ElevenLabs Pay As You Go: A Cheaper Voice Agent Minute Is Also a Contract Change You Are Asked to Accept

ElevenLabs has lowered the price of its API and its agent platform and introduced a pay as you go model, and the announcement page shows a last update of 28 September 2026. The headline numbers are straightforward. Text to speech is up to 55 percent cheaper, speech to text is up to 45 percent cheaper, and ElevenAgents is up to 20 percent cheaper. On the Starter agent plan the rate moves from 0.10 to 0.08 dollars a minute. The company says performance, quality and support are unchanged, and that customers can pay only for what they use without an upfront commitment. For anyone running a voice agent, that is a real saving, and it is worth taking. It is also worth reading as a change to a contract rather than as a discount.

Start with the mechanism, because it is the part most teams will skim. Existing subscribers do not move automatically. They go to their subscriptions page and choose Switch to new pricing, and free plan users can start pay as you go by adding credits. That means the cheaper price is an opt in, and the switch is a decision somebody inside your company makes, usually the person who happens to be logged in as the account owner. A pricing change that you accept with a button is an amendment to the commercial terms of a supplier, and the supplier controls under SOC 2 and ISO 27001 expect a named owner to approve those. Almost nobody records that approval, so the evidence trail for one of your voice vendors quietly ends at the original signup.

The second thing to notice is the words up to. Each of the three reductions is a ceiling, not a rate. The 55 percent applies to a specific model tier of text to speech, the 45 percent to a specific speech to text model, and the 20 percent to the agent minute on the plan where the saving is largest. A team on a different model or a higher plan will see a different percentage, and a team that was already on a discounted annual commitment may see nothing at all, or may see that pay as you go is now cheaper than the commitment it is locked into. Do not forecast from the headline. Take last month usage, price it under both schemes line by line, and compare the two totals before anyone clicks anything.

Pay as you go changes the shape of the risk as well as the size of the bill. A subscription gave you a ceiling, because when the included allowance ran out the agent either stopped or you topped up on purpose. Usage based billing removes that ceiling unless you build one. A voice agent that loops, a campaign that dials more numbers than planned, or a prompt that makes the agent talk for twice as long will now simply cost more, quietly, until someone looks at the invoice. Set a spend limit and an alert threshold on the account, write down who receives the alert, and test that the alert reaches a real person. This is the same exposure our piece on AI credits covered, except that here the unit is a minute of live customer conversation, which is harder to pause than a batch job.

The cost per minute is also only one line of a voice agent bill. The language model that drives the conversation, the telephony carrier, and any recording and transcription storage sit on separate invoices, and a 20 percent cut to one of them moves the total by much less than 20 percent. Use the saving as a prompt to work out your fully loaded cost per completed conversation, including the calls that fail or transfer to a human, because that is the number a business case should be built on. If the agent minute drops but the average call gets longer because the new settings make it more talkative, you have saved nothing, and you will only see it if you track minutes per resolved call and not just spend.

From a compliance view, a repricing is a good moment to refresh the facts about the supplier, not just the rate. Check that the data processing terms, the retention settings for audio and transcripts, and the list of subprocessors still match what your records say, since the new plan may carry updated terms and you are the one accepting them. If your agent speaks to customers, the EU AI Act Article 50 disclosure duty that applies from 2 August 2026 is yours as the deployer, and a cheaper minute does not change it. ISO 42001 asks for an inventory of AI systems with an owner and defined limits, and the new spend cap is exactly the kind of limit that belongs in that register. Vanta, Drata and Secureframe can hold the vendor review as evidence if you attach the decision and the date.

There is a quieter benefit to doing this carefully. A voice vendor that cuts prices this quickly is signalling that the market is competitive, which gives you leverage at renewal and a reason to keep your agent portable. Keep your prompts, your call flows and your test scripts in your own repository rather than only inside the platform, and keep a short list of the alternatives you would evaluate if the terms moved against you. You do not need to leave. You need to be able to, because that is what makes a vendor negotiate and what makes an exit plan a real control rather than a line in a policy.

A short list for this week. Price last month usage under the old and new schemes and record the difference. Decide who owns the Switch to new pricing choice and write that approval down with the date. Set a spend limit and an alert on the account and confirm a person receives it. Re-read the data terms, retention and subprocessor list that apply after the switch. Add the vendor and its new limits to your AI system register. Do those five things and a price cut becomes a clean, evidenced improvement. Skip them and you have changed a supplier contract by clicking a button and told nobody.