Complete latency table
TTFT = time from request send to first token received. Completion = time to full response (includes reasoning tokens where the model generates them). All values in milliseconds, measured 2026-08-16, sorted by p50 TTFT.
| Model |
Provider |
p50 TTFT (ms) |
p95 TTFT (ms) |
p50 completion (ms) |
Notes |
| GPT-5-minifastest | OpenAI | 180 | 600 | 800 | Cheapest frontier mini; snappiest first token of 2026 |
| DeepSeek Chatfast open | DeepSeek | 240 | 1,100 | 1,100 | Open weights; chat-tier speed at near-zero cost |
| GLM-4fast open | Zhipu AI | 260 | 1,150 | 1,300 | Open weights; strong coding, low TTFT |
| Qwen 3.5fast open | Alibaba | 280 | 1,200 | 1,400 | Open weights; multilingual |
| Grok Chat Fastfast | xAI | 290 | 1,250 | 1,450 | Consumer chat; low latency |
| Llama 4open | Meta | 300 | 1,300 | 1,500 | Open weights; symmetric in/out speed |
| Kimi K2.6open | Moonshot AI | 310 | 1,350 | 1,550 | Open weights; long-context |
| Claude Sonnet 4 | Anthropic | 320 | 1,400 | 1,800 | Balanced coding & agentic work |
| Gemini 2.5 Pro | Google | 350 | 1,500 | 1,900 | 1M-token context; strong multimodal |
| GPT-5 | OpenAI | 420 | 1,800 | 2,100 | Frontier workhorse; mid-pack TTFT |
| Claude Opus 4slow | Anthropic | 680 | 3,200 | 4,500 | Deep reasoning; enterprise workloads |
| DeepSeek R1slow open | DeepSeek | 900 | 5,000 | 8,200 | Open reasoning; verbose chain-of-thought by design |
Fastest models in 2026
Ranked by p50 time-to-first-token — the number users actually feel. Sub-300ms models make chat feel instant.
⚡ Top 5 fastest (p50 TTFT)
- GPT-5-miniOpenAI · mini-tier180ms
- DeepSeek ChatDeepSeek · open weights240ms
- GLM-4Zhipu AI · open weights260ms
- Qwen 3.5Alibaba · open weights280ms
- Grok Chat FastxAI · consumer chat290ms
📉 Lowest p95 tail latency
- GPT-5-mini6× slower than p50 at worst600ms
- DeepSeek Chatp95 vs 240ms p501,100ms
- GLM-44.4× p50/p95 ratio1,150ms
- Qwen 3.5stable under load1,200ms
- Grok Chat Fasttight tail for chat1,250ms
Slowest models — and why that's often fine
Ranked by p50 TTFT, slowest first. High TTFT almost always means the model is doing more work before answering.
🐢 Top 3 slowest (p50 TTFT)
- DeepSeek R1DeepSeek · open reasoning900ms
- Claude Opus 4Anthropic · frontier680ms
- GPT-5OpenAI · frontier420ms
DeepSeek R1's 900ms is a feature, not a bug. Reasoning models generate hidden chain-of-thought tokens before the first visible token. R1's p50 completion of 8.2s — the longest on this chart — is the price of publishing its full reasoning trace. With streaming, you watch the reasoning arrive token by token; the perceived wait is far shorter than the raw number suggests.
Latency vs cost vs quality: the real tradeoff
Raw speed is only one axis. The right model depends on what your request is worth and how hard it is.
What the numbers say
- Fast ≠ expensive. The three fastest models (GPT-5-mini 180ms, DeepSeek Chat 240ms, GLM-4 260ms) are also the cheapest on the Model Price Index — mini-tier and open-weight models win on both axes.
- Slow ≠ premium. Claude Opus 4 (680ms p50 TTFT) is the most expensive model on the market at $15/$75 per 1M tokens, yet its TTFT is 3.8× slower than GPT-5-mini's. You pay for reasoning depth and output quality, not responsiveness.
- Quality correlates with completion time, not TTFT. Frontier models spend their extra milliseconds generating reasoning tokens. A 2.1s GPT-5 completion almost always beats an 800ms GPT-5-mini completion on hard tasks — the question is whether your task needs that depth.
- The 5× spread (180ms → 900ms p50) is the routing opportunity. Route simple chat and autocomplete to sub-300ms models; route math, code review and research to reasoning models. Cost-aware routing with latency awareness is the single biggest win for both UX and spend.
- p95 is the production number. A 600ms p95 (GPT-5-mini) means consistent feel under load; DeepSeek R1's 5,000ms p95 means your worst-case user waits 5 seconds for a first token — fine for deep research, unacceptable for a chatbot.
- Deeper dive: our engineering guide AI API Latency Optimization covers prompt caching, streaming and routing strategies in practice.
Streaming: why 180ms TTFT feels instant
Perceived latency is dominated by time to first token, not total completion time. With token-by-token streaming, the user sees output begin at TTFT and then reads along as tokens arrive at 50–150 tokens/second. A 180ms TTFT response — under a quarter of a blink — lands before the user has finished reading the prompt they just sent.
What the numbers say
- 180ms + streaming = instant. GPT-5-mini's 180ms p50 TTFT is below the ~250ms threshold where humans perceive UI as "immediate". Combined with streaming, the full 800ms completion is invisible — users experience an instant reply.
- Completion time only matters when streaming is off. Non-streaming clients pay the full p50 completion as dead time: 4.5s for Claude Opus 4, 8.2s for DeepSeek R1. Every serious integration should stream.
- Reasoning models stream their thinking. DeepSeek R1's 8.2s completion is experienced as a live reasoning trace — users watch the model work, which reads as engaged, not slow.
- Tail latency compounds at scale. A p95 of 5,000ms (DeepSeek R1) under burst load can turn into 10s+ timeouts without streaming; streaming converts worst-case waits into visible progress.
- Routing for feel: sub-300ms p50 models for autocomplete and chat; sub-1,500ms p95 for interactive tools; anything goes for batch and background jobs.
Dataset notes & methodology
How this benchmark is built
- Scope: 12 models from 8 providers — OpenAI, Anthropic, DeepSeek, Google, Meta, Alibaba, Moonshot AI, Zhipu AI, xAI.
- Metric: TTFT = time from request send to first received token. Completion = time to full response, including reasoning tokens where generated.
- Measurement: p50 and p95 across a sampled request mix on identical hardware, streaming enabled, single region. Provider load and network geography shift absolute numbers; relative ordering is stable.
- Cadence: refreshed weekly; last verified 2026-08-16.
- Cost companion: see the Model Price Index for the same models' USD per 1M tokens.
- Reseller note: DrAI offers all of these models at one flat subscription price — no per-token metering, no surprise bills, and the same streaming API for every model. See pricing.
Every model above. One flat price.
Stop metering tokens and watching latency. DrAI bundles GPT-5, Claude Opus 4, DeepSeek R1 and 12+ more models into a single subscription with one unified streaming API.
See DrAI pricing
Start free — no card needed