# 6 Smallest AI alternatives in 2026, before lightning_v3.1 retires

By Samuel Oram-Jones

Smallest AI retires its lightning_v3.1 model on 30 November 2026 at 18:30 UTC, tells teams to move to lightning_v3.1_pro before then, and names voice clones made on lightning-v3.1 among the things the retirement affects. A team that cloned a voice on the older model has to decide before that date whether to re-clone, and a re-clone is a decision about consent and recordings more than about a model name.

The shortlist for a phone agent is Bland Speech, Cartesia, Rime, Deepgram, Inworld, and Fish Audio, with every rival detail as of November 2026.

## Why do teams look for Smallest AI alternatives?

Teams look for Smallest AI alternatives when a voice agent needs latency, language, and cloning figures that agree from page to page, because Smallest AI says "sub 100ms latency across 70+ languages" and also states 31 languages and about 200 ms to first byte at 40 concurrent requests, and its model schedule moves on 30 November 2026.

Smallest AI sells a realtime voice AI suite for enterprise: voice agents, Lightning text to speech, Pulse speech to text, Hydra speech to speech in beta, and the Electron language model. The current text to speech model is Lightning v3.1 Pro, at about $0.195 per 10,000 characters, which is about $19.50 per million. Pay as you go starts with $10 in free credits, and Enterprise is custom with a 99.99% uptime SLA.

Smallest AI's own figures differ from place to place. Smallest AI states 70+ languages, 15 languages, 14+ languages, and 31 languages in different places. Smallest AI marks voice cloning No on pay as you go and Custom on Enterprise, while Smallest AI's instant clone endpoint takes a 5 to 15 second sample with one POST request. Smallest AI also names SOC 2 as an Enterprise feature and sells a HIPAA zero data retention add-on at $1,000 a month on pay as you go.

Lightning can be a fine model and still cost a buyer a first week of reconciling those figures.

## What should I check in a Smallest AI alternative?

A team comparing Smallest AI alternatives should check what each provider's latency figure measures, which phone formats come back, and what each provider's model schedule says, because the providers here publish those three answers differently.

Latency figures measure different things, and several providers publish more than one. Bland publishes 199ms to first audio on a streamed request, Cartesia says under 90ms, and Deepgram says as low as 80ms for Flux TTS. Rime says Coda v2 reaches first audio at 30 ms P50 at the server without network, and separately publishes 96 ms P50 for Coda with one request at a time. Inworld says as low as 50 ms for TTS-2 Flash on a third-party benchmark and puts TTS-2 at 100 ms P90, measured at its servers without network. Fish Audio states "<70 ms" as latency without labelling it time to first audio, and puts its s2-pro model at 100ms to first audio. A figure measured at the server leaves out the network, and a figure at 40 concurrent requests answers a different question than one at a single request.

Phone formats come down to 8 kHz mu-law, the 8-bit telephone encoding that Twilio uses. Bland Speech, Cartesia, Rime, Deepgram, and Inworld each document it. Fish Audio's text to speech API returns mp3, wav, pcm, and opus.

Model schedules show how a provider handles change. Cartesia set sunsets for sonic-2 and sonic-turbo on 20 October 2026 and says Sonic 3.6 is "fully backwards compatible" with Sonic 3.5. Rime retired Arcana on 15 August 2026 and retires Mist v1 and v2 on 3 December 2026. Fish Audio retires s1 on 31 December 2026, and its recommended model, s2.1-pro, serves the old s1 requests afterward.

## What are the best Smallest AI alternatives?

The best Smallest AI alternatives are Bland Speech, Cartesia, Rime, Deepgram, Inworld, and Fish Audio, with Bland Speech first because it returns 8 kHz mu-law audio, bills one standard per-character rate with no monthly plan, and ranked first of 13 AI voices in Design Arena's Audio Realism Bench results of 4 August 2026.

| Provider | First-audio figure, as the company states it | Published price | Phone audio |
| --- | --- | --- | --- |
| Bland Speech | 199ms to first audio on a streamed request | $0.04 per 1,000 characters, pay as you go | Mu-law at 8 kHz, PCM at 8 to 48 kHz |
| Cartesia | Under 90ms | Pro $5 a month for 100K credits; 1 credit per character | Mu-law and A-law at 8 kHz |
| Rime | 30 ms P50 for Coda v2 at the server | Coda v2 $0.050 a minute on Flex | G.711 mu-law at 8 kHz |
| Deepgram | As low as 80ms for Flux TTS | $0.0450 per 1,000 characters on Flux TTS | Mu-law and A-law at 8 kHz |
| Inworld | As low as 50 ms for TTS-2 Flash | $25 per million characters on Realtime TTS-2 | Mu-law and A-law at 8 kHz |
| Fish Audio | "<70 ms", not labelled as time to first audio | $15 per million UTF-8 bytes on s2.1-pro | mp3, wav, pcm, and opus |
| Smallest AI (for comparison) | "Sub 100ms"; about 200 ms at 40 concurrent requests in a benchmark | About $19.50 per million characters on Lightning v3.1 Pro | Mu-law, 8 kHz to 44.1 kHz |

<details><summary>What these rows mean</summary>

- **First-audio figure:** the number each company publishes, with its own definition, so rows aren't a race. Several companies publish a second figure as well.
- **Published price:** each provider's own rate with the plan or model named. Rime's minute converts at about 1,000 characters, Rime's own figure.
- **Phone audio:** the telephone-ready output each provider documents.

</details>

### 1. Bland Speech

Bland Speech is Bland's own text to speech model, built for phone calls and sold through a studio and an API with HTTP and WebSocket streaming. The WebSocket takes text as an LLM writes it, and a new context ID interrupts the turn that is playing. Bland publishes 199ms to first audio, and Bland Speech has a Pipecat service and a LiveKit Agents plugin in Python. More on the [text to speech API](/text-to-speech-api).

Bland Speech suits a team that wants phone-ready audio and a clone it can start itself: an instant clone takes about 10 seconds of audio, and a professional clone is fine-tuned on 30 minutes or more after Bland's consent and rights review. Bland Speech costs $0.04 per 1,000 characters, or $40 per million, pay as you go with no monthly plan, and a new account gets 50,000 free characters once.

<div class="cmp-proscons cmp-proscons-own"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>One standard per-character rate in the studio and the API, with no monthly plan</li><li>Mu-law output at 8 kHz for phone lines</li><li>A Pipecat service and a LiveKit plugin</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>More than a hobby project needs, because Bland Speech is built for production calls</li></ul></div></div>

### 2. Cartesia

Cartesia sells Sonic 3.6, released on 27 August 2026 in 44 languages, and says first audio arrives in under 90ms. Cartesia bills credits at 1 credit per character: Pro is $5 a month for 100K credits, Startup is $49 for 1.25M, and Scale is $299 for 8M. Cartesia returns mu-law and A-law at 8 kHz, and an instant clone needs 10 seconds of audio on Pro. Cartesia suits a team that wants a first-audio figure under 90ms as Cartesia states it and will manage a credit allowance.

<div class="cmp-proscons" data-rival="cartesia"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>44 languages on one model</li><li>Sonic 3.6 is backwards compatible with Sonic 3.5</li><li>SOC 2 Type II, with BAAs on Enterprise</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Speech draws on a monthly credit allowance, with overage at $65, $45, or $38 per million credits</li><li>Emotion tags work in English only, as a Beta feature</li></ul></div></div>

### 3. Rime

Rime sells conversational voice models for enterprise, aimed at production voice agents, IVR (interactive voice response phone menus), and telephony. Coda v2 costs $0.050 a minute and Mist v3 $0.030 a minute on Flex, and Free includes $50 of usage a month. Rime returns G.711 mu-law at 8 kHz, says it has been HIPAA compliant since February 2024 and SOC 2 Type 2 compliant since May 2025, and signs a BAA. Rime suits a team that wants telephony-focused models with compliance paperwork already in place and will pay by the minute.

<div class="cmp-proscons" data-rival="rime"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>SOC 2 Type 2 and HIPAA statements with a BAA</li><li>LiveKit, Pipecat, Daily, and Vapi guides</li><li>Dedicated VPC or on-premises for enterprise customers</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Mist v1 and v2 retire on 3 December 2026</li><li>Voice cloning is an enterprise clone ordered from Rime</li><li>Coda's cloud API covers 10 languages, with each voice speaking one</li></ul></div></div>

### 4. Deepgram

Deepgram sells text to speech for voice agents beside its speech to text. Flux TTS, Deepgram's current model, costs $0.0450 per 1,000 characters on pay as you go, and Aura-2, the previous generation, costs $0.030. Deepgram says Flux TTS starts audio in as low as 80ms, returns mu-law and A-law at 8 kHz, and ships English plus 10 languages with built-in text normalization. Deepgram says it is SOC 2 Type 1 and Type 2 certified and signs BAAs for Enterprise. Deepgram suits a team that wants speech to text and text to speech from one vendor, with self-hosted options.

<div class="cmp-proscons" data-rival="deepgram"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Speech to text and text to speech from one vendor</li><li>A $200 free credit with no card and no expiry</li><li>Self-hosted and private cloud options</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>SSML is stripped from input, with a warning</li><li>32 of Flux TTS's 42 languages are in Early Access</li></ul></div></div>

### 5. Inworld

Inworld sells realtime voice AI aimed at consumer apps, with Realtime TTS-2 at $25 per million characters and TTS-2 Flash at $15 on its On-Demand plan, which costs $0 a month. Instant cloning starts from 3 seconds of audio on every plan, and Inworld returns mu-law and A-law at 8 kHz. Inworld says it is SOC 2 Type II certified, HIPAA and GDPR compliant, and offers a BAA on Enterprise or as an add-on on its $1,500-a-month Growth plan. Inworld suits a team that wants plain-language direction of delivery and a low per-character rate.

<div class="cmp-proscons" data-rival="inworld"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>Instant cloning from 3 seconds on every plan</li><li>A commercial licence on the $0 On-Demand plan</li><li>Mu-law and A-law output at 8 kHz</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Inworld states 500+ languages in one place and 200+ in another</li><li>Professional cloning is a beta that runs on TTS-2 only</li></ul></div></div>

### 6. Fish Audio

Fish Audio is a voice AI platform with a web app, a REST API, and SDKs. Fish Audio's s2.1-pro model costs $15 per million UTF-8 bytes and supports 83 languages, an instant clone needs at least 10 seconds per clip, and Drama 3, a directable model launched on 13 October 2026, costs $50 per million bytes at list price. Fish Audio says a SOC 2 audit is underway. Fish Audio suits a team that wants a low per-byte rate and a directable model, and will convert audio for the phone itself.

<div class="cmp-proscons" data-rival="fish-audio"><div class="cmp-proscons-pros"><p><strong>Pros</strong></p><ul><li>$15 per million UTF-8 bytes on s2.1-pro</li><li>83 languages</li><li>A free API model for development</li></ul></div><div class="cmp-proscons-cons"><p><strong>Cons</strong></p><ul><li>Output formats are mp3, wav, pcm, and opus</li><li>The free model carries no time-to-first-audio guarantee</li></ul></div></div>

## Which Smallest AI alternative sounds most human?

Among the alternatives that were scored, Bland Speech v3 sounded most human, ranking first of 13 AI voices at 1365 in Design Arena's Audio Realism Bench results of 4 August 2026, while a human recording scored 1500.

In those results Cartesia's Sonic 3.5 ranked 6th at 1096, Smallest AI's Lightning v3.1 Pro ranked 11th at 952, and Inworld's TTS-1.5 Max ranked 13th at 763. Rime, Deepgram, and Fish Audio weren't in those results, so no score places any of them against Bland Speech v3. Inworld's Realtime TTS-2 Flash (9 August 2026), Cartesia's Sonic 3.6 (27 August 2026), and Fish Audio's Drama 3 (13 October 2026) were released after those results and aren't scored in them. The voices in Design Arena's [4 August 2026 article](https://notes.designarena.ai/audio-realism-benchmark-measuring-realism-in-audio-models/) were stock voices, one female and one male per company, chosen as its most naturally conversational American-English pair, so no cloned voice was scored.

## Which Smallest AI alternative should I pick?

Bland Speech is the Smallest AI alternative for phone calls and voice agents. A team whose clones were made on lightning_v3.1 faces a re-clone decision before 30 November 2026 wherever it goes, and Bland Speech v3, which placed first of 13 AI voices in Design Arena's results of 4 August 2026, makes an instant clone from about 10 seconds of clean audio.

Neither price nor latency makes the case for Bland Speech. At 1 million characters a month, published rates come to $15 on Fish Audio s2.1-pro, about $19.50 on Smallest AI Lightning v3.1 Pro, $25 on Inworld Realtime TTS-2, $30 on Deepgram Aura-2, $40 on Bland Speech, $45 on Deepgram Flux TTS, and about $50 on Rime Coda v2 on Flex by Rime's conversion of 1,000 characters a minute, with Cartesia's Startup plan at $49 a month for 1.25M credits. Smallest AI says sub 100ms and Bland publishes 199ms, so a team buying on speed or price alone has cheaper and faster-sounding options on this list.

[Bland Speech vs Smallest AI](/compare/bland-speech-vs-smallest-ai) goes deeper on the head-to-head, the [Smallest AI review](/blog/smallest-ai-review) covers Smallest AI on its own, and [Cartesia vs Smallest AI](/blog/cartesia-vs-smallest-ai) puts two of these providers side by side. The [alternatives guides](/compare/text-to-speech/alternatives) cover other providers, the [text to speech hub](/compare/text-to-speech) lists every comparison, and [Bland Speech](/speech) has a studio for a first test.

## Frequently asked questions

### What is the best Smallest AI alternative?

Bland Speech is the best Smallest AI alternative for phone calls and voice agents, because it bills $0.04 per 1,000 characters with no monthly plan, returns 8 kHz mu-law audio, and ranked first of 13 AI voices in Design Arena's results of 4 August 2026.

### What happens to Smallest AI clones when lightning_v3.1 retires?

Smallest AI retires lightning_v3.1 on 30 November 2026 at 18:30 UTC and names voice clones made on lightning-v3.1 as affected by the retirement. A team that wants the voice on another model or another provider clones it again from the original recordings, with the speaker's consent.

### Is there a cheaper Smallest AI alternative?

Fish Audio's s2.1-pro at $15 per million UTF-8 bytes costs less than Smallest AI's Lightning v3.1 Pro at about $19.50 per million characters, and Inworld's TTS-2 Flash at $15 per million characters equals Fish Audio's rate. Bland Speech costs $40 per million, and Deepgram's Aura-2 costs $30.

### Does Smallest AI have a free tier?

Smallest AI's pay as you go plan starts with $10 in free credits. Bland Speech gives a new account 50,000 free characters once, Deepgram gives a $200 credit with no expiry, and Rime's Free plan includes $50 of usage a month.

### Can a cloned voice move from Smallest AI to another provider?

No, a cloned voice stays with the provider that trained it, so a switch means cloning again from the original recordings, with the speaker's consent. [How to move a cloned voice to Bland Speech](/blog/move-cloned-voice-to-bland-speech) covers the steps for Bland.
