ElevenLabs Agents Review (2026)

Paying $990 a month buys exactly the same rate per minute as paying $6.

Divide each published price by the minutes it includes and the answer is $0.08 at every paid tier. Starter is $6 for 75 minutes. Business is $990 for 12,375. Both are eight cents, and eight cents is also what an extra minute costs on any plan, including the free one.

So the ladder is not a volume discount. It is a concurrency ladder with seats and voice clones attached, and it is worth knowing that before someone buys the $299 tier expecting the unit cost to fall.

What the eight cents actually includes

This is the question the site exists to ask, and here the answer is unusually good. Speech-to-text and text-to-speech are both inside the rate. The language model and the telephony are not.

The pricing page states it plainly in its FAQ, and its own cost calculator itemises the same split, showing platform voice at $0.08 alongside a separately-billed model line and a footnote that telephony is billed separately.

Add the two exclusions at this site's list rates and the computed floor is $0.1048 a minute against an advertised $0.08. A 31% gap, which puts it fourth of the six platforms costed here and exactly where its architecture says it should be.

The ranking is not a quality judgement, it is a reading of what each rate covers. The three above it bundle the model as well and exclude only the phone line. Vapi and Millis quote orchestration alone, which is why their multiples run to 1.9× and 3.3×. ElevenLabs sits in the middle because it bundles two components of four, and it is the only one of the six bundling synthesis made by the company selling the agent.

The model rate card nobody quotes

The pricing page carries a calculator with a model picker, and behind that picker is something no ranking page reproduces: the vendor's own conversion of token pricing into cost per call minute, model by model.

Two things follow. The model is the largest swing factor in the bill, and choosing a fast cheap model rather than a frontier one is worth more than any tier upgrade on the page. And a custom model of your own is billed by ElevenLabs at zero, so bringing your own key removes that line entirely.

It is also worth noting where the vendor's assumptions differ from ours. Working backwards from the calculator, it appears to assume roughly 1,240 input tokens a turn where this site's model assumes 1,800. Their per-minute figure is the more optimistic of the two, and long calls with growing transcripts push it further apart.

The bundled minute undercuts the vendor's own synthesis

Here is the arithmetic that makes this product genuinely unusual. Standalone text-to-speech on the Creator plan costs $0.18 per minute of generated audio. An agent speaks roughly half of a two-party call, so that is $0.09 per minute of call.

The agent minute is $0.08, and it includes the transcription as well.

Buying the whole agent from ElevenLabs is cheaper than buying only the voice from ElevenLabs. That inverts the pattern this site was built to expose, where the platform layer marks up components you could buy directly. Here the platform is where the discount lives.

Burst, and what a spike costs

Exceed your concurrency limit and calls do not fail. They burst, up to three times the plan's limit, at double the rate. Sixteen cents a minute, which is $0.1848 all-in on this site's model.

The part worth reading twice is not the price. Burst calls are deprioritised for speech processing and may run at higher latency, which in a voice agent is the failure that matters. You pay double for a worse call, and past the burst ceiling calls are rejected outright. Concurrency is the number to size correctly, and it is the only thing the tier ladder actually sells.

The promise that was kept

In February 2025 the vendor published a post cutting conversational AI prices, and included a sentence worth preserving: the prices did not include model costs, which it was absorbing at the time and would eventually pass on.

It has. The pricing table now has an External Providers section, and the model is in it. That is a company doing exactly what it said it would do, and it still means a buyer comparing today's rate to a 2025 review is comparing two different products. Every page ranking for this product's pricing describes the current table as though it had always looked that way.

The byline audit, which is the most one-sided yet

Nine results for this product's pricing query. Six are ElevenLabs' own pages. Two more are written by companies selling to the same buyer: one sells billing infrastructure to AI companies, the other sells testing and observability for voice agents. The ninth is a Reddit thread inside the vendor's own subreddit.

Not one independent editorial page ranks. On a query that is purely about what something costs.

Self-serve to $990 a month, then a sales conversation. The free tier gives fifteen minutes with no commercial licence, which is enough for about seven two-minute test calls and nothing you can ship.

• Size concurrency, not minutes. Minutes cost the same on every plan. Concurrency is the only thing that changes, and exceeding it means double rate at degraded latency.
• Pick the model before the plan. It swings the all-in cost by 47% on the vendor's own figures, which is more than any tier decision on the page.
• Bring your own model key if you have one. Custom models are billed at zero by ElevenLabs, so that line disappears from the bill entirely.
• Arrange telephony separately and price it. Twilio, Telnyx, Plivo, Vonage and SIP are all documented. Static-IP SIP is enterprise-only, which matters if your carrier requires it.
• Do not buy annually on lumpy volume. The annual discount applies to the included block and not to the marginal minute, which stays at $0.08.

Unusually complete, and the completeness is the reason the flat rate is visible at all. Two things on the page still need asking about.

The first is credits. Agent plans are described as billed by call minutes rather than from the shared credit pool, and then the same FAQ says model fees are deducted from credits and that you are limited by your monthly credit limit. Minutes and credits appear to be separate allowances on one subscription, and the page never says how many credits an agent plan carries. Ask.

The second is the enterprise rate card, which publishes a fallback of $0.070 a minute for agent audio where a contract does not specify one. Set that against the self-serve annual effective rate of $0.0667 on the Business plan and the published enterprise default is the more expensive of the two. Enterprise buyers negotiate, so this is not evidence that enterprise is worse value. It is evidence that the only enterprise number published is above the self-serve one, which is a useful thing to have read before the call.

One more comparison worth carrying: the same vendor sells a receptionist product whose overage runs $0.40 a minute, five times the agent rate, on the same voice over the same phone network. The packaging decides the price far more than the technology does.