Voice AI's price war is being fought with expiry dates

I.Three launches, three expiry dates

On 23 September, Google released Gemini 3.8 Flash TTS and Flash-Lite TTS, covering more than 100 languages and over 2,000 ready-made voices. Its audio output is priced at $9 per million tokens through 31 December, rising to $18 from 1 January 2027. The doubling appears in the pricing table, not the announcement.

Five days later, ElevenLabs shipped Eleven v4 and v4 Turbo, with Turbo reaching audible speech in a median of about 150 milliseconds. Digital Applied lists v4 at $0.08 per 1,000 characters and Turbo at $0.04, both cut 72% until 12 October.

On 1 October, Microsoft AI released MAI-Transcribe-2-Streaming at $0.54 per audio hour "through the end of the year", alongside MAI-Voice-2.1 at $22 per million characters and MAI-Voice-2.1-Flash at $15.

Read as a price war, this is a war of introductory offers. Every headline number above has an end date inside the next three months.

II.The speech layer became a hyperscaler feature

The bigger change is who sells the voice. With the October releases, Microsoft can offer streaming transcription, a language model and speech synthesis inside its own cloud. Tech Times describes this as completing the voice-agent pipeline, so developers no longer need a third-party transcription vendor. Google already bundles TTS into the Gemini API and named LiveKit, Pipecat, Agora and Vercel as launch partners.

That squeezes specialists in theory. In practice the leading specialist is growing fast. TechCrunch reports ElevenLabs' annualised revenue rose from about $330 million at the start of 2026 to more than $600 million, with 55% from large companies. A hyperscaler entering a market and a specialist doubling revenue in the same year are not contradictory. The market is expanding faster than share is being taken.

Each vendor also claims first place on a different scoreboard. ElevenLabs cites the Artificial Analysis voice arena. Google cites Hume AI's Voice Design Benchmark, where Flash TTS scored 71.4. Microsoft cites Artificial Analysis for transcription accuracy. These are separate tests of separate things, and no single ranking settles who leads.

III.The honest technical read

The latency gains are real. A full conversational turn on Microsoft's stack takes an estimated 580 to 780 milliseconds, with the language model consuming most of it. Speech is no longer the slow part of a voice agent.

The production case is weaker than the launch posts suggest. Rajesh Beri's 3 October review notes that MAI-Transcribe-2-Streaming is a preview Microsoft does not recommend for production, with no service-level agreement, no word-level timestamps or confidence scores, and preview terms that let Microsoft store submitted data for abuse monitoring. For calls carrying financial or health details, that is a compliance problem, not a footnote.

Microsoft is also not the cheapest transcriber. Its 2.5% word error rate edges Grok Voice Transcribe 2.0 at 2.73%, but at 2.7 times Grok's $0.20 per hour. Benchmark error rates are a ceiling. Tech Times cites academic work finding real-world error rates above 13% for systems that test under 5%.

Two of the vendor claims need correcting. ElevenLabs' preference figures come from its own blind tests. And its own launch post describes the promotion as "3x credits" for Creator plans and above, which works out to roughly a two-thirds cut, while Digital Applied's per-character API rates imply 72%. We report both rather than reconcile them.

IV.The economic ripple runs through prices, not yet payrolls

Prices. On the assumptions in Fig. 2, the voice layer of an agent costs roughly $1 to $3 per conversation-hour at list prices, against the $21.53 median hourly wage for US customer service representatives. That is an estimate, and it leaves out the larger costs: language-model tokens, telephony, integration and supervision. But it means that halving or doubling TTS prices barely moves the business case. The introductory discounts matter for developer adoption, not for the labour arithmetic.

Labour. The exposed pool is large. BLS counts 2.67 million US customer service jobs and projects a 5% decline over 2025–35, citing automation. Gartner put the global agent workforce at about 17 million when it predicted in 2022 that conversational AI would cut agent labour costs by $80 billion in 2026. The fine print in that forecast was that only one interaction in ten would be fully automated.

Survey evidence since then cuts against fast displacement. In a survey of 321 service leaders in October 2025, only 20% reported AI-driven headcount reductions. 55% held staffing flat while handling more volume. Gartner separately predicts that half of organisations planning major service-workforce cuts will abandon those plans by 2027.

Market structure. Cheap speech hurts vendors who sell only speech. Transcription is the most exposed: Microsoft's batch model costs $0.10 per hour, and rivals stream for $0.18 to $0.20. Value shifts to whoever owns the agent platform, the customer relationship and the telephony. ElevenLabs is already moving there with ElevenAgents. Single-model API vendors without a platform are the likeliest consolidation targets.

Capital. Investors are still betting that specialists win. ElevenLabs raised $500 million at an $11 billion valuation, and TechCrunch reports talks at $22 billion. If hyperscaler bundling were compressing specialist margins, that valuation would be hard to defend. The next funding disclosure is a direct test.

V.Second-order effects: the bottleneck moves

Once speech is cheap, the constraints that remain are reliability, trust and law. Google's voice cloning is unavailable in Illinois, Texas, the EEA, the UK, Switzerland and India, and all output carries a SynthID watermark. Biometric and consent rules, not model quality, now decide where cloned voices can ship. For a Spanish or European deployer, the cloning feature does not exist yet.

The second effect is that cost moves upstream. With the speech layer near a dollar an hour, the language model handling each turn becomes the dominant variable cost in most voice agents. Labs that own both the model and the voice can price the bundle as one product. This is a structural advantage for Google and Microsoft over stand-alone voice vendors that pay someone else for reasoning.

The third effect is on offshore service hubs. Countries whose outsourcing industries rest on voice work face the same arithmetic as US call centres, at lower wages. We found no recent primary data sizing that exposure, so we flag it as a direction rather than a figure.

VI.Precedent: Klarna priced the labour, then priced the quality

In February 2024, Klarna said its AI assistant was doing the work of two-thirds of its customer service chats. By May 2025 it was recruiting human agents again, and its chief executive acknowledged that cost had driven the rollout and quality had suffered. The lesson for voice is direct. The cost of the machine was never the binding constraint. The cost of a bad interaction was.

VII.How it plays out: three scenarios

Probabilities are our analytical judgment, not model forecasts, and they sum to 100%.