joey.hersh

home / methodology

How the token number is built.

The quarter-trillion figure on the front page, broken down by model family, harness, and year.

modeled estimate · input plus output tokens · not a prompt count

A quarter trillion tokens, and where they went

Input plus output tokens across every model I've used since GPT-3, reconstructed from subscription ceilings, API bills, and one measured agent ledger. Cached context reads count, because agentic coding re-reads enormous context and that is where most of the volume lives. It is a modeled estimate, not a meter. The shape is the point: a lot of tokens through a lot of different models, harnesses, and providers.

Horizontal bar chart of modeled tokens processed by model family, in billions, plus one measured agent-environment ledger. OpenAI 190B Anthropic 38B Google 9.0B Open-weight and self-hosted 8.0B One agent environment, measured 18B
0¼ trillion300B

About 155B to 263B. The bars sum to about 263B. The headline rounds that to "more than a quarter trillion" and treats it as a modeled ceiling.

Fifteen harnesses and providers

  • ChatGPT
  • Codex CLI
  • Claude Code
  • claude.ai
  • Cursor
  • OpenCode
  • Gemini CLI
  • Hermes
  • OpenClaw
  • n8n
  • Cyrus
  • AWS Bedrock
  • OpenRouter
  • ComfyUI
  • Unsloth

Seven years of models

  1. 2019GPT-2 fine-tunes on Colab
  2. 2022GPT-3 fine-tuned classifier, OPT-1.3B document QA, GPT-4 API early access
  3. 2023GPT-3.5 Turbo in production, Falcon-40B RAG inside a bank, Cyrus on the first LangChain release
  4. 2024Claude 3.5, Llama 3, Bedrock agents, LangGraph, AutoGen at launch
  5. 2025Codex, Claude Code, Cursor, Gemini CLI, Hermes and OpenClaw, thousands of scheduled agent runs
  6. 2026GPT-5 family, Claude 5 family, Qwen 3 fine-tunes as a custom model router

How the number is built

  • Shared reconstruction of the OpenAI subscription usage. Models the OpenAI family alone at 225 to 250 billion at the ceiling. The split across the other families is my reconstruction; no provider export was reconciled.
  • Hermes scheduler ledger, August 19 to September 17, 2026. 920 recorded runs and 1.34 billion prompt-plus-completion tokens in thirty days, one profile, scheduled jobs only.
  • Independent measurement of Pro-plan API-equivalent value. Third-party. Used as the ceiling anchor for the largest family.
  • Read-only inspection of one Hermes environment's usage ledger. Measured, but limited to retained rows in one of twelve environments.
FamilyModelsHarnessesBasisBillions
OpenAIGPT-3 fine-tunes, GPT-3.5 Turbo, GPT-4 early API access, GPT-4o, o-series, GPT-5 through 5.6ChatGPT, Codex CLI, the API, Cursor, Cyrussubscription ceilings plus API spend, 2022 to 2026190B
AnthropicClaude 3.5 through Claude 5, including Fable and OpusClaude Code, claude.ai, Cursor, Hermes, Bedrocksubscription ceilings plus API spend, 2024 to 202638B
GoogleGemini 1.5 through Gemini 3, Vertex AIGemini CLI, AI Studio, an internal spec-driven CLI forkAPI spend and CLI usage, 2024 to 20269.0B
Open-weight and self-hostedGPT-2 fine-tunes, OPT-1.3B, Falcon-40B, Llama 2 and 3, Mistral, Qwen 3 fine-tuned with Unslothlocal GPUs, ChromaDB and pgvector RAG stacks, OpenRouter, ComfyUIlocal runs and OpenRouter spend, 2019 to 20268.0B
One agent environment, measuredmixed, routed by taskHermesread-only ledger of input plus cache-read tokens in one environment, September 202618B
Modeled total263B