Source: raw/Everything_You_Need_to_Know_about_AI_Tokens.md — Nofar Gaspar with Nathaniel Whittemore, The AI Daily Brief (“Operator’s Cut”), transcript fetched 2026-08-05. Refreshed 2026-09-29 with raw/OpenAI_Cuts_Off_Elon_s_Cursor_Humanity_s_First_Star_Probe_and_Trump_s_Nuclear_Mars_Ship_EP_285.md (Peter H. Diamandis Moonshots EP285, https://www.youtube.com/watch?v=JywXvB8PpTs) on outcome-based pricing. Refreshed 2026-11-01 with two Nate B Jones videos, raw/Enterprise_AI_Pricing_-_Why_Everything_You_Know_Is_Wrong.md and raw/Anthropic_Just_Filed_Its_S-1._Here_s_What_The_Headlines_Are_Missing..md.
The wiki holds several token-optimization articles (Claude Code token optimization, token-optimizer, cost intelligence levers) but no primer on what a token actually is and where the money goes. This fills that gap, and its most useful contribution is a rough order-of-magnitude ladder — because the practical mistake is almost never “my emails cost too much.”
Key Takeaways
- A token is a chunk of text bigger than a character and smaller than a word. It is the unit under every bill, quota, and rate limit. In English the ratio is roughly ¾ of a word per token, so a page of text ≈ 1,000 tokens.
- There is a language tax. Hindi, Thai, Greek and other non-Latin-script languages can consume 2–5× more tokens for the same content. Because billing is per token, the identical question costs materially more in some languages.
- Code tokenizes badly. Indentation, brackets, and whitespace all become tokens. Newer code-friendlier tokenizers exist, but numbers remain a problem — digit sequences get chopped mid-number.
- Most “AI is dumb” memes are tokenizer artifacts. The strawberry-letter-counting failure is the canonical case: the model never saw individual letters, only
straw+berry. Worth knowing before treating a tokenization quirk as a reasoning failure. - Conversations compound, and this is the quiet cost driver. The model has no memory between turns, so the entire prior session is resent with each message. By turn 10 the total has grown much faster than the turn count suggests — long sessions are expensive independently of what is in them.
- Agentic work is a different sport. Agents run autonomously in loops and consume, by widely-cited industry estimates, 5–30× the tokens of a simple chat; a typical agentic task involves 10–20 model calls each carrying instructions, history, and tool results. Poorly designed harnesses do worse.
- The leadership framing is the real argument. Token anxiety pushes people toward less sophisticated use cases — exactly backwards, since the high-value work is the token-heavy work. The goal is routing, not rationing.
The cost ladder
| Workload | Rough token cost | Note |
|---|---|---|
| Everyday chat / drafting an email | ~1,000 tokens, ~half a cent | ”Nobody should ration emails. It’s not where the money goes.” |
| Search / light research | Multiplies quietly | The band worth auditing for efficiency |
| Deep research | 70,000 to hundreds of thousands | Spawns subagents |
| Data analysis | 1M+ per task | |
| Heavy coding / agentic workflows | Comparable or worse |
Email → agentic coding spans a factor of ~1,000×.
A per-seat benchmark from inside a lab (added 2026-10-27). Epoch AI, reading OpenAI’s September research post, reports that the median OpenAI researcher’s daily coding-agent usage, valued at API prices, grew from under 600 a day by mid-August, and that 90th-percentile users spent **over 2.6-5.3 trillion a year in revenue (raw/newsletter-epoch-ai-b02c234fec.md; see Epoch compute economics).
The cautionary anecdote: a learner with a yes/no question reached for the deep-research tool instead of a web search. It spawned roughly 100 subagents and the answer cost over 4 million tokens. This is the shape of nearly every “our AI bill exploded” story — not steady overuse, but a cheap question routed to an expensive tool.
What this implies
- Routing beats rationing. The lever is making sure the right question reaches the right tool, not capping usage. Deep research where deep research is warranted; a web search — or Google — for a yes/no.
- Session length is a cost decision. Because context compounds, starting fresh is a cost control, and compaction/pruning features are economic features. Compare Hermes’s per-turn micro-compaction in v0.20.0 and Claude Code’s
/contextbreakdown. - Measure per task, not per token. The 5–30× agentic multiplier is only alarming without a task-value denominator — the framing used throughout cheap-executor delegation.
- Vendors may start billing that way (2026-09-29, host-reported). On Peter H. Diamandis’s Moonshots (EP285), Diamandis says Salesforce prices Agentforce “based on customer revenue generated, not tokens consumed.” He adds that OpenAI “has also done the same, letting some of their largest customers pay only when its AI actually completes the job,” with no charge for tokens, compute time or API calls. His framing: a company selling tokens is “selling you compute”; one selling results “is selling you labor.” No primary source is named for either vendor claim; verify both before using them in a pricing negotiation (
raw/OpenAI_Cuts_Off_Elon_s_Cursor_Humanity_s_First_Star_Probe_and_Trump_s_Nuclear_Mars_Ship_EP_285.md).
- Vendors may start billing that way (2026-09-29, host-reported). On Peter H. Diamandis’s Moonshots (EP285), Diamandis says Salesforce prices Agentforce “based on customer revenue generated, not tokens consumed.” He adds that OpenAI “has also done the same, letting some of their largest customers pay only when its AI actually completes the job,” with no charge for tokens, compute time or API calls. His framing: a company selling tokens is “selling you compute”; one selling results “is selling you labor.” No primary source is named for either vendor claim; verify both before using them in a pricing negotiation (
- Audit the middle band. The top of the ladder is obvious and the bottom is free; search and research are where cost “multiplies quietly.”
Update 2026-11-01: where outcome pricing fits, and cost per task in the S-1
From Nate B Jones, “Enterprise AI Pricing — Why Everything You Know Is Wrong” (raw/Enterprise_AI_Pricing_-_Why_Everything_You_Know_Is_Wrong.md, https://www.youtube.com/watch?v=RvnQZQAHUDc), and his reading of Anthropic’s S-1 (raw/Anthropic_Just_Filed_Its_S-1._Here_s_What_The_Headlines_Are_Missing..md).
- Tokens meter compute, not useful thinking. Jones argues that per-token and per-seat contracts leave it to the buyer to turn intelligence into return. He says per-token pricing “is here to stay”: it suits buyers who already know how to apply AI, because it leaves room to experiment and to reassign work.
- Outcome pricing works for closed loops. His named example is Sierra, which he says is actively making per-outcome deals. He pairs it with Cedar, a healthcare-payments company. He does not say outright that Cedar is a Sierra customer. ^[inferred] Cedar’s voice agent, Cora, takes a 2 a.m. billing call, verifies the patient, explains the charges, confirms insurance and takes payment on the call. “The hospital doesn’t care how many tokens that call took.” He also cites Fin in customer service. He says outcome pricing “will not work as cleanly when the work is more complex”, such as redesigning a customer’s whole maintenance programme. He adds: “I don’t know Cedar’s exact contract terms.”
- Between the two sits partnership pricing. Jones points to OpenAI’s ownership stake in Thrive Holdings, with teams embedded in its accounting and IT-services firms, and to Anthropic’s joint venture with Blackstone, Hellman & Friedman and Goldman Sachs. He describes both as cases where the supplier shares the work of turning tokens into value. See Forward-Deployed Engineer Playbook.
- The S-1’s evidence for measuring per task. Jones reports that, according to the filing, Opus 5.5 cut customer cost per task by about 40% against Opus 5. He also cites Ramp’s token-management sample, where token volume rose about 52% over four weeks in September–October while spend rose about 23%. (The caption is garbled; the pairing is approximate.) Per OpenRouter, nearly 90% of September’s agent tokens were cached. Cheaper tasks did not shrink the bill, which is the pattern this primer predicts: measure value per task, not tokens. See Anthropic’s Public S-1.
- A test of ROI that Jones proposes. Put a team freed up by AI on a revenue project you could not previously staff, and attribute that revenue to AI. He offers two public reference points, both as he reports them: Shopify has held headcount flat and deferred hiring for about two years while revenue grew 34% (COO Jess Hertz), and Block’s gross profit grew 25% after it cut staff.
Try It
- Paste a representative sample of your real work into a public tokenizer and look at the chunking — especially if you work in a non-Latin script or in code.
- Find your organization’s equivalent of the 4-million-token yes/no question. Tool misrouting, not volume, is usually the bill.
- Put a cheap default on the fast path and reserve deep-research/agentic modes for tasks that earn them.
- Treat long-running sessions as a line item; restart or compact rather than accumulating.
- Stop rationing the half-cent workloads — it buys nothing and suppresses adoption of the valuable use cases.
Open Questions
- The 5–30× agentic multiplier is attributed to “widely cited industry estimates” with no source named. Directionally consistent with what this wiki has measured elsewhere, but treat the range as folklore until traced.
- The ~half-cent-per-email figure is model- and price-dependent and no model is specified. It is an order of magnitude, not a rate.
- No measured before/after for any of the strategies discussed — the episode is a primer, not a study.
- The 4-million-token anecdote is unaudited (one learner, self-reported).
- Outcome-based pricing is unconfirmed. The Salesforce Agentforce and OpenAI pay-on-completion claims come from a podcast host with no source named. Which customers qualify, and how “completes the job” is judged, are not stated. 2026-11-01: Nate B Jones names Sierra as doing per-outcome deals, with Cedar’s billing calls as the example of a closed-loop outcome, but he says he does not know Cedar’s contract terms. There is still no published outcome-priced contract to cite.