Cheapest AI Voice APIs in 2026
Compare AI voice API cost per minute in 2026. Cartesia and ElevenLabs price differently; convert credits into real project spend before you pick.
The cheapest AI voice API in 2026 is not the lowest sticker price. Credits, concurrency, and cloning fees change the per-minute number. Convert each vendor into cost per finished minute for your script length before you lock a stack.
Per-1,000-character and per-1,000-credit prices are meaningless until you know how many minutes of finished voiceover a thousand units actually buys. Here's that conversion, shown, with the assumptions stated plainly so you can check them against your own script. Prices were read from each vendor's own pricing page on 8 October 2026: Deepgram, Cartesia and ElevenLabs.
The short answer
- Cheapest per minute: Deepgram, at roughly $0.014 per minute on Aura-1 and $0.027 on Aura-2 (pay as you go). Then ElevenLabs Flash / Turbo at about $0.036, Cartesia Sonic 3.6 on its Pro plan at about $0.045, and ElevenLabs v3 at about $0.072.
- A 10-minute narration: about $0.14 on Deepgram Aura-1, $0.27 on Aura-2, $0.36 on ElevenLabs Flash / Turbo, $0.45 on Cartesia Pro and $0.72 on ElevenLabs v3.
- A 100-video batch (1 min each): about $1.35 on Aura-1, $2.70 on Aura-2, $3.60 on Flash / Turbo, $4.50 on Cartesia Pro and $7.20 on ElevenLabs v3.
- Hume Octave 2 is off the table for new work: Hume ends access to its TTS API on 13 November 2026, so don't budget a new project on it.
- Cheapest isn't always right — see the decision rule at the bottom before you default to the lowest number.
The conversion, and the assumption behind it
Providers meter differently — characters, credits — so any per-minute figure requires an assumption about how much spoken audio a thousand units actually produces. This uses a standard voiceover pace: roughly 150 spoken words per minute, which comes out to about 900 characters including spaces — call it 900 units per minute as a clean planning number. That's a modeling assumption, not a number every vendor publishes (Cartesia says a minute of speech takes 750-800 credits, and ElevenLabs' own tables assume about 1,000 characters a minute), so treat the per-minute figures below as planning estimates and confirm the exact metering with each provider before committing a large batch.
The pricing table
| Provider | Price per 1,000 | Estimated cost per minute* | 10-min narration | 100 × 1-min batch |
|---|---|---|---|---|
| Deepgram Aura-1 (pay as you go) | $0.015 | ~$0.014 | ~$0.14 | ~$1.35 |
| Deepgram Aura-2 (pay as you go) | $0.030 | ~$0.027 | ~$0.27 | ~$2.70 |
| ElevenLabs Flash / Turbo | $0.04 | ~$0.036 | ~$0.36 | ~$3.60 |
| Cartesia Sonic 3.6 (Scale plan, $299 for 8M credits) | ~$0.037 | ~$0.034 | ~$0.34 | ~$3.36 |
| Cartesia Sonic 3.6 (Pro plan, $5 for 100K credits) | $0.05 | ~$0.045 | ~$0.45 | ~$4.50 |
| ElevenLabs v3 | $0.08 | ~$0.072 | ~$0.72 | ~$7.20 |
| Hume Octave 2 | API closes 13 November 2026 | Not a pick for new work | n/a | n/a |
*Estimated at ~900 units per minute of spoken narration. Actual cost scales with your real script length and pacing, not this average. Cartesia counts one credit per character. ElevenLabs v4 is also $0.08 per 1,000 characters at list price, with a promotional $0.022 until 12 October 2026.
What that actually means for two real jobs
A single 10-minute narration — an explainer, a course module, a long-form video voiceover — costs somewhere between about 14 cents on Deepgram Aura-1 and 72 cents on ElevenLabs v3. At this scale, the dollar difference is trivial either way; the decision should be made on quality and control needs, not price, because you're arguing over well under a dollar.
A 100-video batch, at one minute of narration each — the shape of a daily short-form posting schedule run for about three months — costs roughly $1.35 to $7.20 depending on the vendor. That's still not a large absolute number, but it's about a 5x spread between the cheapest and the dearest row, and it compounds at genuine scale: a team generating narration for 1,000 videos a month is looking at roughly $13.50 on Aura-1, $27 on Aura-2, $45 on Cartesia Pro or $72 on ElevenLabs v3 — the gap that actually shows up on an invoice.
Plans change this math for steady, moderate usage. ElevenLabs' Free plan includes 10,000 characters a month and its paid plans (Starter $6, Creator $22, Pro $99 a month) include a monthly allowance, so if your narration volume fits inside it you're paying the flat subscription rate, not the per-1,000 rate used above. Cartesia works the same way: the Pro plan's $5 buys 100,000 credits a month, and overage on Pro is $65 per million credits. The overage math matters most once you're consistently exceeding plan limits or running bursty, unpredictable volume.
Why the cheaper option isn't automatically the right one
Cost per minute is one axis. It's not the only one, and defaulting to the cheapest API on every job is how a project ends up sounding wrong for the content even while coming in under budget.
The fuller decision rule, covered in more depth in the best text-to-speech API roundup: pick by latency need, expressiveness need, and language need first, and let price break the tie only when none of those bind. A script that needs the emotional range of ElevenLabs v4 or v3, or ElevenLabs' language and voice breadth, is a case where the cheaper option produces a worse result, not just a cheaper one — and a bad read that has to be redone erases the savings fast. Hume Octave 2 was the other emotional pick, but its API closes on 13 November 2026; what to use instead of Octave 2 covers the alternatives.
Budgeting it inside a real pipeline
If your voiceover is headed into a published video rather than a standalone audio file, the more relevant cost isn't the raw per-1,000 API rate anyway — it's the credit cost of the whole pipeline: script to voice to video to captions to publish. Versely's AI text-to-speech tool runs on the same credit-based model as its video and captioning tools — Cartesia Sonic 3.6 is 4 credits per 1,000 characters, and Gemini 3.8 Flash TTS and Inworld TTS 2 are 2 — so testing a script costs the same whether you're auditioning three takes or committing to a final one, which matters more for your actual budget than the underlying per-1,000 rate, since retakes are usually where content budgets quietly blow past estimates. For projects that also need a consistent cloned voice across a batch, AI voice cloning on Versely costs no credits to create (up to three saved clones per voice engine); you pay per 1,000 characters whenever the cloned voice speaks, like any other text-to-speech.
FAQ
What's the cheapest AI voice API in 2026? Of the APIs covered here, Deepgram: Aura-1 is $0.015 and Aura-2 is $0.030 per 1,000 characters pay-as-you-go. ElevenLabs Flash / Turbo is $0.04, Cartesia's Pro plan works out at $0.05 per 1,000 credits, and ElevenLabs v3 is $0.08.
How much does a 10-minute voiceover actually cost? Roughly $0.14 to $0.72 depending on the API, using a standard ~150-words-per-minute narration pace (about 900 characters a minute).
How much would voiceover for 100 short videos cost? At one minute of narration per video, roughly $1.35 on Deepgram Aura-1 up to $7.20 on ElevenLabs v3 — a real gap at volume, though a small one in absolute dollars at this scale.
Is Hume Octave 2 cheaper or more expensive than Cartesia? Neither: Hume is closing its TTS API on 13 November 2026, so it is no longer a pricing option for new projects.
Is a subscription or credit-based pricing cheaper overall? It depends on your volume pattern. Steady usage inside a plan's included allowance favors subscriptions like ElevenLabs' Creator ($22) and Pro ($99) plans; bursty or unpredictable volume favors per-unit pricing, since you're not paying for quiet months.
Should I always pick the cheapest TTS API? No. Price should break a tie between options that already meet your latency, expressiveness, and language needs — not override them. A cheaper voice that's wrong for the script costs more once you count the redo.
The takeaway
On list price, Deepgram is the honest answer to "cheapest AI voice API in 2026", with ElevenLabs Flash / Turbo and Cartesia close behind and ElevenLabs v3 the dearest of the set. But the numbers above are planning estimates built on a stated assumption, not invoiced totals — run your actual script through the finalists before committing a large batch, and check the pick-by-constraint rule before assuming the cheapest option is also the right one for the job.
Versely AI Weekly
New AI models, every Thursday.
The image, video, audio and LLM models released that week, what each is good for, and what's new on Versely.