Executive view
The bill, at a glance
This is the most defensible estimate for paying for the observed workload through the APIs—not a subscription-price comparison.
Provider share
OpenAI
37.8%
$7,537.47
Anthropic
62.2%
$12,381.87
Caching changed the economics
$19.9K
Recorded cache behavior
$138.4K
All input priced fresh
Estimated avoided input cost: approximately $118,447 in this 30-day window.
Token profile
Why the volume looks enormous
Long-running agents repeatedly replay histories, repository context, and tool results. Almost all of that input was served from cache.
OpenAI
Codex usage profile
$7,537
89,210 unique usage events after deduplication
Anthropic
Claude usage profile
$12,382
53,080 unique API-like assistant requests after deduplication
Cost concentration
Which models drove the bill
Opus-class and Fable usage dominate Anthropic; GPT-5.6 Sol dominates OpenAI.
OpenAI model mix
$7,537
Anthropic model mix
$12,382
| Model | Input | Cached | Output | API estimate | No-cache control |
|---|---|---|---|---|---|
| GPT-5.6 Sol | 6,695.2M | 6,523.8M | 18.46M | $4,724.64 | $34,471.35 |
| GPT-5.5 | 2,248.9M | 2,129.4M | 8.02M | $1,902.70 | $11,485.00 |
| GPT-5.6 Terra | 2,794.8M | 2,737.7M | 3.89M | $892.52 | $7,102.79 |
| GPT-5.6 Luna | 79.3M | 74.1M | 0.07M | $16.48 | $101.68 |
| GPT-5.4 mini | 5.4M | 4.6M | 0.04M | $1.14 | $4.23 |
| Total | 11,823.7M | 11,469.5M | 30.47M | $7,537.47 | $53,165.06 |
| Model | Requests | Cache reads | Output | API estimate | No-cache control |
|---|---|---|---|---|---|
| Claude Fable 5 | 10,255 | 3,415.6M | 8.48M | $5,075.43 | $35,208.76 |
| Claude Opus 4.8 | 19,788 | 6,027.0M | 21.34M | $4,558.41 | $31,186.44 |
| Claude Opus 5 | 12,821 | 3,191.1M | 12.13M | $2,249.08 | $16,433.80 |
| Claude Sonnet 4.6 | 8,411 | 684.7M | 5.56M | $469.33 | $2,228.26 |
| Claude Haiku 4.5 | 1,175 | 65.7M | 0.45M | $16.14 | $72.49 |
| Claude Sonnet 5 · promo | 401 | 34.0M | 0.14M | $13.48 | $71.96 |
| Synthetic / non-billable | 229 | 0 | 0 | $0.00 | $0.00 |
| Total | 53,080 | 13,418.1M | 48.09M | $12,381.87 | $85,201.71 |
Attribution
Largest identifiable workloads
Names were inferred only from invocation-folder paths. Prompt and response text was not inspected.
OpenAI workloads
01Unattributed sessions$2,372.84
02Content Pipeline Runner$912.76
03Deep Work Strategist (Codex)$846.53
04Vertical Entry Strategist$601.96
05Content Agent$497.42
06Project PM (Codex)$457.46
07Project Manager$406.62
08Agent System Builder$316.06
09SEO Daily Update Owner$289.09
10AgentDeck Developer$274.81
Anthropic workloads
01Deep Work Strategist (Fable)$3,217.62
02Content-to-Page / Bricks Implementer$2,293.09
03AgentDeck Developer (Claude)$2,186.30
04Deep Work Session (Fable)$2,169.67
05Agent System Builder$423.34
06Contra / Review Specialist$395.15
07Keyword Research Owner$287.13
08Visual QA Reviewer$237.73
The top four Anthropic workloads account for $9,866.68—79.7% of Anthropic cost.
Rate card
Pricing used in the estimate
USD per million tokens (MTok), using standard global, non-batch API rates.
OpenAI model rates
| Model | Input | Cached input | Output |
|---|---|---|---|
| GPT-5.6 Sol | $5.00 | $0.50 | $30.00 |
| GPT-5.6 Terra | $2.50 | $0.25 | $15.00 |
| GPT-5.6 Luna | $1.00 | $0.10 | $6.00 |
| GPT-5.5 | $5.00 | $0.50 | $30.00 |
| GPT-5.4 mini | $0.75 | $0.075 | $4.50 |
Anthropic model rates
| Model | Base input | 5m write | 1h write | Cache hit | Output |
|---|---|---|---|---|---|
| Claude Fable 5 | $10.00 | $12.50 | $20.00 | $1.00 | $50.00 |
| Claude Opus 5 / 4.8 | $5.00 | $6.25 | $10.00 | $0.50 | $25.00 |
| Claude Sonnet 4.6 | $3.00 | $3.75 | $6.00 | $0.30 | $15.00 |
| Claude Haiku 4.5 | $1.00 | $1.25 | $2.00 | $0.10 | $5.00 |
| Claude Sonnet 5 · through Aug 31 | $2.00 | $2.50 | $4.00 | $0.20 | $10.00 |
| Claude Sonnet 5 · from Sep 1 | $3.00 | $3.75 | $6.00 | $0.30 | $15.00 |
Action plan
Four practical cost controls
Use high as the normal ceiling
Reserve xhigh for bounded, high-value work. Historical Sol/xhigh usage
alone represents $4,054 at API-equivalent prices.
Protect cacheable prompt prefixes
Stable histories and workspace prefixes are the difference between roughly $20K and the $138K no-cache control.
Gate Fable and Opus-class work
Fable 5 plus Opus 4.8/5 account for approximately $11,883 of the $12,382 Anthropic estimate.
Install a monthly API-style meter
Record model, input, cache writes, cache reads, output, effort, and invocation so future reports become a repeatable budget control.
Audit trail
Method, quality, and exclusions
Only usage metadata was parsed. No prompt, response, email, document, or tool-result content was included.
Method and data quality
- OpenAI: 3,454 candidate rollout files inventoried; 520 recently modified files could contain in-window events.
- OpenAI parser retained 89,210 unique cumulative events, removed 230 repeats, and ignored 22 malformed JSONL lines.
- Anthropic: all 795 Claude project JSONL files were streamed; 53,080 unique requests remained after global message-ID deduplication.
- OpenAI long-context premiums were applied to 450 requests above 272K input tokens.
- Anthropic base input, five-minute writes, one-hour writes, cache reads, and output were priced separately.
- Three Claude messages had TTL subtotals exceeding the top-level cache-creation field by 274,418 tokens (0.12% of cache writes). TTL subtotals were used; the conservative impact is under $3.
Uncertainty and exclusions
- Local usage logs may not match provider invoice aggregation around retries or product-side orchestration.
- Deleted, off-device, or otherwise unavailable session logs are not included.
- US-only Anthropic inference would add 10% to the Anthropic portion.
- Batch discounts were not assumed because these were interactive agent workloads.
- Taxes, marketplace markups, separately billed tools, storage, containers, hosted search, images, and fast-mode premiums are excluded.