Best Speech-to-Text APIs for AI Agents in 2026

Voice & Transcription APIs · Updated 2026-11-03

How this comparison works

Transcription vendors meter differently — per minute, per hour, per credit — so the table shows each vendor's own published unit honestly. Two traps to avoid. First, headline rates often belong to the cheapest model in the lineup, not the most accurate one: AssemblyAI's $0.15/hour is Universal-2; its flagship Universal-3.5 Pro is $0.21/hour. Second, streaming and pre-recorded audio are usually priced differently — AssemblyAI's streaming runs $0.45/hour against $0.15-$0.21 for files. If your agent listens live, budget from the streaming column, not the batch column.

Concurrency limits matter as much as price for voice agents: a published cap on simultaneous streams is a hard ceiling on simultaneous conversations. Deepgram, Gladia, and Speechmatics all publish theirs, and that transparency is itself a selection criterion.

The six APIs compared

Prices verified 2026-11-03 against vendor pricing pages. ElevenLabs figures are from its published pricing as carried by multiple sources — its site blocked automated reading during verification, which the dataset notes.

Deepgram — the real-time voice stack

Deepgram is the specialist: speech-to-text (Nova-3 and the Flux conversational models), text-to-speech, and a voice-agent API under one roof. Pay-as-you-go transcription starts at $0.0077 per minute for Nova-3 Monolingual — $0.46 an hour, mid-pack on price — and the starter credit is the most generous in this list: $200 free, no card. The published concurrency (150 simultaneous streaming connections on pay-as-you-go, 225 on Growth) tells you Deepgram expects production voice traffic, and its MCP support is real: the Deepgram CLI ships a built-in MCP server. The catch is the Growth tier's shape — prepaid, $4,000+/year — so the jump from hobby to committed volume is a step, not a slope.

Best for: live voice agents and call-center workloads where streaming latency and concurrency are the product.

AssemblyAI — transcription plus understanding

AssemblyAI's pitch is that the transcript is the beginning, not the end: speaker diarization, entity detection, summarization, and an LLM gateway over spoken data. Pricing is pay-as-you-go per hour, billed per second, from $0.15/hour (Universal-2), with $0.21/hour for the flagship Universal-3.5 Pro and $0.45/hour for real-time streaming; new accounts start with $50 of credit. Streaming limits are published (5 new streams/minute free, 100/minute paid). One clean, verifiable fact for budget planning: AssemblyAI's own FAQ states it does not offer an affiliate program, so nobody recommending it is earning a commission on it.

Best for: agents that analyze recorded audio — meetings, calls, podcasts — rather than talk live.

ElevenLabs Scribe — the two-way voice platform

Most people arrive at ElevenLabs for text-to-speech and stay when they notice Scribe, its speech-to-text model, at $0.22 per hour on the API pricing. The platform entry is the $6/month Starter plan (30,000 credits, commercial license), and the free tier carries 10,000 credits a month without a commercial license. For an agent that both listens and speaks, one vendor, one bill, and one credit pool is a real operational simplification. ElevenLabs also runs one of the few published affiliate programs in this category — 22% of a new subscriber's payments for their first 12 months via PartnerStack — which is worth knowing when you read glowing comparisons anywhere, including this site.

Best for: agents that need speech in and out, and creators already on the platform.

OpenAI Speech-to-Text — the default you already have

If your agent already calls OpenAI, its transcription models need no new vendor relationship: gpt-4o-mini-transcribe at an estimated $0.003/minute is among the cheapest published rates in this comparison, with gpt-4o-transcribe at an estimated $0.006/minute. One correction to common knowledge: the current official pricing page no longer lists the old whisper-1 line — the transcribe family has replaced it. The trade-offs are packaging: no published free tier on the API pricing page, rate limits governed by your general usage tier rather than a transcription-specific table, and no MCP server that calls the API — OpenAI's official MCP server is documentation-only.

Best for: teams already on OpenAI who need solid transcription without another vendor.

Gladia — multilingual, euro-priced, generous sandbox

Gladia (Solaria models) covers 100+ languages with async transcription at 0.16 EUR/hour and real-time at 0.68 EUR/hour, a 50 EUR starter credit good for a year, and clearly published self-serve limits (25 concurrent async jobs, 30 real-time sessions). Its official MCP server (@gladiaio/mcp) makes it a first-class tool for MCP-based agents. Euro pricing is a minor budgeting wrinkle for US buyers and a plus for European ones; either way, the rate card is public and simple.

Best for: multilingual transcription and European teams; agents built on MCP.

Speechmatics — the quiet price leader

Speechmatics lists the lowest entry rate in this comparison — Melia 1 at $0.12/hour for multilingual pre-recorded transcription. Its pricing-page headline says "from $0.24/hr" while the model table lists $0.12, so we record both rather than pick one. New accounts get $100 of free credit, limits are published (50 concurrent real-time sessions, 10 files/second pre-recorded), and it also does text-to-speech with the first million characters free. The gap: no official MCP server — only community ones — so MCP-native stacks need a wrapper.

Best for: cost-sensitive batch transcription across many languages.

Choosing by workload

• Live conversation (voice agents, calls): Deepgram or AssemblyAI streaming — check the streaming rate and concurrency cap, not the batch rate.
• Recorded media at volume: Speechmatics or AssemblyAI on per-hour price; Gladia if language coverage leads.
• Talks and listens: ElevenLabs, one credit pool both directions.
• Already paying OpenAI: its transcribe models are the path of least resistance.
• MCP-native stack: Deepgram, AssemblyAI, ElevenLabs, and Gladia ship official servers; Speechmatics and OpenAI do not (for API calls).

What the price does not tell you

Three costs hide behind the per-hour rate. Add-on features: diarization (who said what), entity detection, summarization, and redaction are bundled at some vendors and metered extras at others — an agent that needs speaker labels on every call should price the bundle, not the base rate. Minimums and steps: pay-as-you-go with no minimum (AssemblyAI, Deepgram, Gladia's wallet) scales down to hobby projects gracefully; prepaid enterprise-style tiers (Deepgram Growth at $4,000+/year) only make sense once volume is proven. The failure meter: streaming connections held open by a stuck agent bill by duration at several vendors, so client-side timeouts are a pricing feature, not just good hygiene. Finally, accuracy: vendors publish accuracy claims measured on their own chosen test sets, which is why this comparison quotes none of them. Every vendor here offers enough free credit to run your audio — accents, noise, domain vocabulary and all — through two or three finalists. An afternoon of that is worth more than any leaderboard, including ours.

FAQ

What is the cheapest speech-to-text API?

On published entry rates, Speechmatics' Melia 1 at $0.12/hour and AssemblyAI's Universal-2 at $0.15/hour lead for batch transcription, with Gladia at 0.16 EUR/hour close behind. OpenAI's gpt-4o-mini-transcribe works out near $0.18/hour at its estimated per-minute rate. The cheapest headline rate is not always the model you will actually run — flagship models cost more at every vendor.

Which APIs give free credit to start?

Deepgram ($200), Speechmatics ($100), AssemblyAI ($50), and Gladia (50 EUR) all publish starter credit, and ElevenLabs has a free monthly credit allotment. OpenAI publishes no free tier on its API pricing page.

Does it matter whether an STT API has an MCP server?

If your agent framework speaks MCP, an official server means transcription becomes a tool call with no integration code. Deepgram, AssemblyAI, ElevenLabs, and Gladia ship official ones. Without one, you call the REST API through your own wrapper — fine, just more code.

Why do streaming and batch prices differ so much?

Streaming reserves live capacity and returns partial results as audio arrives; batch processing can be scheduled and packed efficiently. AssemblyAI's published spread ($0.15-$0.21/hour batch vs $0.45/hour streaming) is typical. Budget voice agents from the streaming rate.

Which is best for non-English audio?

On published language coverage, Gladia (100+ languages) and Speechmatics (55+) lead this list, and both price multilingual models at their entry rates rather than as premium add-ons. Verify your specific language pair on the free credit before committing volume.