Avatar API latency
benchmark
Eighty people each held sixteen live, unscripted conversations across Anam Cara-4, HeyGen LiveAvatar, LemonSlice 2.1 and Tavus Phoenix-4, and rated what they experienced. 1,280 rated conversations. Every product ran its own default stack.
Measured July 31, 2026
Data from 80 participants
Includes 1,193 sessions
Benchmark snapshot
Median TTFF, default-stack study
Anam
1.550s
HeyGen
4.712s
Tavus
5.720s
🍋
LemonSlice
9.395s
Vendor
Median
P95
P99
n
1 Anam
1.550s
3.606s
8.290s
301
2 HeyGen
4.712s
8.692s
10.624s
300
3 Tavus
5.720s
8.092s
16.366s
304
4 LemonSlice
9.395s
15.130s
39.241s
288
For Anam and Tavus measured connection time included dynamically configuring the avatar used in the call. For HeyGen and Lemonslice, pre-configured avatars were used. Anam lets you configure your avatar at the same time as requesting your session token at no extra cost. For Tavus, this requires a separate API call which we measured independently at 1.4s. To ensure a fair comparison between the competitors we subtracted 1.4s from the measured connection times for Tavus.
median 5.7 s; n=304
Time from pressing start to the first video frame, every rated session. Anam connected fastest for every rater; our 95th percentile, 3.6 s, is quicker than any rival's median.
TTFF is the time from pressing Start to first visible avatar video frame. The benchmark used 80 raters, 1,280 rated conversations, and provider default stacks.
Median TTFF
Typical wait before the avatar appears.
Default-stack realism
Provider production path, not a synthetic best case.
P95 / P99
How bad slow sessions get.
Sample size
How many measured sessions support the number.
Measured on 31 July 2026. These products change continuously and we make no claim about how any of them performs today, the method is published so the run can be repeated against whatever is being served now.
For Anam: Creating a persona (ephemeral), requesting a session token and connecting to the call
For Heygen: Requesting a session token and connecting to the call with a pre-defined persona
For Tavus: Creating a persona, creating the room, joining the room
🍋
For Lemonslice: Creating a room, joining the room with a pre-defined persona
Every rating, session telemetry, raw audio timing, provider configuration, harness code, and reproduction script is published.
At Anam we take take fairness and transparency seriously. We recognise that there is a conflict of interest inherent in running and publishing our own benchmark results. This is why we make all the code and data associated with the numbers we publish available so that anyone can check our work and reproduce it.
© 2026 Anam Labs
SOC 2 Type II report · HIPAA-aligned controls.