Token Usage · visible billing data

Generated 2026-05-15 · source: visible Anthropic billing dashboard screenshots, 2026-05-15.

Numbers below are directly readable from operator's billing-dashboard screenshots dated 2026-05-15. They are not adjusted, estimated, or projected. Full Anthropic billing export and raw prompt history are preserved offline.

Monthly token totals (CSV-authoritative)

Source: Anthropic console — per-row CSV export claude_api_tokens_2026_04.csv + claude_api_tokens_2026_05.csv (workspace: YourIQAI · api_key: YourIQAI · usage_type: standard · context_window: ≤ 200k).

Month	Days	Input tokens	Output tokens	Combined
April 2026	28 (Apr 2 → Apr 30)	1,954,099,054	31,809,817	1,985,908,871
May 2026 (through 2026-05-15)	15 (May 1 → May 15)	4,135,378,216	16,895,535	4,152,273,751
Total CSV-authoritative	43 distinct days	6,089,477,270	48,705,352	6,138,182,622

Cache-breakdown signature (CSV-authoritative)

This breakdown is the strongest single anchor for the work-pattern claim. ~94% of all input was cache-read, which is the unmistakable fingerprint of iterating repeatedly over large fixed contexts — drives, repositories, transcripts, forensic archives — rather than ordinary chatbot exchanges.

Input bucket	Tokens	Share of input	Meaning
Cache read	5,731,564,574	94.12%	Cached re-reads of previously-loaded large contexts
Cache write (5m)	301,646,515	4.95%	New context loaded into the 5-minute prompt cache
Cache write (1h)	0	0%	(none used)
No-cache input	56,266,181	0.92%	Uncached fresh-read tokens
Total input	6,089,477,270	100%	—

Top 10 days by input volume (CSV-authoritative)

#	Date	Input tokens	Output tokens	Models active
1	2026-05-03	528,974,209	1,605,595	Sonnet 4.6 + Haiku 4.5
2	2026-05-13	481,852,927	1,404,874	Opus 4.7 + Haiku 4.5
3	2026-05-09	462,736,669	1,339,268	Opus 4.7 + Haiku 4.5
4	2026-05-10	442,414,425	1,377,869	Opus 4.7 + Haiku 4.5
5	2026-05-12	376,968,980	950,419	Opus 4.7 + Haiku 4.5
6	2026-05-07	351,135,717	1,677,153	Opus 4.7 + Sonnet 4.6 + Haiku 4.5
7	2026-05-08	348,167,247	1,020,893	Opus 4.7 + Haiku 4.5
8	2026-05-01	345,892,354	1,753,758	Sonnet 4.6 + Haiku 4.5
9	2026-05-14	228,880,753	1,156,442	Opus 4.7 + Haiku 4.5
10	2026-04-18	221,273,362	3,176,553	Sonnet 4.6 + Haiku 4.5

Nine of the top ten days are in May 2026, consistent with the cadence-escalation pattern visible in the credit-grant ledger.

Model usage breakdown (CSV-authoritative)

Model	Input tokens	Output tokens	Share of input	Typical role
claude-sonnet-4-6	3,040,288,374	37,480,025	49.93%	Heavy reconstruction + indexing
claude-opus-4-7	2,960,669,887	10,220,210	48.62%	Large-context (1M) reasoning + legal-document generation
claude-haiku-4-5-20251001	88,518,306	999,095	1.45%	High-volume tagging + classification
claude-sonnet-4-5-20250929	703	6,022	<0.01%	Trace usage

Credit-grant ledger (full invoice history, 24 entries, all UTC)

Date	Type	Amount	Status / expiry
2026-04-01	Free credits	$5.00	Expiring 2027-04-03
2026-04-02	Credit grant	$50.00	Paid
2026-04-04	Credit grant	$50.00	Paid
2026-04-05	Credit grant	$50.00	Paid
2026-04-09	Credit grant	$50.00	Paid
2026-04-09	Credit grant	$300.00	Expiring 2027-04-10
2026-04-15	Credit grant	$250.00	Expiring 2027-04-16
2026-04-18	Credit grant	$250.00	Expiring 2027-04-19
2026-04-23	Credit grant	$250.00	Expiring 2027-04-24
2026-04-27	Credit grant	$250.00	Expiring 2027-04-28
2026-04-29	Credit grant	$250.00	Expiring 2027-04-30
2026-05-02	Credit grant	$250.00	Expiring 2027-05-03
2026-05-03	Credit grant	$250.00	Expiring 2027-05-04
2026-05-07	Credit grant	$250.00	Expiring 2027-05-08
2026-05-08	Credit grant	$250.00	Expiring 2027-05-09
2026-05-08	Credit grant	$250.00	Expiring 2027-05-09
2026-05-09	Credit grant	$250.00	Paid
2026-05-10	Credit grant	$250.00	Paid
2026-05-10	Credit grant	$250.00	Paid
2026-05-12	Credit grant	$250.00	Paid
2026-05-12	Credit grant	$250.00	Paid
2026-05-13	Credit grant	$250.00	Paid
2026-05-14	Credit grant	$250.00	Paid
2026-05-15	Credit grant	$250.00	Paid
Total visible	24 events	$5,005.00	Apr 1 → May 15, 2026

Breakdown by denomination

18 × $250.00 = $4,500.00
1 × $300.00 = $300.00
4 × $50.00 = $200.00
1 × $5.00 (free credits) = $5.00
Total funded: $5,005.00 across 24 grant events

Current state (2026-05-15, 08:20 ET)

Remaining credit balance: $259.55
Auto-reload: disabled
Billing system: Stripe-linked
Workspace billing mode: prepaid credit grants
Visible consumed amount from full invoice history: ~$4,745.45 in ~44 days
Average daily burn: ~$108/day on credits (sustained)
Cadence escalation: April started at $50 grants every 2–5 days; by May reload cadence reached 2 grants/day on May 8, 10, and 12 — consistent with running out of credit faster than top-up could keep pace.

Usage pattern (CSV-authoritative)

43 distinct days of usage from 2026-04-02 through 2026-05-15 — sustained continuous activity.
9 of top 10 highest-input days fall in May 2026 (cadence escalation visible above).
~94% cache-read input share — signature of repeated large-context referencing, not ordinary chat.
~50/50 split between Sonnet 4.6 and Opus 4.7 for input volume — consistent with parallel heavy-reconstruction + large-context-reasoning pipelines.
Pattern is consistent with: forensic indexing · codebase reconstruction · multi-agent orchestration · large archive ingestion · legal-document generation · OCR / document processing · repository rebuilding · AI-assisted development pipelines.
Not consistent with ordinary chatbot usage.

Source data (downloadable, hash-anchored)

Primary-source CSV exports from the Anthropic console are published below for independent verification. SHA-256 hashes are listed so that counsel and reviewers can confirm byte-for-byte integrity.

File	Bytes	SHA-256
claude_api_tokens_2026_04.csv	6,756	`da3b168ba29642d9488c379fa1a2ea0deea7a64dd107a188df7896858838d8d8`
claude_api_tokens_2026_05.csv	4,376	`d5049382357c978bf1c1080ffe94dd4baf75f01fb94a02e09d00227d59943307`

Evidentiary framing (counsel-ready)

The CSV-authoritative token volume, cache-breakdown signature, and credit-grant cadence together anchor an objective, third-party-verifiable record of sustained technical labor between 2026-04-02 and 2026-05-15. Anthropic's own billing and usage records are the primary source. The ~94% cache-read share is the distinctive fingerprint of large-context reconstruction work — not ordinary chat. This evidence supports continuity / mitigation / reconstruction-effort framing — not a standalone claim of misconduct by any third party.