API Response Time Comparison

Compare API latency and response times across different LLM providers and models.

Note: These are approximate values based on typical performance. Actual latency varies by region, load, and network conditions. Values represent streaming mode performance.
How many tokens you expect in the response
RankModelProviderFirst TokenPer TokenTotal TimeTokens/Second
#1
Gemini 3 Flash
Google170ms16ms
8,170ms
8.17s
61.2
tok/s
#2
Grok 4 Fast
xAI190ms19ms
9,690ms
9.69s
51.6
tok/s
#3
Claude Haiku 4.5
Anthropic200ms20ms
10,200ms
10.20s
49.0
tok/s
#4
GPT-5 mini
OpenAI230ms22ms
11,230ms
11.23s
44.5
tok/s
#5
Claude Sonnet 4.6
Anthropic380ms40ms
20,380ms
20.38s
24.5
tok/s
#6
GPT-5.1
OpenAI400ms45ms
22,900ms
22.90s
21.8
tok/s
#7
Gemini 3 Pro
Google440ms50ms
25,440ms
25.44s
19.7
tok/s
#8
GPT-5.5
OpenAI450ms50ms
25,450ms
25.45s
19.6
tok/s
#9
Claude Opus 4.8
Anthropic480ms55ms
27,980ms
27.98s
17.9
tok/s
Fastest Model
Gemini 3 Flash
8,170ms total
Avg Response Time
17938ms
Across all models
Output Tokens
500
Expected generation

Latency Optimization Tips

  • Use streaming mode to show partial results faster and improve perceived performance
  • Choose models with lower latency for real-time/interactive applications
  • Consider using faster models (like Gemini 3 Flash or Claude Haiku 4.5) for non-critical tasks
  • Implement caching for common queries to bypass API calls entirely
  • Deploy in the same region as your API provider for lower network latency

Comparing LLM latency

For anything interactive — chat, autocomplete, voice — latency matters as much as price. Two numbers drive the experience: time to first token (how long before the response starts streaming) and per-token speed (how fast the rest arrives). Smaller models are dramatically faster on both, which is why they're often the right choice for user-facing features even when a larger model is slightly more accurate.

This tool estimates the total response time for a given output length across models, so you can see the speed-versus-capability trade-off in concrete terms. Latency varies with region, server load, prompt size, and reasoning effort, so treat these as representative benchmarks rather than guarantees — but the relative ordering between models is a reliable planning guide.

How to use this tool

  1. Enter how many output tokens you expect the response to contain.
  2. Compare first-token latency, per-token speed, and total time across models.
  3. Favor a faster model for real-time, user-facing features; reserve slower frontier models for background reasoning.

Frequently asked questions

What is time to first token?+

The delay between sending your request and receiving the first piece of the response. It's the number users feel most in chat interfaces, because it's how long the screen sits empty before text starts appearing.

How can I make responses feel faster?+

Stream the output so text appears as it's generated, use a smaller or faster model for interactive paths, and keep prompts and outputs concise. Streaming alone makes a big perceived-speed difference.