Someone in the meeting says it: "AI is a hundred times faster than us at this." Heads nod. The plan is written.
Nobody asked the only questions that make the number mean anything. Faster than a person doing what, measured how, and as of when? Without those three answers, "a hundred times faster" is not a number. It is a vibe in a suit, and it will silently price every decision in the plan.
This post is the discipline, plus the current board of numbers worth knowing, each carrying its date like a passport.
The board, as of spring 2026
| Capability | Number | Source and date |
|---|---|---|
| Human typing | About 40 words a minute average, about 100 for the fast | Standard measures, stable for decades |
| Human reading | About 238 words a minute, silent reading | Brysbaert meta-analysis, 2019 |
| Machine writing | About 280 tokens a second on mainstream APIs; 476 on Groq-class hardware; 2,000 to 3,000 on wafer-scale inference | Published inference benchmarks, 2026 |
| Task horizon | Length of task completed unaided doubling roughly every 7 months, possibly faster for the newest frontier models | METR, March 2025, updated 2026 |
| The inversion | Experienced developers took 19 percent longer with AI tools on real repositories | METR randomized trial, early 2025 |
| Swarm gain | Multi-agent system beat a single agent by 90.2 percent on broad research evaluations | Anthropic Engineering, 2025 |
| Swarm bill | About 15 times the tokens of a single chat for the same task | Anthropic Engineering, 2025 |
| Parallel voice | One thousand simultaneous calls at a few hundred milliseconds of latency | Voice agent platforms, 2026 |
| Cost per call | Roughly 0.30 to 0.50 dollars with AI against roughly 6 to 7 with a human | Voice platform pricing, 2026 |
| Price trend | Inference cost for constant capability falling about 10 times a year | a16z, November 2024 |
Read the table twice and notice the two rows that refuse to obey the speech. The 19 percent row and the 15x row are in the same board as the 90 percent, because they are all true at once. A board with only good news is a brochure.
Why the date is the number
Capability numbers here move on a doubling clock, not a quarterly clock. A benchmark quoted without a date is worse than stale. It is unfalsifiable: nobody can check it, including the person who said it.
Three habits make numbers usable in plans:
- Attach the source and the date, out loud. "Agents complete longer tasks every seven months, METR, March 2025" is a fact you can re-check. "AI is improving fast" is a mood.
- Name the comparison class. Faster than a person typing? Than a person thinking? Than your team's actual current process, including its meetings? Those are three different multiples, and plans built on the wrong one fail quietly.
- Give it a shelf life. In this field, a number older than a year is history, and a number older than two is decoration. Write the expiry next to the figure, and re-check the load-bearing ones on a calendar, not on a hunch.
The row that spoils the speech
Every table of AI multiples should carry one row of embarrassment, and the 19 percent is the honest one. In METR's randomized trial, experienced open-source developers working on real repositories finished their own familiar tasks slower when using AI tools. Early 2025, confidence interval 2 to 39 percent slower.
It does not say the tools are weak. It says the multiple depends on where the speed lands. Speed that arrives at the wrong node of a workflow does not multiply the workflow. It multiplies the queue, the review load, and the cleanup. That single row is why capability tables belong in planning meetings next to constraint analysis, and why the review layer is where the money goes is the sentence that follows every multiple.
Numbers in plans, not in speeches
The test of a number is whether it survives contact with a plan:
- "A thousand calls a minute" becomes real when it names the platform, the date, and the cost per call, and the plan shows who reads the transcripts.
- "Ten times cheaper every year" becomes real when the plan prices next year's experiments at next year's prices and still finds them worth doing.
- "It doubles every seven months" becomes real when the plan asks what the venture looks like at double, and at double again, because two doublings are fourteen months.
A number that cannot be converted into a check is a speech. A number with a date, a source, and a decision attached is evidence, and evidence still decides, even when it moves monthly.
FAQ
These numbers will be wrong by next year. Why publish them?
Because they are useful now and wrong later, and knowing which is which requires the dates. A dated table teaches the discipline this post is about: quote, source, expiry, re-check. The table is expected to age. That is its feature.
Which of these numbers should a founder actually track quarterly?
Three. The task horizon, because it sets what can be delegated to agents at all. The cost per unit of work in your own operation, because unit economics compound quietly. And your own team's constraint location, because the multiples only matter where the queue is. The rest is scenery until one of the three moves.
Is the 19 percent study not just early tooling being immature?
Partly, and that is the point of dating it. It measures early 2025 tools on mature developers' own codebases. Whether the number has improved since is exactly the kind of question a dated number lets you ask again. What does not age is the lesson: multiples are conditional on where the speed lands.
How do I stop my team from quoting undated numbers?
Ask the two questions in public, every time, without tone: measured how, and as of when? Teams stop quoting vibes the third time the questions get asked. Start doing it yourself, especially when the number supports your own plan. That is what makes it a norm instead of a politics.
Related reading
- The Desk Assumption
- Ten People, One Hundred Agents
- From Pull Requests to Proof
- Validation at Machine Speed, Evidence Still Decides
- Confident Wrong Output Marked Done Is the Agentic Failure Mode
- Human Judgment Is the Scarce Layer
- The AI Pressure Map: Twenty Shifts, Four Zones
If you want your plans built on dated evidence instead of meeting-room multiples, book a discovery call.