Model Independent AI API monitoring
How fast are AI models today?
Compare real-world inference speed across leading AI models. Track time to first token, generation throughput, latency and uptime — in one transparent dashboard.
OpenAI · Anthropic · Google · More providers coming soon
Performance overview
AI inference at a glance
Latest measurement Sun, 01 Nov 2026 21:00:08 GMT
Speed leaderboard
Fastest AI models
Ranked by median output generation speed. Compare first-token latency and total response time.
| Rank | Model | Tokens/sec ↓ | TTFT | Total time | Uptime |
|---|---|---|---|---|---|
| — | GPT-6 Astra insufficientOpenAI · gpt-6-astra | 38.3 t/s | 2.84 s | 6.50 s | 100.0% |
| — | GPT-6 Luna insufficientOpenAI · gpt-6-luna | 113.2 t/s | 1.71 s | 3.52 s | 100.0% |
| — | GPT-6.1 Sol insufficientOpenAI · gpt-6.1-sol | 70.5 t/s | 1.26 s | 3.30 s | 100.0% |
Window: 24h · 3 models tracked
Historical monitoring
Performance over time
Median throughput of the current speed leader. Per-model history lives on each model page.
Output generation speed
median tokens/sec
Loading…
What do these metrics mean?
Understanding inference performance
TTFT — Time to first token
How long until the model begins returning generated text. Lower is better.
TPS — Tokens per second
How quickly the model streams its output after generation begins. Higher is better.
Total response time
Elapsed time from request start until the response completes. Depends on output length.
Uptime
Share of scheduled test attempts that succeed within the published timeout.
Why ModelSpeedTest
Built for transparent AI performance comparisons
⚡
Streaming-based timing
Measure first visible text, output throughput and end-to-end response time using provider API streams.
◎
Repeatable test protocol
Versioned prompts, documented settings, fresh contexts and repeated samples for comparable measurements.
▥
Historical trends
Explore 24-hour, 7-day and 30-day changes, including performance variance and service interruptions.
Our methodology
How AI model speed testing works
ModelSpeedTest uses official provider APIs and scheduled cloud probes. Each run sends an identical, versioned prompt with streaming enabled and documented settings.
01
Send a standardized request
A versioned, simple English text-generation prompt is sent with streaming enabled and documented model settings, with no conversation history.
02
Measure the response
Each probe records time to first generated text, generation duration, total response time and token usage where available.
03
Publish comparable results
Measurements are summarized using medians, with region, exact model ID, sample count and staleness disclosed. See full methodology JSON →
API performance is not identical to consumer chat-app performance. Network routes, reasoning settings, output length and provider token accounting can affect comparisons. Scheduled probe regions are observed, not geographically guaranteed, and measurements from different regions are never mixed.
Frequently asked questions
AI model speed test FAQ
What is TTFT in AI?
Time to first token is the delay between sending a request and receiving the first generated text token. It reflects perceived responsiveness but also includes network and request-processing time.
What are tokens per second?
Tokens per second is the rate at which a model generates output tokens. Visible output throughput should be distinguished from billed output tokens and any hidden reasoning tokens.
Does a faster AI model reason better?
No. Inference speed and reasoning accuracy are different dimensions. This website focuses on speed, latency and availability rather than reasoning scores.
Are these measurements from ChatGPT or Claude's websites?
No. The benchmarks use official model APIs, not the consumer chat interfaces. Results may differ from what you experience in an app.
How often do results update?
Every 30 minutes per model, subject to API limits, cost caps and deployment configuration.
Can token count show which AI model is greener?
Not by itself. Energy consumption also depends on hardware, model architecture, batching, input processing, datacenter efficiency and electricity supply.