Response Time
How fast does the model respond after the caller stops speaking?
Caller speech ends00:12.49
Model starts00:13.07
P50
P90
Measures response latency with P50 for the typical caller experience and P90 for slower tail responses.
Better conversations start with the right model.
Compare leading AI models across response quality, speed, and reliability. Find the right balance for your voice agent and choose with confidence.
What we evaluate
Model performance, measured where voice agents fail.
Retell tests the behaviors that make a voice agent feel fast, reliable, and ready for production.
Grounding
Does the model stay within the provided context?
Grounded
Answers stay within supported information.
Unsupported
Answers contain unsupported or unverifiable claims.
0%50%100%
Averaged across every model scored on Retell's grounding suite: answers supported by the prompt, tools, or retrieved context.
Tool Calling
Does the model choose and execute the correct tool?
Tests whether the model selects the right action, passes valid arguments, and uses tool results correctly.
Instruction Following
Does the model follow the system and task instructions?
Checks whether the model follows role constraints, business rules, formatting requirements, and refusal rules.
Task Completion
Does the model complete the user's goal?
Evaluates whether the model resolves the task, asks for missing information, and avoids dead ends.
Model Benchmarks
We benchmarked the leading models for you.
Compare quality, latency, and reliability across top providers, with filters that make the tradeoffs easier to see.
All models
Whether answers stay within the prompt, tools, and retrieved context.
#Position in the current sort. Each model's pass-rate rank within the suite is shown in its details.ModelEach row is one reasoning variant of a model, exactly as the benchmark scored it.
ScoreScore is the model's pass rate — the percentage of tested scenarios it handled correctly in this suite.
LatencyLatency is the P90 of Retell's live provider health probes over the last 24 hours. It is measured per model, so every reasoning variant of a model shares one value.
CostCost is the provider's list price in US dollars per minute of audio. A dash means no rate-card entry is published for this model.
Revolutionize your call operation with Retell
