Search & Audio AI API Routers

The category groups endpoints that bill outside the LLM-token regime. The 27 entries split into two populations: speech — text-to-speech (Polly, ElevenLabs, Cartesia, PlayHT, Fish Audio) and speech-to-text (Deepgram, AssemblyAI, Soniox, Gladia, Rev) — and search — result APIs (Tavily, Exa, Brave, Serper, SearchAPI, You.com, Perplexity Search, x-search), web data and crawling (Firecrawl), and search infrastructure (Google PSE, SearXNG, Context7, Linkup). What they share is the billing unit: characters, minutes, credits, queries, pages — never tokens — so the directory's per-1M-token columns read empty for these rows, or carry flagged conversions. That is a feature of the taxonomy, not missing data.

It matters to buyers because the two regimes fail differently. A speech API's cost scales with output volume and quality tier (neural voices cost multiples of standard); a search API's cost scales with query volume, and its value scales with result quality on your queries — a cheap per-1K rate that returns irrelevant results is the most expensive price in the category. The entries separate the two concerns: the credit column quantifies the documented free allowance at the provider's published rates, the notes carry the native unit economics, and the status probe tells you whether the endpoint is answering at all.

Three unit shapes do the billing. Per character: AWS Polly prices by 1M characters of speech — $4.00 standard, $16.00 neural, $100.00 long-form, $30.00 generative — with a documented free tier of 5M standard, 1M neural (first 12 months), 500K long-form and 100K generative characters per month. Per minute: Deepgram bills transcription by the second — Nova-3 at $0.0048 per minute streaming and $0.0043 pre-recorded on current published rates — and its documented free credit is $200, good until used and equivalent to tens of thousands of minutes of Nova-3 transcription. Per credit: ElevenLabs runs a credit model — 10,000 credits per month on the free plan, paid tiers up to 600,000 credits per month, with the effective cost per minute of TTS falling as the tier grows — and its enterprise tier adds DPA/SLA, a HIPAA BAA and SSO per the pricing page.

The unit determines the math, so keep it native. A podcast pipeline is a minutes problem: price per-minute rates against your monthly audio hours. A voice product is a characters problem at low volume and a credits problem at scale, where the tier's effective per-minute rate — the pricing pages break it out per minute — beats the headline credit price. Where the dataset converts a speech unit into the site's normalized columns it flags the conversion , because any such conversion must assume a usage shape; the provider's own unit is always the budgeting number.

Search APIs bill per query or per 1,000 queries, and several add result-level or page-level pricing: Exa's published pay-as-you-go runs from $4 per 1,000 search requests for instant search, with contents fetched per page and answers billed separately; Tavily bills search credits at $0.008 per credit on pay-as-you-go with a free 1,000 credits per month; Firecrawl bills scrape pages as credits (1,000 free per month, 1 credit per page on the current table); Serper and SearchAPI bill per 1,000 queries. Google PSE is the outlier with a free daily allocation of search results and paid volume above it — which is why its row carries a quantified allowance rather than a per-unit price.

The value variable here is result quality on your queries, and no dataset column measures it: the entry's description and notes carry what research found, and the community rating carries what users report. The practical test is small and cheap — take 50 of your real queries, run them on the two finalists' free allowances, and score recall and freshness yourself. That one hour of testing is most of the comparison, and it is why the evaluation section below is built around query sets rather than price tables.

25 of the 27 entries carry a quantified allowance, and the allowances differ in kind: monthly resets (Polly's 5M characters, Tavily's 1,000 credits, ElevenLabs' 10,000 credits, Firecrawl's 1,000 pages reset monthly or per billing period), one-time grants (Deepgram's $200 credit, good until used), and daily caps (PSE's free daily allocation). The kind changes the plan: a monthly reset supports a standing background job sized below the allowance; a one-time grant is a prototype budget; a daily cap supports a daily job. The credit column on the entry is the quantified value of that allowance at the provider's published rates — compare rows as USD, not raw units, but never budget from the USD figure alone: read the kind in the notes first.

Speech: pick the unit that matches your pipeline (minutes for audio processing, characters or credits for generation), then compare quality tiers at that unit — neural and generative tiers cost multiples of standard, and the entry's note says which tier the tracked rate belongs to. Check the compliance line if the audio is sensitive: the providers in the sources document SOC 2 and, for enterprise, HIPAA options (Deepgram and ElevenLabs both). Search: shortlist two or three, run your own 50-query test on free allowances, price the winner at your real query volume including result-level extras, and set a monthly query budget before you go live — search cost is the one that silently scales with traffic, because each user request can fan out into several API calls.

Speech: (1) unit economics at your volume — compute cost per minute of audio or per 1M characters at the tier you need, not the headline rate; (2) quality-tier availability — the rate your task requires (neural vs standard, multilingual) may sit on a different row of the provider's table than the minimum shown on the entry; (3) latency shape — streaming vs batch changes both price and architecture; (4) compliance documentation where the audio is sensitive. Search: (1) result quality on your own queries — the only metric that matters, and the only one you can measure yourself; (2) coverage — does the index include the sources your task needs; (3) cost per query including result extras; (4) rate and fair-use limits on the tier you will actually run.

What this page does not measure: speech quality (voice similarity, word-error rate on your domain's audio) and search relevance (recall on your queries) — both are your test set's question, run on the free allowances before you commit. It does establish the smaller facts: the documented allowances and their reset kind, the published unit prices at the verification date, and each endpoint's live status.