An Operator's Cut with Nufar Gaspar on the term that's dominating every enterprise AI conversation: the token. From the all-inclusive era to token-maxing leaderboards to today's anxiety-driven self-censorship, this is the primer on what a token actually is, why the bill is so hard to read, and how to spend wisely rather than sparingly.
Spend tokens wisely, not sparingly — and measure cost per accepted task, not cost per token.
After the all-inclusive, token-maxing, and now token-anxious eras, the goal is the token-smart era. Understand what a token really is, recognize that tokens aren't born equal across labs and layers, and stop reading the bill in tokens — which are an unmeterable moving target — and start reading it in dollars per accepted task. Kill the tokens that spin, tune the tokens that produce, and fearlessly defend the tokens that teach.
Practitioners feel watched when they use an expensive model, regular users wonder if one ambitious prompt will eat their weekly allowance, and leadership sees a bill growing faster than expected. Nufar's aim is to flip that anxiety into literacy — understand what the bill means, then spend wisely rather than sparingly.
The proposed metric is built around one question: what does each successful task actually cost? That reframes tokens away from a purely financial number and toward where usage creates value versus where it quietly leaks it.
We started token-oblivious in the all-inclusive era, where flat subscriptions hid the meter. Then came token-maxing (usage as a badge of AI maturity), then today's token-anxious backlash — and the goal now is the token-smart era.
Meta tracked employee AI usage on an internal leaderboard, burning 60-74 trillion tokens in a single month with its top user hitting 280 billion (~2.3M books). Uber blew through its entire 2026 AI coding budget in four months, and one unnamed company ran up a $500M Claude bill with no limits in place.
NLW: the hand-wringing over gamed leaderboards is overwrought — of course people game systems with real stakes. He'll bet any amount that a company wildly overspending via a token leaderboard will be farther ahead than one that underspends out of fear of proving ROI.
Meta went from a leaderboard to a memo constraining AI usage — the press now calls it token-minimizing — and Uber caps employees at 1,500. When employees self-censor, every prompt becomes an ROI conversation, which is exactly the wrong era to stay in.
— Nufar Gaspar. Nufar's framing for why token anxiety is dangerous: self-censorship pushes your most capable people back toward low-stakes work instead of the high-value workflows that actually move the needle.
A token is a chunk of text — bigger than a character, usually smaller than a word — that the model reads and writes. In English the ratio is roughly three-quarters of a word per token, so a page of text is about 1,000 tokens. OpenAI's public tokenizer page lets you see how your own text gets chunked.
Because billing is per token, languages like Hindi, Thai, and Greek can generate two to five times more tokens for the same content — so the same question costs much more in another language. Code has its own quirks, with indentation, brackets, and whitespace all becoming tokens.
Models fail to count the R's in 'strawberry' because they never saw the individual letters — just 'straw' and 'berry' as tokens. Many of the 'AI is so dumb' memes are literally just tokenization artifacts.
An email is ~500-700 tokens (about half a cent — nobody should ration those), a page is ~1,000, images run just over 1,000, and deep research can hit tens or hundreds of thousands. One learner accidentally sent a yes/no question to a deep-research tool that spawned ~100 sub-agents and cost over 4 million tokens.
The model doesn't remember prior messages, so it re-sends the entire session each turn. By turn ten it may be reprocessing so much earlier context that total tokens grow far faster than the number of turns suggests — immortal sessions quietly balloon the bill.
NLW: the direction of travel is clear — more advanced, more useful work requires more intelligence and thus more tokens. Unaddressed token anxiety incentivizes people to stay swimming in daily emails instead of doing the deep research and agentic work leaders actually want.
Agents work autonomously in loops, and a typical task involves 10-20 model calls carrying instructions, history, tool definitions, and prior results. McKinsey estimates ~60% of an agentic task's cost is tied to checking, refining, and regenerating answers — the expensive part is getting from the answer to the accepted result.
Every lab has its own tokenizer — OpenAI's vocabulary is ~200k, Gemini's ~256k, Llama's about half, Claude's unpublished. Since price per million is denominated in each lab's own tokens, the same document can cost 10-20% more on one provider than another, making the sticker price meaningless for comparison.
When Opus 4.7 shipped in April, the per-million price sheet was identical but a new tokenizer produced ~30% more tokens for the same text. Independent analysis of over a million requests found native token counts up 32-45%, with real-world bills up 12-27% after caching. Same sticker, smaller candy bar.
Input tokens (prompt, history, files, tool definitions) are cheapest but accumulate fast; reasoning tokens are the model's invisible internal thinking, billed at the high output rate and adding 4-20x; output tokens are what you see, typically 3-5x pricier than input. The reasoning layer — like 'kitchen time' on a restaurant bill — catches everyone by surprise.
High reasoning effort can drive 10-12x more tokens, but for simple questions lower reasoning often produces a lower cost per task and better quality by preventing the model from overthinking. The GPT-5-class gap is widening too — $10 per million input vs. $50 output.
Databricks tested coding agents on its own code base: Sonnet 5 was 1.7x cheaper per token than Opus 4.8, but cost ~$2.09 per task versus $1.94 for Opus because it needed more iterations. The lesson: reach for the right model, not the cheapest. Different agent harnesses at the same quality also showed a 2x+ difference in cost per task.
Include retries, review, and every correction, then divide by accepted results — that's your cost per accepted task. The practical audit: run 5-10 representative tasks through your model and tool options, hold input and quality constant, and compare first-pass success, human correction, elapsed time, and total cost.
Tokens that teach (experiments, failed workflows, identity files, memory) are tuition and should be defended fearlessly. Tokens that produce ship real work. Tokens that spin — machines talking to themselves, idle agents, bloated context, wrong-model use — are activity without output. The token-smart move: kill the spin, tune production, protect teaching, in that order.
Her disabled 'Chloe' chief-of-staff agent ran on the Anthropic API with bills going to a secondary inbox. Traveling and not using it, she still got charged — $1,500 in two weeks, ~400 million tokens in and near-zero out (a ~2,600:1 ratio) from cron jobs compacting empty sessions every 30 minutes. If it happened to an AI expert, it can happen to anyone whose company owns the card.
NLW ran an OpenClaw agent perpetually researching new AI adoption data sources; it did exactly what it was told but simply wasn't valuable enough for the cost. Auditing spin means catching processes that accidentally became spin — like hourly Slack miners or a morning brief nobody reads — not just outright errors.
Watch idle or over-frequent agents, unused automations, the 'pre-prompt tax' of always-on rules and tool definitions, immortal conversations, unfiltered data retrieval (500 rows when you needed 20), scattered context, and rework loops. Diagnostics include the weekend test, extreme input-to-output ratios, and spend rising while value stays flat.
New task means new session; be intentional about model choice; rightsize context (Goldilocks, not too much or little); build reusable skills and automations instead of ad hoc re-asking; filter everything you can; and kill jobs early when the model's reasoning shows it's off, rather than letting it spin.
Anthropic's Claude Code /doctor command audits token distribution, stale skills, redundant tool configs, and overly long or overlapping instructions, returning concrete kill/reduce recommendations. Non-Claude users can have their AI recreate the same audit — it isn't complex.
NLW expects routing norms to be solved first for deterministic software engineering and to stay far messier for knowledge work, where preference is about the nature of a response, not benchmark scores. The GPT-5 auto-router frustrated super users — proof people want control — so understanding model capabilities remains a high-leverage skill.
The point isn't only driving the bill down but improving return — and that often means spending more on tokens that teach: running the same task three ways, experimenting with tools and skills, and building context so the model gives personalized, organization-aware results. A 20k-developer study found the heaviest AI users were roughly twice as productive.
Make consumption visible and teach people to spend smartly, not sparingly. Budget by workload and individual — someone building reusable skills for the whole team deserves far more than someone using AI as an extended Google — and audit on a schedule for stale automations and instructions.