FOCUS export

The file focus-2026-10-01-to-2026-10-08.csv holds the platform's metered costs in the Cost and Usage columns of FOCUS 1.4. It has 35 rows: 24 for Claude Code (one per UTC day and model) and 11 for GitHub Actions (one per UTC day and public repository). scripts/focus_export.py wrote it on 2026-10-09.

Check sums: the ListCost of the Claude Code rows adds up to USD 1,021.01, the Loki total for the range; the Actions rows add up to 47.0 runner minutes over 63 runs.

Run

Loki must be reachable; on the workstation that means an SSH tunnel to the home server. The GitHub token only raises the API rate limit.

GITHUB_TOKEN="$(gh auth token)" LOKI_URL=http://localhost:13100 \
  python3 scripts/focus_export.py --start 2026-10-01 --end 2026-10-09 \
  --out inside/finops/focus-2026-10-01-to-2026-10-08.csv

The script runs one instant LogQL query per day, metric and model (the request_sum() pattern from grafana/build.py with a [1d] range) and reads job durations from the Actions API.

Column mapping

Column	Claude Code rows	GitHub Actions rows
`BilledCost`	0: the subscription covers the calls	0: standard runners are free for public repositories
`ListCost`	Sum of `cost_usd`: token counts at Claude API list prices, as Claude Code computes them	0: the list price for public repositories
`ContractedCost`	Equal to `ListCost`: no negotiated rates	0
`EffectiveCost`	Empty (deviation 1)	0
`ChargeCategory`	`Usage`	`Usage`
`PricingQuantity`, `PricingUnit`	All tokens of the day and model, `Tokens`	Job minutes, `Minutes`
`ConsumedQuantity`, `ConsumedUnit`	Same as pricing	Same as pricing
`ServiceProviderName`, `HostProviderName`, `InvoiceIssuerName`	Anthropic	GitHub
`ServiceCategory`	`AI and Machine Learning`	`Developer Tools`
`ServiceName`	Claude Code	GitHub Actions
`SkuId`, `SkuPriceId`	The model ID	`actions_linux`, `actions_linux:public-repository`
`ResourceName`	Empty	The repository
`ChargePeriodStart`, `ChargePeriodEnd`	The UTC day	The UTC day
`BillingPeriodStart`, `BillingPeriodEnd`	The calendar month	The calendar month
`x_Requests`	API calls	Workflow runs
`x_InputTokens`, `x_OutputTokens`, `x_CacheReadTokens`, `x_CacheWriteTokens`	Token counts by class	Empty
`x_PrivateRepoEquivalentCost`	Empty	Minutes at USD 0.006, the Linux 2-core rate for private repositories

Columns that start with x_ are custom columns, which FOCUS allows with that prefix.

Declared deviations from FOCUS 1.4

EffectiveCost is empty on the Claude Code rows. The specification requires the share of the covering purchase (the subscription fee) applied to each charge, and the fee is not published.
No Purchase rows. The subscription, the domains and the reverse proxy are missing. The Bedrock eval runs are missing too: BilledCost must be the invoiced amount, and only the harness's estimates are published (Unit economics lists them per run).
BillingAccountId and BillingAccountName read withheld.
ListCost on the Claude Code rows uses the per-token prices of the Claude API, the prices Claude Code applies. The plan in use has no per-token price.
Each Claude Code row sums four token classes with different prices into one PricingQuantity, so the file gives no ListUnitPrice. The FOCUS 1.5 draft adds TokenDirection and TokenCacheAction properties for this split; until then the counts by class are in the x_ token columns.
The file has not been run through the FOCUS Validator.

Why export at all

FOCUS gives one schema for every provider. The same file can join the cloud provider's own FOCUS export, once the invoice lines are added, without a mapping per vendor. It also makes the deviations explicit: each empty cell above marks a figure that the owner keeps private or that no meter provides. The FinOps framing of the platform: FinOps.