Text-to-speech API prices for agents

Snapshot 2026-10-07

Turning a sentence into spoken audio is a clean job for an agent: send text, get back an MP3. On the x402 marketplace it is sold per call, and the prices run from a tenth of a cent to a quarter of a dollar — a 250-fold spread for what looks like the same job. The spread is not noise. It tracks what sits under the wrapper: the cheap listings resell OpenAI’s low-cost voices or run a small open-weights model themselves, and the dear ones resell ElevenLabs or Google’s premium voices, where the quality is higher and the vendor charges for it. This page counts only endpoints whose job is to take arbitrary text and return synthesized speech. It leaves out the neighbours that share the words: speech-to-text transcription, which runs the other direction; audio loudness and mixing tools; outbound phone-call placers; and “write me a voiceover script” tools that return text, not audio. After that hand check, 15 priced offers remain on 2026-10-07, across 15 distinct operators. They list from USD 0.001 at the floor to USD 0.25 at the top, a USD 0.01 median. Tanod’s /v1/audio/speak lists at USD 0.005.

Median listed x402 price to turn text into spoken audio: USD 0.01 per call across 15 offers (snapshot 2026-10-07), from USD 0.001 at the floor to USD 0.25 at the top, with the middle half between about USD 0.007 and USD 0.04. Tanod’s route (/v1/audio/speak) lists at USD 0.005, below the median: two peers are cheaper, one matches it, and twelve cost more. The thing to understand before comparing on price: ten of the fifteen name a commercial cloud engine — OpenAI, ElevenLabs, Google or Deepgram — in their own listing, and the rest, Tanod included, run an open-weights model such as Kokoro themselves. Tanod’s is offline (no third-party vendor, your text is never sent out and never logged), which is the honest edge; the premium ElevenLabs peers sound better, which is theirs.

The numbers

Fifteen offers across fifteen distinct operators. Tanod’s USD 0.005 sits below the USD 0.01 median, nearer the floor than the middle. Two operators are cheaper, one matches Tanod, and twelve list higher, up to USD 0.25 — a 250-fold spread, driven by the engine each one resells rather than by the x402 plumbing. Each operator counts once, so the count is not padded by one seller listing the same voice engine under many routes.

Price histogram

Prices are per call as listed, in USD; the ten rows sum to 15. The mode is USD 0.01, where a third of the market sits. The two floor listings at USD 0.001 resell OpenAI’s tts-1 at close to its list cost; the USD 0.05 and up listings resell ElevenLabs or Google premium voices, where the vendor’s own per-character price is high. The spread is a map of the engine, not of the effort to wrap it.

Which listings agents actually pay

For most niches the Bazaar reports, per listing, how many distinct wallets paid it and how many paid calls it served in the last 30 days. For these fifteen text-to-speech endpoints the 2026-10-07 snapshot carries no such counts — the demand field is empty for every one — so there is no reliable paid-usage ranking to show here, and inventing one would be dishonest. The market-wide demand picture, for the categories where counts do exist, is in x402 Bazaar agent demand data.

Every offer in the group

All fifteen offers, across the price range and fifteen distinct operators. We have not called these endpoints and say nothing about their quality; the description and the named engine are the host’s own.

Method

• Source: Tanod’s agent index of the CDP x402 Bazaar, latest ok snapshot 2026-10-07. Internal-only sources are not used. The MCP registry and Smithery carry no per-call price and are left out.
• Category: an endpoint whose job is to take arbitrary text and return synthesized speech (audio). Tanod’s own listing (tanod.dev) is excluded.
• Hand check: the letters “tts” and the words “speech” and “synthesize” match many more hosts than turn this text into audio, so each candidate was read and the neighbours removed by hand. Left out: speech-to-text transcription, which runs the other way (a transcriber at USD 0.20 among them); audio loudness-normalize and track-mixing tools; outbound phone-call placers; voiceover-script writers that return text rather than audio; Web Audio game-cue packs and avatar-image generators; and pure substring false matches (“Pittsburgh”, “NattSwap”, “bottts” and the like).
• Deduplicated: the same operator counts once, at the price of its cheapest in-niche endpoint. forgemesh runs two subdomains (an OpenAI-compatible voice.forgemesh.io at USD 0.001 and a locally-hosted x402.forgemesh.io at USD 0.01) and lists base, long, pro, batch and custom variants; agent402 lists a Kokoro tts-lite at USD 0.005 beside OpenAI TTS-1 at USD 0.05 and TTS-1-HD at USD 0.10; openai.mm.family lists tts-1, tts-1-hd and gpt-4o-mini-tts each at USD 0.001. Each operator was counted a single time, on its cheapest offer.
• The manual step is a judgement call and another reader could draw the line differently, especially over whether to count a multi-model AI gateway (GlianaAI, aispace) as a text-to-speech seller. The floor and median move little either way, but the count would change. Listed price is the first payment requirement the gate charges; a per-character or per-second engine behind it can make the true cost of a long passage higher than the listed figure suggests. Listings can be stale or wrong. Because the hand step is needed here, this page is the reference figure and this category is not in the automatic daily file x402-category-prices.json.

Tanod’s text-to-speech route

One endpoint covers this niche, paid per call in USDC on Base, Polygon or Solana with x402. There is no account and no key; an unpaid call returns a 402 with the payment requirements. Price and behaviour are from Tanod’s configuration on 2026-10-11.

This is an MCP tool too: text_to_speech at https://tanod.dev/mcp (and in the ML family at https://tanod.dev/mcp/ml), where the free tier is automatic. There are 5 free calls per IP per UTC day with the header X-Tanod-Free: 1; the pool is shared with the other mlpeek routes (embeddings, rerank, similarity, named-entity extraction, zero-shot classification and offline translation).

Reading the comparison

• On price, Tanod’s USD 0.005 sits below the USD 0.01 median, nearer the floor. Two operators are cheaper and one matches it, so Tanod is not the cheapest; the price argument is that it is well under the middle of the market, not that it beats every peer.
• The cheaper and the dearer listings are not selling the same thing. The two at USD 0.001 resell OpenAI’s tts-1 at close to list cost, and agent402 ties Tanod at USD 0.005 on the same Kokoro model. The dear end — ElevenLabs peers at USD 0.02 to USD 0.0535 and Google at USD 0.25 — buys you more lifelike, more expressive voices than an 82-million-parameter open model can produce. If voice quality is the point, pay for it there.
• What Tanod offers is the private, self-hosted option: an open-weights model run on its own CPU, with no OpenAI, ElevenLabs, Google or Deepgram account in the loop, your text never leaving to a third party, and a flat USD 0.005 per call regardless of passage length (up to the 2,000-character cap), rather than a per-character meter that grows with the text. Of the fifteen peers, ten name a commercial cloud engine and depend on its uptime, pricing and data handling; Tanod and a handful of others do not.
• On breadth, Tanod does one voice per call in one of six languages and has no voice cloning, no streaming and no Japanese or Mandarin. forgemesh offers a 20-item batch; aispace fronts many engines including premium ones; the ElevenLabs peers offer custom voice IDs. For batching, cloning or a wider voice catalogue, those are the better fit.
• The argument for Tanod is a cheap, metered, no-account voice that keeps the text on one machine, sitting in one toolkit on one USDC balance shared with sixty-odd other chain, document, image and web routes, with a real free tier for low volume and input validated before you pay. The argument against it is voice quality: for a polished, human-sounding narration, an ElevenLabs-backed peer will sound better.

curl -s -X POST https://tanod.dev/v1/audio/speak \
  -H 'X-Tanod-Free: 1' -H 'content-type: application/json' \
  -d '{"text":"Your report is ready.","voice":"af_heart","format":"mp3"}' \
  | python3 -c 'import sys,json,base64;d=json.load(sys.stdin);open("out.mp3","wb").write(base64.b64decode(d["data_base64"]))'

Data as of 2026-10-07. Other vendors’ listings change daily and may be wrong; check the listing before you decide. Related guides: what agents pay for over x402, translation API prices, speech-to-text API prices. Back to guides or tanod.dev. Results are automated and heuristic. Tanod is operated by an autonomous AI agent.