Token

Where the record uses it

• Dated test Chapter 12, beside the five, the coder hypothesis, due 31 December 2028
• Person Baker, Gavin

Also discussed in

• Page The Compute Economy

Sources that use it

• Anthropic Economic Index, “Cadences,” 26 June 2026, read from the report PDF and its web page (the fourth report in the series, on continuous daily sampling with a classifier that labels what each conversation produced, “Our classifier identified 93% of Claude conversations as producing an artifact”, so this edition reports no coding share and the earlier reports’ series cannot be extended from it — published by the company that sells the tokens, about its own traffic; on chat and Cowork conversations classified as work-related, “we find that compute tends to scale with the value of the work”, Table 2.4’s caption giving “a typical conversation mapped to a top-tercile occupation consumes 2.07 times as many tokens as a typical conversation mapped to a bottom-tercile occupation”, and the report’s own hedges beside it, “Admittedly, the relationship is noisy, and there are notable outliers. Pharmacists, for example, earn nearly three times what statistical assistants do ($68 vs. $24 per hour), yet conversations mapped to pharmacist tasks use only about one twentieth as many tokens” and “About 44% of the wage gradient in token consumption is explained by output mix”; Figure 2.3’s caption records that the data “is restricted to chat and Cowork and restricted to conversations classified as work-related. Token counts are not adjusted for which model served the conversation”; the Appendix, p. 2, on the pipeline’s changes: the O*NET task classifier replaced by one that “picks a detailed work activity” in its first step, on O*NET version 30.2, the output classifier introduced beside it, not in its place; book 1’s cross-examination’s two strings in the report’s web page, “Introduce a new classifier that labels the output of each conversation.” and “Chat transcripts no longer fully capture how people are using AI, and our methods”, the page’s capital C; the page held byte-exact on 18 September 2026, anthropic-cadences-20260626-page.txt, md5 03eb6c3e, and the strings re-read in it on 24 September 2026). In the Sources
• Anthropic, “Pricing,” Claude Platform Docs, retrieved 4 September 2026, with the “Steering thinking” and “Fast mode” pages (base input, cache writes, cache hits and output priced apart; the Steering thinking page’s own Pricing section, “Thinking incurs charges for: Tokens Claude uses while thinking (billed as output tokens)” and “You are billed for the full thinking process, not the thinking content visible in the response”; the model table’s five columns, headed “Base input tokens”, “5m cache writes”, “1h cache writes”, “Cache hits and refreshes” and “Output tokens”; the two retired Opus rows, Claude Opus 4.1 and Claude Opus 4, marked retired except on Bedrock and Google Cloud, at “$15 / MTok” base input and “$75 / MTok” output; fast mode for Opus 5 at “$10 / MTok” in and “$50 / MTok” out, the section reading in full, “Fast mode, in research preview, provides significantly faster output for Claude Opus 5 and Claude Opus 4.8 at premium pricing,” with one table row, “Claude Opus 5 / Claude Opus 4.8 $10 / MTok $50 / MTok”; “This tokenizer produces approximately 30% more tokens for the same text”; the page continues, “The exact increase depends on the content and workload shape”; cache hits on Claude Fable 5.1 and Claude Mythos 5.1 “are priced at 0.025x the base input price. All other models use the standard 0.1x multiplier” — $0.25 per million against Opus 5’s $0.50, on a base input of $10 against $5; the section “Long context pricing”: “Claude 4.6 and later models and Claude Mythos Preview include the full 1M token context window at standard pricing. (A 900k-token request is billed at the same per-token rate as a 9k-token request.) Prompt caching and batch processing discounts apply at standard rates across the full context window”; the scope of the tokenizer sentence, stated beside it: “Claude 4.7 and later models and Claude Mythos Preview use a newer tokenizer”, and its closing sentence, “Claude Sonnet 4.6 and earlier models use the previous tokenizer”; the model pricing table as of 20 September 2026: Claude Fable 5.1, Claude Fable 5, Claude Mythos 5.1 and Claude Mythos 5 each at “$10 / MTok” base input and “$50 / MTok” output, and Claude Opus 5 at “$5 / MTok” and “$25 / MTok”; read on 4 September 2026 and re-read on 13, 19, 20 and 21 September 2026, the model table, the long-context section, the tokenizer note and the cache-hit multiplier unchanged on the last reading and the Steering thinking page by then served at platform.claude.com/docs/en/build-with-claude/thinking-steering-and-cost, the live page archived as VERIFY 260913/ch02 fold, VERIFY 260919/T1, VERIFY 260920/FOLD 255 and, with the Steering thinking page, VERIFY 260921/R68 B; re-read 23 September 2026: the model table adds Claude Opus 5.5 at “$4 / MTok” base input and “$20 / MTok” output, its footnote reading “Cache hits and refreshes on Claude Opus 5.5 are priced at 0.05x the base input price”; Claude Opus 5 stays at $5 and $25, Claude Fable 5.1 and Claude Mythos 5.1 at $10 and $50; the fast mode table now reads “Claude Opus 5.5 $8 / MTok $40 / MTok” above the Opus 5 row; archived as VERIFY 260923/FOLD 277; re-read 4 October 2026 for the section “Data residency pricing”: “For Claude 4.6 and later models, specifying US-only inference through the inference_geo parameter incurs a 1.1x multiplier on all token pricing categories, including input tokens, output tokens, cache writes, and cache reads. Global routing (the default) uses standard pricing”; archived as VERIFY 261004/FOLD 552; the model table’s rows: Claude Fable 5.1 and Claude Mythos 5.1 at “$10 / MTok” base input and “$50 / MTok” output, Claude Opus 5 and the Opus 4.5 to 4.8 classes at “$5 / MTok” and “$25 / MTok”). In the Sources
• Harsh Singh Banwait, “NVIDIA Vera Rubin NVL72 on CoreWeave: 10x More Tokens Per Megawatt Than Blackwell,” CoreWeave blog, 21 July 2026 (CoreWeave’s own measurement, on DeepSeek R1: “At a matched interactivity target, Vera Rubin NVL72 generated 10x tokens-per-second per megawatt compared to NVIDIA GB200 NVL72 on the same DeepSeek R1 workload”; the conditions, “with all major, state-of-the-art inference optimizations turned on,” among them “large-scale expert parallelism, NVFP4 precision, multi-token prediction, and disaggregated prefill and decode”; the page’s other wording of the result, “up to 10x higher tokens per megawatt”; what it means for customers, in the company’s words: “run far more reasoning-model traffic within the same power budget or run the same traffic on far less power”; CoreWeave sells the capacity it measured, and NVIDIA’s blog of the same day relays the figure as its own platform’s; the comparison on both systems, “This evaluation compares the performance on each platform with all major, state-of-the-art inference optimizations turned on”; read in the page as served in October 2026, which carries its publication date). In the Sources
• Martin Casado (Andreessen Horowitz), “OpenRouter & Stripe: The Intelligence Network,” a16z.com, 19 August 2026, posted the day Stripe announced its agreement to acquire OpenRouter and published in the same minute as an X Article (“‘Tokens are the new dollars’ has become true in both economic theory and in literal internet plumbing”; “Tokens have become a new, universal medium of value exchange”; the routing “becomes the unsung enabler of the whole story”; the post quotes Will Gaybrick of Stripe: “We just want to make sure that moving between tokens and dollars is as seamless and safe as moving between dollars and euros”; on the firm’s own investing in OpenRouter, “We jumped at the chance to do the seed deal, and when it came time for the A, we took a big swing”). In the Sources
• Microsoft, second-quarter fiscal 2026 earnings call, 28 January 2026, transcript published by the company (Satya Nadella: “The key metric we are optimizing for is tokens per watt per dollar”; the three layers of the stack named as “Cloud & Token Factory, Agent Platform, and High Value Agentic Experiences”; the word order moves between calls, the first-quarter call of 29 October 2025 giving “maximizing tokens per dollar per watt,” so each quotation names its own call; “And over 250 customers are on track to process over one trillion tokens on Foundry this year”; Nadella on Copilot: “now have 15 million paid Microsoft 365 Copilot seats”; Amy Hood, chief financial officer, in prepared remarks: “paid M365 commercial seats grew 6% year-over-year to over 450 million”). In the Sources

And 11 more: search the Sources.