Speechify

Our full take

Speechify’s developer platform grew up in mid-2026. It moved onto its own domain as SpeechifyAI (speechify.ai; the consumer reading app stays at speechify.com), swapped the flat pay-as-you-go rate for subscriptions, and, the bigger news, launched a full Voice Agents API with phone numbers, SIP and one all-in per-minute price. This page covers the developer platform, not the consumer app.

Text-to-speech is still the value play, it just prices differently now. Free gives 50,000 characters a month. Starter is $10 a month for 1 million characters, then $10 per million. Pro is $99 for 3 million, then $8 per million. Scale is $499 for 10 million, then $6 per million. Work that out per 1,000 characters and the overage runs $0.010 down to $0.006 by tier, so the old flat rate is now the entry rate and volume genuinely gets cheaper. The old $5,000 Enterprise minimum has disappeared from the page, and voice cloning still arrives with the paid tiers rather than being locked to Enterprise.

Now the agent product, because its pitch lands directly on the sorest point in this market. Most platforms quote a platform fee and then pass through the language model, the transcription, the voice and the phone line as separate metered bills. SpeechifyAI sells one line: from $0.07 a minute, covering the model, speech-to-text, text-to-speech and telephony orchestration, with the pricing page promising “no passthrough, no token math”. The tier detail: Free includes 60 minutes a month (hard cap, bring your own carrier for numbers), Starter includes 120 minutes then $0.075 a minute with one phone number, Pro includes 1,200 minutes then $0.07 with five numbers and the premium models (GPT-4o and GPT-4.1), Scale includes 6,000 minutes then $0.068 with twenty numbers, batch campaigns and bring-your-own-model, and Enterprise quotes from $0.06. For context against the platforms we track, Deepgram’s standard bundled agent rate is $0.075 a minute plus your own phone line, and a normal Vapi stack lands anywhere from $0.05 to $0.30 once the parts add up. One predictable line item at $0.07 is a genuinely strong offer if it holds up on real calls, which we have not yet tested.

On the voices: the library now advertises 1,500-plus voices in 30-plus languages. Note the language number: the site used to say 60-plus, and we have cut our stored figure to match the smaller current claim rather than keep the flattering old one. The model line-up is Simba 3.2 (the recommended flagship), Simba 3.0 still available, a multilingual model, and an English model that carries the cloned voices. The site banner claims Simba 3.2 is number one on the Artificial Analysis TTS leaderboard; that is the vendor’s claim about a third-party board. The dated result we verified ourselves is SIMBA 3.0 placing seventh of seventy-six in May 2026, which was already the cheapest model in that top ten, so the quality-per-pound story does not depend on the newer boast.

The caveats have moved but not vanished. Latency for plain TTS is quoted around 300ms, fine for narration and apps, not in the sub-100ms class the realtime specialists occupy; for live agents the bundled product now manages the whole loop itself, and there is no published end-to-end latency figure for it, let alone a measured one. Compliance is the bigger gap: SOC 2 Type II now appears by name on the Enterprise tier next to SSO and a custom DPA, which is firmer wording than before, but there is still no certificate you can open and no HIPAA mention anywhere, so every compliance box on this page stays unticked. And the agents product is young: it reached general availability in 2026, the concurrency ceilings on self-serve tiers are modest (3 to 30 calls), and there is no public track record yet.

My read: as a text-to-speech building block Speechify keeps the value crown, top-ten quality at the lowest credible price, now with volume discounts on top. The Voice Agents API earns an immediate shortlist spot for a cost-driven buyer who wants one predictable line item instead of five pass-through bills, with the standard young-product caution: no reviews, no measured latency, thin paperwork. We will fold it into the test-call run. The scores are an editorial preview as everywhere on this site, and voiceRange eases a point with the language claim halving. Everything above comes from SpeechifyAI’s pricing, voice-agents and docs pages, captured 2026-07-11, with the screenshot in evidence.