the same as ownership. For most creators the distinction never bites; for anyone assigning rights downstream, registering a work or relying on exclusivity, it does. Two related points: §9.2.3 — Murf Dub only — states “Outputs provided to You may be similar or identical to outputs independently provided by Murf to others”; and the help centre says voice-cloning customers “own the rights to the custom voice and can also use it for commercial purposes”, a claim with no counterpart in any of the three contracts, which we attribute to the help centre and flag as unsupported.

The carve-outs, and who they apply to

• No resale of Murf’s offerings, including its voice library (Terms of Service §5.1 self-serve; Master Service Agreement §9.3(a) and (e) enterprise). A separate reseller programme is advertised but we did not read it — resale evidently exists only under an agreement we cannot see.
• No training any AI model on Murf-created voices (§5.1).
• No robocalls violating the U.S. Telephone Consumer Protection Act (§11.2.14); no impersonation, stalking or defrauding (§11.2.11); no illegal, obscene, defamatory or pornographic material (§11.2.12).
• No High Risk Activities — failure causing death, bodily injury or environmental damage, expressly including medical life-support, emergency response and air-traffic control (Master Service Agreement §9.2 and §21.8, enterprise only).
• Murf Dub outputs only: §9.2.4 forbids representing outputs “as being approved or vetted by Murf”, “as being an original work or a wholly human-generated work”, or using them “for automated decision-making that has legal or similarly significant effects on individuals, unless it does so with adequate human review”. Not a general text-to-speech obligation — a distinction other reviews collapse.
• Attribution: silence, reported as silence. No attribution or credit requirement on any plan, having checked §5.1, §§1–13, the API terms and the help centre’s commercial-rights articles (August 11, 2026). We report the silence; we do not upgrade it into a guarantee.

Redistribution, broadcast and platform risk

Redistribution to third parties is expressly permitted by §5.1. Broadcast is neither carved out nor expressly permitted — the terms are silent on it as a category, and we say so rather than manufacturing a clause. Before building an audiobook or YouTube business on Murf, read Terms of Service §6.3, which disclaims liability for, verbatim (the capitals are Murf’s):

“THIRD-PARTY POLICIES (SUCH AS YOUTUBE REFUSING TO ALLOW MONETIZATION OF VIDEOS WITH SYNTHESIZED VOICES OR ACX DECLINING TEXT TO SPEECH BASED AUDIO BOOKS)”

— Terms of Service §6.3, checked August 11, 2026: Murf names ACX and YouTube specifically as risks it will not carry. Its help centre is equally frank, providing “commercial licensing for the voiceovers created for videos on YouTube or any other streaming platform… as long as your content is original” while noting “the policies of YouTube and other third-party platforms regarding the use of AI voices are constantly changing, and we recommend that you check with them”.

Indemnity: a marketing assurance and a contract are not the same thing

Murf’s Ethical AI page states: “Coverage for Claims — Every Murf voice emerges from active collaboration and explicit consent from its creators. We’re dedicated to supporting our commercial customers against any copyright claims for using our voices” (murf.ai/ethical-ai, undated page, checked August 11, 2026). But the self-serve Terms contain no Murf-to-customer indemnity: §6.2 disclaims all warranties and §6.1 runs indemnity the other way, the customer indemnifying Murf. A real indemnity exists only in Master Service Agreement §16.1, and only for a “Provider-Covered Claim… that the Service, when used by Customer as authorized in this Agreement, infringes or misappropriates a third party’s copyrights”. Self-serve buyers get an assurance on a web page; enterprise buyers get a clause in a contract. If indemnity is why you are choosing a vendor, that difference is the whole decision.

Model training: the promise is enterprise-scoped

Murf commits contractually not to train on customer data — in its enterprise agreement only. Master Service Agreement §5.1, verbatim:

“Subject to this Agreement, Provider and its authorized third party service providers may copy, display, modify, and use Customer Data solely to provide and maintain the Service and Support under this Agreement (‘Use of Customer Data’). It is clarified Use of Customer Data expressly excludes training of Provider’s AI models or any other AI models used in providing the Service.”

— Master Service Agreement §5.1, document dated April 5, 2024, checked August 11, 2026.

Murf’s marketing scopes the same promise identically, and the qualifier is in the last five words: “You own your data — Any data you provide to us is used only for providing the service. We’re simply custodians of your data and do not use it for training our AI models in any way for our Enterprise customers” (Ethical AI page, checked August 11, 2026).

The self-serve Terms and Product Specific Terms contain no equivalent commitment. We read the Terms in full: §3.1’s licence to “use, process and transmit” carries no training carve-out, and §4.1’s confidentiality clause permits use “for providing the Services” without addressing training. Murf Dub runs the other way outright — §9.2.2(b) requires irrevocable consents “insofar as they allowed for learning by and within Murf’s AI model”, surviving termination. So: never state “Murf does not train on your data” unqualified. The commitment exists in the enterprise contract, is limited to enterprise customers on Murf’s own marketing page, has no self-serve counterpart, and is expressly inverted for Murf Dub data. If your scripts are confidential and you are buying self-serve, that is a question for sales before purchase. One real privacy lever does exist for API users on any plan: encodeAsBase64 on the Synthesize Speech endpoint returns audio inline and is documented as giving “zero retention of audio data on Murf’s servers” (changelog, July 29, 2025).

Retention, deletion and security claims

Murf’s two sources on post-termination deletion do not agree. The Privacy Policy (dated November 3, 2023) says Murf “permanently delete[s] account information, including any project, media or text information uploaded by you, within 30 days of account termination”; the Security & Trust page says a “grace period of up to 90 days” and “data is automatically deleted within 90 days”. The more recently dated official source would normally win, but the Security page carries no date at all, so that test fails and we publish both figures with their sources (both checked August 11, 2026). If a 30-day deletion guarantee is a compliance requirement, get it in writing.

The remaining security statements are Murf’s own, unaudited by us, reported as its claims (undated Security & Trust page, checked August 11, 2026): data stored and processed only in AWS US-East-2 (Ohio); AES-256 at rest, TLS 1.2 in transit; “ISO 42001, ISO 27001, SOC 2 Type II, CCPA and GDPR compliant” with a SOC 2 Type II audit completed “with no exceptions”; “no security incidents or data breaches till date”. Note the tension with the eleven-region Falcon 2 residency offering: the API documentation reconciles it (account metadata stays in the US, processing moves to the chosen region) while the Security page states the US-only position without that qualifier. For residency questions the API documentation is the more specific source.

Pros

• Unusually checkable for a vendor this size. A dated changelog, clean-Markdown twins of every API documentation page, published concurrency limits and per-model prices let you verify what you are buying without a sales call, and re-verify cheaply later.
• A commercial-use grant you can quote in one sentence, including express permission to make output available to third parties, and rights that survive cancellation (§5.1, quoted).
• Cheap, transparent API list pricing with a real free quota — and a 50-million-character startup grant for eligible companies.
• The deepest documented expression controls in this category — pause tags at 0.1–5 s, per-word Emphasis on a two-axis grid, IPA and alternative-spelling pronunciation, per-word locale switching, Variability levels, “Say It My Way” delivery mimicry. On the Gen2 non-streaming path, at least.
• A script editor built for revision: blocks and sub-blocks with per-block voice, style, speed and pitch, unlimited blocks per project, and free re-voicing of the same text.
• Serious enterprise plumbing where it exists: eleven regional endpoints with Falcon 2 data residency, on-premise deployment, a contractual no-training covenant and a documented copyright indemnity — all enterprise-scoped, and all labelled as such by Murf.
• Zero-retention option for API users on any plan via encodeAsBase64.

Cons

• Control and streaming are mutually exclusive. Every customization feature is documented as Gen2-non-streaming-only, and the Gen2 streaming path is deprecated on August 16, 2026 in favour of Falcon 2 — the path without them. No single documented model does both.
• The replacement model covers less ground than the model it replaces: 20 base languages against 33, 5 documented styles against 30, by our count of Murf’s own tables on August 11, 2026.
• Concurrency outside US-East is 2 on both the free tier and pay-as-you-go, while the Falcon 2 page advertises “10,000+ concurrent calls” without noting there that the figure is an enterprise ceiling.
• The free trial is not an evaluation tool. Ten minutes of generation you cannot download, from a product whose export is paid-gated: you cannot hear how Murf handles your script before paying.
• No self-serve voice cloning at any price; what exists is enterprise-only, sales-gated, one-to-four weeks, five documented languages.
• Word timings only on the non-real-time path, awkward precisely for the voice-agent and avatar use cases Falcon 2 is marketed at.
• Deprecation churn. In eight months Murf launched a streaming model in beta (November 2025), superseded it (June 2026), deprecated a locale parameter (March 2026) and dated a streaming sunset (announced July 2026, effective August 16, 2026) — and the older model’s documentation page now redirects to the newer one with no retirement notice. Integrating here means budgeting for maintenance.
• You cannot plan confidently against Murf’s published specifications without checking which page a number came from — the subject of the table below.

Where Murf’s own pages disagree with each other

Every row is Murf against Murf — two or more of the vendor’s own live pages, all read August 11, 2026, each a place where a buyer could take away the wrong number.

Small things add to the same picture: a heading spelled “Enterpirse”, a doubled word in “the voice model model”, “with undue delay” where “without” is plainly meant in the Security page’s deletion commitment, and one voice table listing “Conversation” and “Conversational”, “NewsCast” and “Newscast” as separate styles. Individually trivial; collectively a fair signal about how reliably a buyer can plan against these specifications. An observation about Murf’s documentation, not a claim about how its audio sounds.

Best use cases

• E-learning and corporate training narration. Long scripts revised repeatedly are what block-based editing and free re-voicing are built for, and the pronunciation tooling handles product names and acronyms. Business tiers add the seats and per-word Emphasis.
• Explainer and marketing video voiceover inside one tool — three-track timeline, video export with SRT/VTT subtitles, and the Canva, PowerPoint and Google Slides integrations cover script to file, provided you do not need non-linear-editor control.
• Real-time voice agents in US-East on a budget. Falcon 2 with WebSocket multiplexing, context IDs and a documented buffer-versus-latency knob is a credible agent backend — if 15 concurrent streams is enough and you need neither pause tags nor word timings.
• Agencies delivering audio to clients: §5.1’s express permission to make output available to third parties answers the question most agency licensing reviews get stuck on, and the rights survive cancellation.
• Regulated deployments with residency requirements — but only on Falcon 2, only on enterprise, and only avoiding the global routing URL Murf itself warns against for strict residency.

And the inverse, because a recommendation without one is not worth much: Murf is a poor fit for audiobook production (Murf itself names ACX declining text-to-speech audiobooks as a risk it will not carry), for anyone whose evaluation depends on hearing output before paying, and for any workflow needing self-serve voice cloning.

Alternatives

Organized by the reason you would leave. Prices checked August 11, 2026 on each vendor’s own pricing page.

• Because you need self-serve voice cloning. ElevenLabs lists Instant Voice Cloning and a commercial license on its $6/month Starter plan (elevenlabs.io/pricing); Cartesia includes instant cloning and a commercial-use licence on its $5/month Pro plan (cartesia.ai/pricing). Murf has none at any price.
• Because you must hear output before paying. Both publish free tiers producing downloadable audio, subject to their own licensing restrictions — check those before shipping. Murf’s Studio trial exports nothing.
• Because you need a lower per-character API price. Murf’s Falcon 2 rate is competitive but not the market floor; our flagship comparison normalizes ten providers’ list prices side by side, all checked August 11, 2026.
• Because you need streaming and fine-grained control — Murf’s structural gap, and the thing worth shopping hardest on. Look for a vendor whose real-time model exposes the same controls as its offline one, and verify that on the API reference rather than the marketing page.
• Because you need an unqualified no-training commitment on a self-serve plan. Murf’s exists only in its enterprise agreement; terms differ enormously across this market (flagship comparison, with dates).

Direct comparisons

Head-to-head pages for Murf AI are not published yet, and we do not link pages that do not exist. Until they are, the side-by-side view lives on our flagship comparison of ten AI voice generators, which puts Murf’s pricing, free tier, licensing and cloning position next to nine competitors, every figure checked August 11, 2026.

Final verdict

Recommended, with the three reservations named at the top — and one date circled. Murf AI is a well-built voiceover product with better documentation hygiene than most competitors and a commercial-use clause you can actually quote; for scripted narration, especially e-learning and corporate video, it is defensible on the documentation alone.

What stops it being unqualified is not quality — we have not listened, and will not pretend otherwise — but coherence. Its streaming model and its control surface sit on opposite sides of a line the company is actively moving: the recommended migration target loses the features, the styles and thirteen base languages the deprecated model has. Meanwhile marketing and documentation publish different names, latency figures, accuracy figures and prices for the same things. None of that makes the product bad; all of it makes it harder to plan around than it should be.

If you stream from Murf, the actionable item is dated: the legacy Gen2 streaming model is documented for deprecation on August 16, 2026. Set your model parameter explicitly — especially in US-East, where the default is currently the model being deprecated — before that date. Murf does not publish what happens afterwards.

This verdict is qualitative and rests entirely on Murf’s published documentation as of August 11, 2026. No numeric score is given, because we have measured nothing that would justify one. When our standardized benchmark runs, this page will be updated with real samples, real costs and a scored assessment.

Frequently asked questions

Sources and verification

Every changing fact on this page — price, model name, quota, limit, licence term, language claim — was read from a Murf-operated source on August 11, 2026, and carries that date beside it. Where Murf publishes a figure we could not verify independently, it is labelled as Murf’s claim. Where we counted something ourselves, it is labelled as a count of Murf’s published document on that date. No third-party blog, roundup or review answered any question on this page, and none was permitted to answer a licensing question.

Principal sources, all read August 11, 2026, beyond those linked beside individual claims. Legal: Terms of Service (dated December 19, 2024), Product Specific Terms (December 3, 2024), Master Service Agreement (April 5, 2024), Privacy Policy (November 3, 2023). API documentation: the overview, both model pages, the Synthesize Speech, Stream Speech and List Voices references, Speech Customization, rate limits, enterprise, FAQ, voice library, data residency and latency pages, plus the complete 15-entry changelog. Pricing: the production pricing endpoint, the API pricing endpoint and the help-centre pricing articles. Marketing and trust: murf.ai, murf.ai/api, murf.ai/falcon, voice cloning, Security & Trust, Ethical AI, startup program. Roughly forty further help-centre articles supplied workflow, trial, limits and integration facts.

Two limits on this page’s evidence. No Murf help-centre article carries a visible publication or revision date, which is why several of Murf’s internal disagreements cannot be resolved by recency and are published as ranges. And its pricing page renders only in JavaScript and served no content to any fetch we made, so we cannot say which annual price set or coupon a real buyer is shown.

What could not be verified, recorded here rather than filled in: whether the API free trial requires a credit card; which speech model the Studio free trial may use; what happens to model: gen2 streaming requests after August 16, 2026; whether the older FALCON identifier is still accepted; Falcon 2’s native sample rate; determinism or reproducibility of output; enterprise pricing, minimum commitments and on-premise licence fees; per-seat Studio pricing; and per-plan project limits for the current plan lineup. Each of these would require an account, an API key or a sales conversation, and we hold none of those.

Published August 11, 2026. Next scheduled recheck: on or immediately after August 16, 2026, to record what actually happened to the Gen2 streaming model — and monthly for pricing thereafter. If we get something wrong, it is fixed visibly under our corrections policy, and no vendor relationship influences what appears here (how we make money).