INTERACTIVE ESSAY / BY REMCO VROOM
The Token Gap
You read this sentence at about four tokens per second. Gemini would have finished the whole article before you reached the comma.
Read the film transcript
A thousand tokens. Play this illustration: the fastest AI takes two seconds while your speaking lane crawls.
Human speech takes two hundred and fifty seconds, an enormous gap. Pause, restart and open the numbers behind it.
The real prize is tokens per insight: explore The Token Gap with AGENTIC. Uncovering how AI works.
You are reading this at roughly four tokens per second.
That is not a metaphor. Human speech runs about 150 words a minute, which converts to three or four tokens per second in the same units we use to measure machines. Think silently and you speed up, maybe to ten. That is your ceiling. It was your grandmother's ceiling. It was the ceiling of the first person who ever explained something to another person.
Claude Fable 5.1 runs at 67 tokens per second. GPT-6 Astra at 60. Gemini 3.5 Flash pushes close to 500. Groq's silicon holds 300-plus with a first token back in under 200 milliseconds.
So the frontier models think out loud between fifteen and twenty times faster than you talk. The fast ones, over a hundred times faster.
Most people have never done this arithmetic. They feel the speed as convenience. It is not convenience. It is a different physics of thought.
AN ILLUSTRATION IN REAL TIME
One thousand tokens.
The same finish line.
Historical, illustrative author assumptions, not live benchmarks. Model output and human token equivalents are different measures. This race compares their assumed pace, not intelligence.
The first lane finishes in about two seconds. At the assumed speaking pace, the same distance takes 250.
Read the numbers without the animation
Gemini 3.5 Flash497 tokens/s · 1,000 tokens in 2.0 seconds
Groq on Llama315 tokens/s · 1,000 tokens in 3.2 seconds
Claude Fable 5.167 tokens/s · 1,000 tokens in 14.9 seconds
GPT-6 Astra60 tokens/s · 1,000 tokens in 16.7 seconds
You, thinking10 tokens/s · 1,000 tokens in 100.0 seconds
You, speaking4 tokens/s · 1,000 tokens in 250.0 seconds
The illustration runs for 22 seconds. At that point the thinking assumption reaches 220 tokens and the speaking assumption reaches 88. It excludes first-token latency and changes in throughput.
The number that should worry you is not the speed
At maximum effort, Astra solves a task using around 27,000 output tokens. Fable 5.1 reaches the same score using 78,000. Same destination. Three times the thinking.
Read that again, because it inverts the whole conversation. The models are not competing on speed. They are converging on speed, all of them sitting between 52 and 75 tokens per second at the top end. What separates them is how much reasoning they burn to arrive.
Which means the real currency is not tokens per second. It is tokens per insight.
And that is the one race where humans are still, for now, absurdly good. You do not need 78,000 tokens to decide whether to hire someone, enter a new market, or back a partner. You need a walk around the block and a hard conversation. Compression, not throughput.
The flip: everyone is anxious about being outrun, because speed is the only dimension you can actually feel. You watch the tokens arrive. You cannot watch the reasoning that produced them. The competition was never the speed. It was the efficiency.
Except the efficiency is moving too
Human reasoning speed has been flat for a hundred thousand years. No version bump. No effort setting. The wetware you run on today shipped in the Pleistocene and has had no meaningful throughput upgrade since.
The machines are on two compounding curves at once.
Moore's Law gave us transistor density doubling roughly every two years. Steady, predictable, the curve that built the modern economy. Huang's Law, coined at Nvidia's 2018 developer conference, is the claim that GPU performance for AI workloads is climbing considerably faster than that, because the gains stack: silicon, architecture, software, precision, decoding strategy, all improving together.
Put those two lines on the same chart and Moore's looks almost conservative. Now add a third line for human cognition and you cannot draw it. It is flat. It has always been flat. It is the x-axis.
THE AUTHOR’S SECOND ILLUSTRATION
Two curves bending.
One line refusing to.
Historical claims and extrapolated trends, indexed to 1× in 2013. The human baseline is a rhetorical illustration, not a measurement of cognition.
2023 claim · ~130,000× projected in 2030
2026 illustration · ~360× projected in 2030
Held constant by the author for this comparison
And the efficiency gap is closing from the other direction. Astra needing a third of Fable's tokens for the same answer is not a hardware story. That is the models getting better at compressing thought. The thing we told ourselves was our edge.
KEEP THE CONTEXT VISIBLE
Sources & assumptions.
The supplied draft names these sources and records a retrieval date of 15 September 2026. Links below provide attribution and context; they do not independently validate every historical input in the essay.
Back to the essay contents- Artificial Analysis model leaderboard, opens in a new tab
The author's attribution for model throughput and token-efficiency comparisons. The live leaderboard changes; the draft's exact historical figures have not been independently reproduced here.
- Groq model documentation, opens in a new tab
Provider reference for model throughput. The draft does not identify the Llama variant or test configuration behind its 315 tokens-per-second assumption.
- Google: Gemini 3.5, opens in a new tab
Provider context for Gemini 3.5 Flash. The race retains the author's 497 tokens-per-second input; this link is not a reproduction of that measurement.
- NVIDIA: Bill Dally, Hot Chips 2023, opens in a new tab
The NVIDIA hardware-performance claim cited by the author, alongside the draft's reference to GTC 2018.
- Epoch AI: Trends in GPU price-performance, opens in a new tab
Marius Hobbhahn and Tamay Besiroglu's 2022 research offers the contrasting price-performance estimate cited in the draft. Hardware performance and performance per dollar are different measures.