AI Speed Test — Which AI Is Fastest Right Now?

AI speed terms, explained

Tokens per second (throughput)

Tokens per second, or throughput, is how many tokens a model produces each second once it has started answering. A token is roughly three quarters of an English word, so 300 tokens per second is about 225 words per second — faster than you can read. It is the number that decides how quickly a long answer finishes.

Time to first token (TTFT)

Time to first token is the pause between sending a prompt and seeing the first word appear, measured in seconds. It is what makes a model feel snappy or laggy, no matter how fast it then types. A model can have a quick first token but slow typing, or the reverse — which is why this tool shows both numbers instead of a single “speed”.

What is a good tokens-per-second speed?

For everyday chat, anything above roughly 50 tokens per second feels smooth, and 200 or more feels genuinely fast. Specialised inference chips — Groq’s LPU and Cerebras’s wafer-scale engine — push open models past 1,000 and even 2,000 tokens per second. Below about 20 tokens per second a long reply starts to drag. The right speed depends on the job: a short chat reply and a 2,000-word draft have very different patience thresholds.

Does faster AI mean better AI?

No. Speed measures how quickly text is produced, not how good it is. A very fast model can still be weaker at hard reasoning, coding or factual accuracy than a slower one. Use a speed test to shortlist models that are responsive enough for your use case, then judge answer quality separately — ideally on your own real prompts.

Why AI speed changes through the day

The same model can be fast one hour and slow the next. When many people hit a provider at once, requests queue and throughput drops; a new model version, extra hardware or a quieter time of day can push it back up. Your own distance from the server and your network add more variation. That is why a live reading beats a fixed number, and why any benchmark should be read as a snapshot under specific conditions.

How the AI speed test works

Every large language model has two different kinds of speed, and this tool measures both. The first is time to first token (TTFT) — the pause between sending your prompt and the moment the first word appears. The second is throughput, measured in tokens per second — how quickly the rest of the answer streams out once it has begun. A token is roughly three quarters of an English word, so a model running at 300 tokens per second is writing about 225 words every second.

Behind the scenes, a small script sends the same short prompt to each model on a schedule, records the two timings, repeats it a few times and keeps the median so a single slow network blip cannot distort the result. Those numbers are combined into a single 0–100 speed score: most of the score comes from throughput, with a bonus for a fast start. The score drives the traffic light — green means the model is quick right now, amber means average, and red means it is running slow compared with its usual pace.

Speeds genuinely move through the day. When millions of people hit a provider at once, queues build and throughput drops; when a lab ships a new model version or adds hardware, it can jump again. That is exactly why a live reading is useful: the model that was fastest this morning may not be fastest this evening. The chart above keeps recent history so you can spot the daily pattern and pick a quiet moment for a big job.

Because free API tiers cover open models plus Google's Gemini, that is the current line-up — Llama, GPT-OSS, Qwen, DeepSeek and Gemma running on speed-focused providers like Cerebras and Groq (whose custom inference chips top the board), alongside Gemini. The two names most people ask about, ChatGPT and Claude, are paid APIs and are not in the free line-up yet; they can be added later. Whatever the mix, the method stays the same, and everything you see is a measured value, never a guess.

Independent estimate for general information only. This page is not affiliated with any AI provider, and speed does not equal quality — a fast model is not automatically the most accurate one for your task.

The fastest LLM in 2026: how the speed landscape looks

Ask “what is the fastest LLM right now?” and the honest answer is that it depends on the hardware behind the model, not just the model itself. In 2026 the LLM speed leaderboard splits into two camps. Specialised inference providers — Groq, with its Language Processing Unit, and Cerebras, with a wafer-scale engine — routinely stream open-weight models such as GPT-OSS, Llama and Qwen at 750 to over 2,000 tokens per second with time to first token under 100 milliseconds. General GPU clouds serving frontier models like Gemini, GPT and Claude usually sit lower on raw output speed because their hardware juggles many different workloads at once.

That is why a fastest-LLM ranking is really a fastest-provider ranking: the same open model can be two to twenty times quicker on a speed-focused chip than on a shared GPU. The live board at the top of this page reflects exactly that — it re-measures each model's tokens per second and TTFT about every half hour, so the order shifts as provider load rises and falls through the day. For a broad, quality-plus-speed view across a hundred-plus models, independent leaderboards such as Artificial Analysis are the reference standard; this tool's niche is the opposite, a small, live, free-tier speed check you can read in a single glance.

Reasoning models add a twist. Models that “think” before answering — including several 2026 Gemini and GPT-OSS variants — spend seconds generating hidden chain-of-thought tokens before the first visible word, so their time to first token can run to ten seconds or more even when raw typing speed is high. A fast typer is not always a fast responder, which is the whole reason this tool reports first-token delay and throughput as two separate numbers instead of one blended “speed”.

Which LLM is fastest for your use case?

“Fastest” means different things for different jobs, so the metric you optimise for should match how people will actually use the app. Human-perception research and MLCommons' MLPerf inference benchmark point to the same thresholds: a time to first token under 500 milliseconds feels instant, above one second is noticeable, and beyond two seconds users start to lose patience — while a typing speed of roughly 30 tokens per second (about reading speed) is already fast enough that extra output speed becomes hard to perceive on screen.

• Chatbots and assistants — prioritise low TTFT. The reply should start almost instantly; once it is streaming, moderate throughput already outpaces reading. Aim for a first token under about half a second.
• Voice agents and real-time apps — the strictest case. Both first-token delay and steady throughput matter, because a spoken pause is far more jarring than a written one; sub-300ms TTFT is the target the fastest providers chase.
• Coding assistants — a mix. You want a quick first token for inline suggestions but also high tokens per second so a long file or refactor completes without a wait.
• Batch, drafting and agents — throughput wins. When work runs in the background or an agent loops through many calls, total tokens per second and cost per million tokens matter far more than the first-token delay.

Use the live board to shortlist models that clear the bar for your case, then confirm answer quality on your own prompts — because the fastest LLM for streaming a chat reply is rarely the same choice as the fastest LLM for a heavy overnight batch job.

LLM API speed vs cost: finding the best value

Raw speed is only half the decision — the other half is price per million tokens. The cheapest LLM API on a list is not always the cheapest in production, because output volume, cached input, retries and answer quality all move the real bill. This tool's best-value figure divides a model's tokens per second by its published output price, so you can spot which providers give the most speed per dollar rather than just the lowest sticker price.

For testing and low-volume projects, several providers offer genuinely free LLM API tiers — Groq, Cerebras, Google's Gemini and OpenRouter among them — typically with a daily token allowance and a per-minute request cap. That is exactly the line-up this speed test uses, which is why every measurement here costs nothing to produce. When a project outgrows the free tier, the fast open-model providers usually stay ten to fifty times cheaper per token than frontier APIs, at the cost of a smaller model catalogue and no fine-tuning.

The practical rule: benchmark speed on the free tier first, check the best-value column, then confirm the current price on the provider's own pricing page before you commit at scale — headline rates change often, and long-context or reasoning workloads can quietly multiply the real cost.