Microsoft and Amazon Just Made Your Multi-Year AI Contract Look Overpriced

On October 1, Microsoft AI released a streaming transcription model and two voice models built for agents and call centers. It priced the faster voice model, MAI-Voice-2.1-Flash, at $15 per million characters. Microsoft says Flash is about 60% cheaper and 55% faster than comparable models. The Information framed the release as a challenge to ElevenLabs. A day earlier, ElevenLabs had closed a $300 million employee tender at a $22 billion valuation and said revenue from its agent platform had more than tripled since February.

The same day, Amazon's Strands Labs open-sourced Strands Decider 2B, a 2-billion-parameter model that picks one answer from a fixed list. It attaches a confidence score and runs on a laptop at a median of about 115 milliseconds. It's a free take on the idea TypeSafe brought out of stealth on September 15 with Jev, which it prices at $0.042 per million input tokens and nothing for output. Amazon distinguished engineer Marc Brooker started the project after Jev's launch. Amazon's was one of four Jev-style releases in 16 days, alongside OpenAI's Decisions API, Cloudflare's Clef and Perplexity's Decisions API. Meanwhile, TypeSafe is in talks to raise more than $1 billion at a valuation above $10 billion, as reported by The Information on September 24, nine days after the launch.

Both releases follow the same pattern. A startup shows that buyers will pay for a narrow AI job, and a company that owns its own compute answers with a cheaper version. Against Jev that took 16 days. Microsoft's voice launch landed the day after ElevenLabs' tender closed. For anyone buying voice-agent minutes or AI routing and scoring calls, that shortens how long a quoted price can be trusted. The market will likely have undercut a rate signed this quarter before the contract reaches its second year.

Microsoft's 60% claim holds at list price and slips against ElevenLabs' discount

The "about 60%" figure checks out against ElevenLabs' standing rate card. Against its current promotional price, it doesn’t. ElevenLabs' API pricing page lists its Flash and Turbo voice models at $0.04 per thousand characters, or $40 per million. Microsoft's $15 is roughly 62% below that. Microsoft's own announcement never names ElevenLabs, though. The comparison came from press coverage, so the honest reading is a Microsoft claim measured against a list price.

ElevenLabs shipped Eleven v4 and v4 Turbo on September 28, and its pricing page shows a 72% launch discount running until October 12. That puts v4 Turbo at $11 per million characters, below Microsoft's Flash. ElevenLabs is also cheaper on streaming transcription. Microsoft's new streaming model carries an introductory rate of $0.54 per hour of audio through the end of the year. ElevenLabs lists its realtime transcription at $0.39.

ElevenLabs' discount and Microsoft's introductory streaming rate both carry an expiry date. Microsoft's batch transcription line shows where expiring prices tend to settle. MAI-Transcribe-1 launched in April at $0.36 per hour, and MAI-Transcribe-2 arrived in September at $0.10. The price fell 72% in five months on Microsoft's own price list. The September price is also labeled a limited-time offer, which shows a buyer how the vendor expects the market to move.

Decision models reached open weights 16 days after Jev launched

Decision models matter to GTM teams for a plainer reason than voice does. Most of what a revenue stack asks AI to do is choose from a short list. That means routing a lead, scoring an account, tagging a ticket, or deciding whether this reply is a meeting request. Paying a chat model to write a paragraph and then parsing a category out of it was always the expensive way to get that answer. TypeSafe's own launch post names classifying, routing and scoring as the jobs Jev is for.

The category didn't stay single-vendor for long. OpenAI announced a Decisions API at DevDay on September 29, built on GPT-6 Luna, then opened it as a public beta on October 6 at $0.10 per million input tokens, the same rate it already charges for Luna input. On October 1, Cloudflare released Clef and Clef-flash under an Apache 2.0 license and called them fully compatible with Jev's API. Clef-flash answers in a median 38.8 milliseconds. Perplexity's Decisions API launched at $0.04 per million input tokens with output free, a little under Jev; its pricing docs now list $0.02, about half Jev's rate. The model behind it is free to download. Ollama added local support for Jev-style models the same week.

Once a model is open-weight and API-compatible with the original, the price of a routing call drifts toward its compute cost. That applies to every vendor in a GTM stack that resells a classification step under a label like "AI scoring" or "intelligent routing." It applies whether or not the vendor has said which model sits underneath.

The counter-case worth taking seriously

ElevenLabs is the strongest evidence against reading any of this as a race to zero. Its agent revenue has tripled since February, while Microsoft was already shipping its own speech models. Its tender announcement says enterprise customers account for 55% of revenue, and it names Stripe and Deutsche Telekom among the customers. An enterprise putting a voice agent on its phone lines pays for more than the model. It pays for latency under load and voices that hold up on a bad connection. It also pays for compliance review and a platform that manages the handoff to a human. The per-character price of the speech model is one input among several. Quality rankings still matter too, and ElevenLabs says Artificial Analysis ranks v4 first.

Decision models have their own version of the objection. Jev's founder, Diogo Almeida, told TechCrunch the clones looked like "ML people wanting to implement a cool architecture." Perplexity's published benchmark table backs part of that. Its model beats Jev overall, 85.71% to 84.51%, but trails it on the hard tier of JevBench, 70.30% to 73.27%. A specialist’s edge still shows on the hardest calls, where a misrouted enterprise lead costs real pipeline.

So the price pressure lands unevenly. The raw model layer, speech priced per character and decisions priced per token, is getting cheap fastest. The layers above it hold their price longer, because that’s where vendors package reliability and someone to call when the agent misbehaves. If a buyer can't say which layer an invoice is paying for, it can't tell whether a quote is fair or not either.

What this means for AI buyers and vendors

Keep usage-priced AI terms short. Any line item billed per minute, per call or per token should run 12 months or less, or carry a right to reprice at each anniversary. A three-year voice-agent agreement at today's per-minute rate bets that prices hold into 2029. Microsoft just cut a transcription price by 72% in five months, and ElevenLabs is running a 72% launch discount.

Where a longer term buys a real discount, write the decline into it. Ask for a step-down clause tied to the vendor's own public list price, so the contract rate falls with list at the next billing cycle. Ask voice-agent and AI SDR vendors which speech and decision models sit underneath their product. A vendor reselling a model whose price just fell has room to pass some of the saving on. RevOps teams paying for AI lead scoring should price the same job against a decision model before the renewal conversation. The routing call behind that line item now has a public benchmark of $0.02 per million input tokens, Perplexity's current list price.

Vendors sit on the other side of the same table. Any company whose gross margin depends on marking up model output should expect buyers to arrive with Microsoft's and Perplexity's price pages printed out. That includes per-minute voice platforms and per-seat scoring tools. The margin that survives is in the work the platforms haven't shipped yet, such as handoff logic and the action log a security team will ask to see before it approves the tool.

Anthropic put that clause to a test on October 7, when it released Claude Haiku 5.5 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100,000 tokens, a tenth of Haiku 4.5's $1 and $5. Anthropic's own estimate of the average saving is around 75%, lower than the list cut because the new model's updated tokenizer uses slightly more tokens per task. The same announcement halved Sonnet 5.5's cache-read price to $0.10, while Haiku 4.5's list price didn't move, so a step-down written against a named model version catches the Sonnet cut and misses the Haiku one unless it also covers successor models. That $0.10 input rate matches what OpenAI charges for GPT-6 Luna, the model behind a Decisions API that costs about 2.4 times Jev's rate for a routing call.

The Information reported on October 24 that OpenAI is cutting prices to win Anthropic's largest customers. Rippling chief executive Parker Conrad told the outlet that OpenAI's share of Rippling's coding work rose from 16% in May to 52%, and DataCamp's chief executive said the company's AI costs fell 80% to 90% after it moved, per the report. One person familiar with the deal said Replit got more than 30% off a contract worth over $100 million a year, which Replit hasn't confirmed; Replit had already made OpenAI its default model in mid-October. That's a model-layer rate undercut within months, not years, which is the risk this piece raised for usage-priced contracts that run into a second year.

My bet is that ElevenLabs keeps v4 Turbo below its $40 list after the launch discount ends on October 12. If it does, this year's rate card is already a ceiling, and any multi-year contract priced off it is priced too high.

If a buyer can't say which layer an invoice is paying for, it can't tell whether a quote is fair or not either.

Keep usage-priced AI terms to 12 months or less, or write in a step-down tied to the vendor's public list price that also covers successor models.

By Business of GTM