Fastest LLM APIs: 13,000 Tokens/Sec Benchmarked

Data as of October 20, 20261 source

Drafted with AI tools and edited by a human. The figures and conclusions were checked by a GPUniq editor.

OpenAI GPT-5 Nano hits 13,060.9 tokens per second (t/s). This model is 100 times faster than the median speed of 130.1 t/s found in this 16 model benchmark. Grok 4.20 reaches 1,634.2 t/s, but flagship models like GPT-6 Astra crawl at 53.6 t/s to prioritize reasoning.

Leaders in the 13,000 t/s race

Throughput is dominated by OpenAI. GPT-5 Nano manages 13,060.9 t/s, while the balanced GPT-5 follows at 8,208.4 t/s. xAI's Grok 4.20 is a distant third at 1,634.2 t/s.

Large gaps exist between these high speed models and flagship versions. GPT-6.1 Sol generates 354.5 t/s. Claude Sonnet 5.5 produces 277.5 t/s. These figures represent raw generation speed for high volume data processing.

Model	Vendor	Speed (t/s)	Input ($/1M)	Output ($/1M)
GPT-5 Nano	openai	13060.9	0.04	0.32
GPT-5	openai	8208.4	1.00	8.00
Grok 4.20	xai	1634.2	1.00	2.00
GPT-6.1 Sol	openai	354.5	1.60	8.00
Claude Sonnet 5.5	anthropic	277.5	1.60	8.00
GPT-4.1	openai	194.2	1.60	6.40
Qwen3.8 Max	qwen	41.2	1.80	5.40

Benchmark scope and Groq results

Groq was not included in this 16 model benchmark set. OpenAI achieved the 13,000 t/s milestone using the GPT-5 Nano architecture. This model uses a fast category optimization to prioritize throughput.

GPT-5 Nano maintains a 400k context window while delivering these speeds. It is built for tasks where latency is the primary constraint. Flagship models like GPT-5.6 Sol drop to 128.7 t/s, showing that speed often comes at the cost of model complexity.

Together AI and Fireworks comparison

Neither Together AI nor Fireworks AI were part of the 16 models measured. This analysis focuses on OpenAI, Anthropic, Google, xAI, and Qwen. OpenAI dominates the high speed tiers among these providers.

The median speed across all measured models is 130.1 t/s. Most providers cluster around this mark for standard models. GPT-4o sits at 131.6 t/s. If you need speeds significantly higher than the median, you must look at specific fast category models like GPT-5 Nano or Grok 4.20.

API latency and streaming performance

High throughput usually results in low per-token latency for streaming applications. GPT-5 Nano is the fastest option. Grok 4.20 is the next best choice, offering 1,634.2 t/s with a 2,000k context window.

Gemini 3.7 Flash provides 161.4 t/s. This is faster than GPT-4o but significantly slower than the top OpenAI models. For applications requiring sub-100ms response times during long generations, fast category models are the only viable choice.

Open source vs proprietary speed

Qwen3.8 Flash hits 75.3 t/s. This is well below the 130.1 t/s median. Qwen3.8 Max is the slowest in the 16 model set at 41.2 t/s, trailing even the heavy Claude Opus 4.8 flagship which reaches 58.7 t/s.

Speed tiers and category deltas

Measuring the delta between categories reveals massive performance shifts. GPT-5 Nano is 101 times faster than the flagship GPT-5.6 Sol. Flagship models like GPT-6 Astra (53.6 t/s) and Claude Opus 5.5 (80.6 t/s) are optimized for output quality rather than velocity.

Best price to speed ratios

OpenAI's GPT-5 Nano is the cheapest high speed option. At $0.04 per 1 million input tokens and $0.32 per 1 million output tokens, it provides the highest measured speed at the lowest price point in the dataset.

Qwen3.8 Flash is also inexpensive at $0.135 for input and $0.423 for output. However, its speed is only 75.3 t/s. OpenAI fast models currently offer the best ratio of tokens per second per dollar.

Future of high throughput models

Industry trends show a divergence between models for human interaction and those for machine to machine workflows. We are seeing a jump from the median speed of 130.1 t/s to GPT-5 Nano's 13,060.9 t/s.

High throughput is becoming a requirement for large context windows. Grok 4.20's 2,000k window needs this speed to be usable. Processing millions of tokens in seconds is the next benchmark hurdle.

Methodology

Speed is output tokens per second measured on the GPUniq gateway for real requests, as a moving average; models with too little traffic to measure are left out rather than estimated. Prices are GPUniq's public retail rates per million tokens, read from the live model catalog at the moment of the snapshot. Ruble figures use the rate GPUniq applies to ruble top-ups at that moment. The discount is against the model vendor's own list price.