AI Activity Ledger

11.4B recorded tokens across 30 active days

Total tokens11.4B11,428,166,067 recorded
Current streak14through last completed day
Daily average380.9Mtokens on active days
Models observed62 agent tools
Cache warmth96.1%cache-served input signal

Daily Usage By Model

Newest day first. Bars extend left-to-right by token count; 1B tokens = 932 px. Stacked by model and tool. Effort stays in the detail table because it is provider-specific and not always captured.

Codex / gpt-5.5 10.7B Claude Code / claude-fable-5 519.8M Claude Code / claude-opus-4-8 156.4M Claude Code / claude-sonnet-4-6 38.4M Codex / Model not captured 13.3M Codex / gpt-5.4 4.1M

Cadence

The strongest proof signal is not one large number. It is repeated usage over time.

Peak Day

Largest completed-day total in this public window.

2026-05-212.7B2,742,413,720 tokens

Model And Tool Mix

Breadth across tools and models is shown as aggregate share, not as private project attribution.

Models

gpt-5.5
10.7B 93.6%
claude-fable-5
519.8M 4.5%
claude-opus-4-8
156.4M 1.4%
claude-sonnet-4-6
38.4M 0.3%
Model not captured
13.3M 0.1%
gpt-5.4
4.1M 0.0%

Tools

Codex
10.7B 93.7%
Claude Code
714.6M 6.3%

Token Anatomy

The split shows what the providers expose: prompt input, cache behavior, output, and reasoning where available.

Prompt inputContext sent to the model.
10.7B
Cache-warm inputInput already served from provider cache where exposed.
10.9B
Cache writeNew cacheable context recorded by providers.
31.1M
OutputAssistant response tokens.
33.2M
ReasoningReasoning tokens when providers expose them.
8.7M

Data Quality

This page should be useful as a receipt, so incomplete fields are named directly and kept out of headline charts.

OpenAI effort coverage 72.9%

Reasoning effort is shown only when the source log captured it. Claude Code usage is marked not applicable instead of forcing a false effort category.

Cache-warm input 96.1%

High cache warmth indicates sustained context-heavy agent sessions, not isolated prompt demos.

Source provenance linked-active / 73.8%

Usage is grouped by whether source logs are linked to active assistant-session records, historical local logs, fixtures, or mixed provenance.

Usage detail table
Provider Tool Model Effort Total Input Cache-warm Output Confidence
AnthropicClaude Codeclaude-fable-5N/A519,768,9891,679,203495,649,9003,651,884local-log-derived
AnthropicClaude Codeclaude-opus-4-8N/A156,411,728124,663148,027,117561,438local-log-derived
AnthropicClaude Codeclaude-sonnet-4-6N/A38,417,95190833,209,951584,179local-log-derived
OpenAICodexModel not capturedNot captured13,283,00613,231,13311,362,04851,873local-log-derived
OpenAICodexgpt-5.4Not captured4,055,0144,039,1533,804,16015,861local-log-derived
OpenAICodexgpt-5.5Not captured2,883,028,0862,870,301,1542,714,890,1129,290,623local-log-derived
OpenAICodexgpt-5.5low148,249147,01761,0561,232local-log-derived
OpenAICodexgpt-5.5medium805,849,990802,670,354756,545,0242,290,155local-log-derived
OpenAICodexgpt-5.5xhigh7,007,203,0546,984,196,3466,779,883,39216,743,311local-log-derived

Method

Source table: usage_rollup_daily

Privacy boundary: public-safe-aggregated-no-prompts-no-paths-no-private-projects

Costs are omitted unless explicitly estimated or reconciled upstream.

This page proves recorded model activity. Shipped work and verification are separate evidence streams.

A token is a model accounting unit, not a shipped-feature metric; shipped output needs separate receipts.

Window: 2026-05-14 to 2026-06-13. Generated at: 2026-06-14T17:28:58Z. Machine-readable data: usage.json.

Neutral Observer Cache Read

A public interpretation of cache behavior should be useful without overstating what token logs can prove.

What It Signals

From a neutral observer's perspective, 96.1% cache warmth is evidence of repeated work inside sustained, context-heavy agent sessions. The ledger shows 10.9B cache-warm input signal against 31.1M cache writes, which suggests the same working context is being reused rather than repeatedly rebuilt from scratch.

What It Does Not Prove

Cache usage is not a shipped-work metric and not a quality score. It does not say whether the work was correct, useful, or efficient end-to-end. It says the interaction pattern is persistent and context-rich; shipped artifacts, commits, reviews, and external outcomes remain separate evidence streams. Provider bucket semantics also differ, so cache-warm input should be read as a directional signal, not a precise cross-provider efficiency benchmark.

Observer Conclusion

The constructive read is that this is a mature usage pattern: large context is being carried forward and amortized across many turns. The risk to watch is context sprawl. A high cache ratio is strongest when paired with pruning, compact task boundaries, and visible output. In this window, output and reasoning tokens total 41.9M, so the page should frame cache warmth as operating continuity, not as a standalone claim of productivity.