Scoreboard
Measured. Not claimed.
Rolling-window aggregates over real runs — p95s and medians, never a hand-picked best. Failures count against the lane that failed, ours included.
PolarGrid · rolling p95 by discipline
Voice pipeline · TTFA
1304ms
p95 · p50 1030ms · 37/42
LLM · TTFT
2031ms
p95 · p50 204ms · 50/50
TTS · TTFB
2745ms
p95 · p50 266ms · 50/50
STT · STT
1173ms
p95 · p50 374ms · 43/46
P95 — rolling window, last 50 runs per lane
Voice pipeline
TTFA — end of user speech → first audio byte
live · rolling · up to 50 runs/lane · Jul 20–Aug 18| Provider | p95 | p50 | Completion | Wins | Runs |
|---|---|---|---|---|---|
| PolarGridPersonaPlex · full pipeline | 1304ms | 1030ms | 88% | 31/42 | 37/42 |
| Vapidefault config | 3261ms | 2641ms | 79% | 0/14 | 11/14 |
| Retelldefault config | 2924ms | 1807ms | 81% | 2/16 | 13/16 |
| ElevenLabsAgents · default config | 1773ms | 1206ms | 100% | 5/12 | 12/12 |
LLM
TTFT — time to first token
live · rolling · up to 50 runs/lane · Jul 27–Aug 11| Provider | p95 | p50 | Completion | Runs |
|---|---|---|---|---|
| PolarGridQwen 3.5 27B | 2031ms | 204ms | 100% | 50/50 |
| Together AIgpt-oss-20b | 2199ms | 354ms | 98% | 49/50 |
| Basetengpt-oss-120b | — | — | 0% | 0/3 |
| Fireworksgpt-oss-120b | 2430ms | 346ms | 100% | 50/50 |
| GroqQwen3.6 27B | 2240ms | 187ms | 100% | 50/50 |
| AnthropicClaude Haiku 4.5 | 3236ms | 726ms | 100% | 50/50 |
| OpenAIGPT-4o mini | 3627ms | 821ms | 100% | 50/50 |
TTS
TTFB — time to first audio byte
live · rolling · up to 50 runs/lane · Jul 22–Aug 17| Provider | p95 | p50 | Completion | Runs |
|---|---|---|---|---|
| PolarGridKokoro 82M | 2745ms | 266ms | 100% | 50/50 |
| ElevenLabsTurbo v2.5 | 3639ms | 502ms | 100% | 50/50 |
| DeepgramAura Asteria | 2730ms | 370ms | 100% | 50/50 |
| GroqOrpheus v1 | 2736ms | 289ms | 100% | 50/50 |
STT
STT latency — audio in → transcript out
live · rolling · up to 50 runs/lane · Jul 27–Aug 10| Provider | p95 | p50 | Completion | Runs |
|---|---|---|---|---|
| PolarGridWhisper V3 Turbo | 1173ms | 374ms | 93% | 43/46 |
| ElevenLabsScribe v1 | 1418ms | 688ms | 96% | 44/46 |
| DeepgramNova-3 | 3198ms | 748ms | 61% | 28/46 |
| GroqWhisper V3 Turbo | 1452ms | 387ms | 95% | 42/44 |
WINDOW — rolling last 50 runs per lane · measuring v1 · live · competitors on their default config
Run it live yourself →Run it yourself — real requests, measured live from your browser
Stream tokens from Qwen 3.5 27B. Watch TTFT live.
Try it →Type text, pick a voice, hear it back instantly.
Try it →Record yourself and see how fast the edge transcribes.
Try it →Talk to an AI — full STT→LLM→TTS on one edge node.
Try it →Race TTS, STT and LLM latency head-to-head.
Try it →