Real-time translation API prices in 2026: translated text costs cents, translated voice costs dollars

A real-time translation API costs between $0.12 and $2.40 per hour of audio at list price, and the gap is mostly about what comes out: translated text is cheap, translated speech is not. ElevenLabs' new Realtime Translation sits at the expensive end and is not self-serve at all - its API reference says enterprise workspaces only.

That single distinction decides most of the shortlist. If you need captions or a translated transcript, you are buying a speech-to-text model with a translation step bolted on, billed per hour. If you need the other person to hear a translated voice, you are buying transcription, translation and speech synthesis in one stream, and vendors price it per minute. Here is what each option charges, taken from vendor pricing pages on October 23, 2026.

What a real-time translation API actually charges you for

Three billing units are in play, and they are not comparable until you convert them. Per-hour-of-audio pricing is the clearest: you pay for the audio you send. Per-minute pricing looks smaller and usually is not - OpenAI's $0.034 per minute for gpt-realtime-translate works out to $2.04 an hour. Per-seat pricing, like DeepL's, caps you on minutes instead of charging for them.

Then there is the unit nobody reads for. AssemblyAI bills streaming on session duration rather than audio duration, according to its pricing page - an open WebSocket with nobody talking still costs money. If your product holds connections open across a whole support call or a whole conference session, that difference can be larger than the price difference between vendors. We wrote about this pattern more generally in comparing the billing unit rather than the headline number, and live audio is where it bites hardest.

The second thing to check is whether translation is included or metered separately. Soniox prices real-time streaming at $0.12 per hour and says it transcribes and translates across 60+ languages in the same API call at no extra cost. Speechmatics charges for the model and then for translation on top: real-time Melia 1 is $0.40 per hour and the translation add-on is $0.65 per hour, covering up to five target languages per job, per its pricing page. On Speechmatics' Enhanced model at $0.80 per hour, translated captions land at $1.45 an hour - roughly twelve times Soniox's list price for the same kind of output.

At 1,000 hours of audio a month, those list prices work out to about $120 for Soniox, $1,050 for Speechmatics on Melia 1 with translation, $2,040 for OpenAI's translation model and $2,400 for Palabra's pay-as-you-go rate. Speechmatics applies volume discounts automatically above 500 hours a month per speech-to-text type, so the published numbers are a ceiling for steady workloads, not a quote.

ElevenLabs Realtime Translation is enterprise-only right now

Eleven Realtime Translation is the one most people are reading about this month, and it is also the one most readers cannot buy. The API reference for the realtime translation WebSocket states plainly that it is currently for enterprise workspaces only, and the product page offers "Contact sales" and documentation rather than a price or a sign-up. Checked October 23, 2026.

What you get for that sales call, per the product page: more than 90 languages, output through the Eleven v4 Turbo model by default, automatic language detection for each caller, a choice from a library of over 17,500 voices or an instant or professional voice clone, translated text alongside interim and final transcripts for captions, and audio formats meant for telecom networks. It lists SOC 2, PCI DSS Level 1, ISO 27001, HIPAA and GDPR compliance plus a Zero Retention Mode, which is the part that matters if you are translating recorded customer calls.

Latency is tuned rather than quoted. Both the product page and the API reference describe an eagerness setting with three values: high starts translating soonest for the lowest latency, medium is the default, and low waits for more of the sentence to produce a better translation at the cost of delay. Neither page gives a millisecond figure, so if latency is your acceptance criterion you will have to measure it yourself under a trial.

One honest limitation worth reading before you promise it to anyone: the product page says the original speaker's voice is not carried over exactly, and that a clone only approximates it. Two-way conversation is also not a single switch - the page says bidirectional setups are handled at the network or product level, for example by connecting each caller to a dedicated endpoint instance. That is architecture work on your side, not a config flag.

Skip it if you want to ship this quarter without procurement, or if your volumes are small enough that an unpublished enterprise price is a bad trade. In our view, an API you cannot price is not a shortlist candidate until someone on your team has a quote in writing.

The self-serve options for translated voice

Two vendors will take a card today and give you translated audio out.

OpenAI gpt-realtime-translate price per minute

OpenAI lists gpt-realtime-translate under transcription models with the use case "live translation" at $0.034 per minute on its API pricing page. For comparison on the same page, live transcription without translation is $0.017 per minute for gpt-live-transcribe and gpt-realtime-whisper, and plain gpt-transcribe is $0.0045 per minute. So the translation model is about twice the cost of live transcription and more than seven times batch transcription. The model sits in the same account as everything else you may already be calling through the OpenAI API, which in our view is often the real reason teams pick it.

Palabra publishes both a per-minute rate and plans

Palabra's pricing page lists speech-to-speech at $0.04 per minute pay-as-you-go with $50 in free credits on sign-up, and 60+ languages on all plans. It also sells capacity plans - Starter at $60 a month for 3 hours, Pro at $200 for 10 hours, Team at $1,000 for 50 hours, with lower figures shown for annual billing. Those plan rates work out to more per hour than the pay-as-you-go price on the same page, which is unusual enough that we would confirm which rate applies to your billing option before committing.

Where AssemblyAI stops

AssemblyAI is a reasonable assumption for this job and, as of October 2026, the wrong one. Its pricing page lists translation as a pre-recorded add-on at $0.06 per hour covering 100+ target languages, and marks it pre-recorded only. Streaming is $0.15 per hour for Universal-Streaming English or Multilingual, with $50 in free credits and no card required - excellent for live captions in one language, not a live translation path. The same caution applies to any speech-to-text vendor whose translation lives in the batch product.

If you would rather not build the pipeline

Buying an API means assembling transcription, translation, synthesis and transport yourself. Teams that only need translated meetings or calls can buy further up the stack.

DeepL now bundles DeepL Voice minutes into its subscription plans: 30 minutes per user per month on the Individual, Team and Business plans, with Individual at €7.49 a month billed annually, Team at €24.99 per user per month and Business at €49.99, per its pricing page. Thirty minutes is a sampler rather than a workload, but it is the cheapest way to find out whether voice-to-voice translation is good enough for your language pair before anyone writes code.

If you do build it, the pieces are listed separately in the directory: real-time transport from a provider like LiveKit, low-latency speech synthesis from one like Cartesia Sonic, and a transcription-plus-translation model in front of both. Upfynd lists 145 products in AI Audio & Voice and 157 in LLMs & AI Infrastructure as of October 2026, which is where the rest of these components sit.

How we would choose

Our view, by what you are actually building:

• Live captions or subtitles in several languages. Buy translated text per hour. Soniox's published rate is the lowest we found and includes translation; Speechmatics costs more but is worth testing beside it, and it sells a free tier with two concurrent real-time sessions plus $100 in credits to test with.
• A product feature where users hear a translated voice. Start on OpenAI's per-minute translation model if your stack is already there, or Palabra if you want speech-to-speech as the whole product rather than one model in a larger API.
• Contact center or telecom deployment with compliance review. This is the case ElevenLabs built Realtime Translation for, and the enterprise gate stops mattering once procurement is involved anyway. Get the quote alongside a per-minute API bill at your real volume.
• Occasional multilingual meetings. Do not buy an API. Use bundled voice minutes in a tool your team already pays for, and revisit only if usage outgrows the allowance.

Prices and plan limits here were read on October 23, 2026 and change often, particularly on the newer voice models. Before you commit, re-check the vendor's pricing page and run your own language pair through a trial - none of these numbers tell you how a model handles your accents, your audio quality or your jargon. To compare the models and SDKs around them in one place, start with the AI Audio & Voice category, and if you build one of these pipelines into a product, you can list it on Upfynd for free.