Which LLM is fastest in 2026 — TTFT, streaming, and voice benchmark

Role-play model in dialogue

I needed a model that holds a role in conversation — no wall-of-text essays, keeps tone, stays on point. Plain chat without a role falls short.

Role-play (especially via a cloud agent) has another upside: you load instructions and context once, then mostly send the user’s question over the wire — no hauling a huge system prompt every turn. Fewer bytes, lower input latency.

Below: which popular APIs handle that role and how fast they respond. Test setup: 17 models in parallel, 5 repeats, one question.

220 ms

best TTFT

17

models

5

runs

Two ways to set a role

System on every request — prompt in the chat/completion body. Agent / Assistant — role lives at the provider; the request carries only the caller’s line.

What we measured

Operator instructions + the question “Hello, what is this service and why are you calling?” Yandex, OpenAI, DeepSeek, Grok — sync and stream where available.

Terms in plain language

Click a row — short and to the point.

Role-play model

Same neural net, but you tell it upfront who it is: “you’re an operator, keep answers short.” Without that it chats like a generic assistant.

You can put the role in every request. Or configure a cloud agent once — then the instruction sits at Yandex or OpenAI and you only send the caller’s question.

The upside: the long prompt doesn’t travel on every turn — less data, faster response. Especially visible in back-to-back dialogue.

Streaming

The answer arrives in chunks — like someone typing in a messenger. A few words first, then the rest.

Handy for voice: you can start TTS immediately without waiting for the full text.

Milliseconds (ms)

Unit for speed measurements. 1000 ms = 1 second.

220 ms — under a quarter second, almost instant. 1700 ms — a noticeable pause. All delays in the table are in ms.

TTFT

Time to first token — how long until the first word of the answer.

Send question → start timer → first text chunk arrives. Lower TTFT means less silence on the phone.

Sync (no stream)

Standard mode: send request, wait, get the full answer at once. No chunks along the way.

In the table, Sync avg is ms until the complete text arrives.

Why a “slow” model can sound faster

The trap: looking only at total response time. A model may deliver the full text in 600 ms sync — sounds fine. But the caller sits in silence the whole time until TTS gets the complete string.

Another model in stream sends the first chunk in 220 ms — voice starts while the rest catches up ~400 ms later.

Sync

1704 ms

OpenAI mini — silence until full text. Then TTS fires all at once.

Stream · TTFT

220 ms

Yandex lite Chat — first words sooner. Tail finishes in the background.

Fastest first chunk (TTFT, average of 5 runs):

Yandex lite · chat

220

Yandex lite · completion

300

OpenAI mini · stream

1192

DeepSeek · stream

1514

Grok · stream

2053

In short: for live dialogue, watch TTFT and streaming. Sync numbers mean “how long in silence until the full answer.” Parallel burst runs inflate ms slightly, but ranking stays the same.

Why the spread

API geography, prompt length, model size, server load. Yandex sends ~2 large chunks, OpenAI — dozens of small ones. In the burst all 17 started at once — network was shared.

One proxy for all

Every request went through the same proxy — Yandex, OpenAI, and the rest. Proxy latency was identical per model: we compare API speed, not “who got a lucky route.”

Full results

Burst 28 Jun 2026 · Mac · one prompt · avg over 5 runs. For stream: TTFT — time to first chunk, Total — end of generation. Cost — estimate per 1000 characters of a typical turn.

LLM benchmark results: sync avg, TTFT, total stream, and cost per 1000 characters
#	Provider	Model	Mode	Sync avg	TTFT avg	Total stream	₽ / 1k chars
1	Yandex	`yandexgpt-lite`	Completion sync · order	429 ms	—	—	~0.067 ₽
2	Yandex	`yandexgpt-lite`	Completion stream	—	300 ms	659 ms	~0.067 ₽
3	Yandex	`yandexgpt-lite`	Chat stream	—	220 ms	595 ms	~0.067 ₽
4	Yandex	`yandexgpt` Pro	Completion sync	838 ms	—	—	~0.40 ₽
5	Yandex	`yandexgpt-lite`	Agent (Responses)	652 ms	—	—	~0.067 ₽
6	OpenAI	`gpt-4o-mini`	Chat sync	1704 ms	—	—	~0.007 ₽
7	OpenAI	`gpt-4o-mini`	Chat stream	—	1192 ms	1661 ms	~0.007 ₽
8	OpenAI	`gpt-4o`	Chat sync	2104 ms	—	—	~0.12 ₽
9	OpenAI	`gpt-4o-mini`	Assistant v2	4330 ms	—	—	~0.007 ₽
12	DeepSeek	`deepseek-chat`	Chat sync	2094 ms	—	—	~0.005 ₽
13	DeepSeek	`deepseek-chat`	Chat stream	—	1514 ms	2267 ms	~0.005 ₽
15	Yandex	`yandexgpt-5-lite`	Completion sync · order	628 ms	—	—	~0.067 ₽
16	Yandex	`yandexgpt-5-lite`	Completion stream	—	569 ms	986 ms	~0.067 ₽
17	Yandex	`yandexgpt-5-lite`	Chat stream	—	550 ms	968 ms	~0.067 ₽
18	xAI	`grok-4.3`	Chat sync	2376 ms	—	—	—
19	xAI	`grok-4.3`	Chat stream	—	2053 ms	2252 ms	—
20	Yandex	`yandexgpt-lite`	Completion sync · operator	676 ms	—	—	~0.067 ₽

Cost: estimate per 1000 characters (75% prompt, 25% reply). Provider rates — June 2026; foreign APIs converted at 78.91 ₽/$. Grok had no public rate in this run.