See what Bedrock and Vertex AI cost, model by model

Two kinds of AI spend appear side by side, and they are measured differently.

• Amazon Bedrock: daily unblended cost from Cost Explorer in your AWS account, grouped by usage type and region. The model and whether the tokens were input or output are read from the usage type.
• Google Vertex AI: daily cost from your BigQuery billing export for services whose name starts with “Vertex AI”, grouped by SKU, region and project.
• GetFinOps’ own AI: each Claude call made for your workspace, such as report narratives, copilot answers and the daily brief, recorded as it runs and priced from a per-model rate table.
• A usage type or SKU that doesn’t match a known model is kept as “unknown” rather than dropped.

The AI Spend page shows month-to-date spend against last month, total tokens, and the change from the previous seven days. A chart plots cost by day, week or month over windows from 24 hours to a year, and a table lists tokens and cost for each model.

Tabs filter the views to one provider: GetFinOps’ own calls, Bedrock or Vertex AI. The same breakdowns are available to AI assistants as read-only MCP tools.

Two rules run on the collected Bedrock and Vertex AI spend in every scan, and their findings are listed on the AI Spend page as well as with your other findings.

• Cost spike: a model’s latest daily spend in a region is above its prior seven-day average plus two standard deviations, or above 1.25 times the average when those earlier days are all equal. The flagged day must cost at least $1, and at least three earlier days are needed. Three standard deviations or more is marked high severity.
• Idle provisioned capacity: over seven days, Bedrock Provisioned Throughput or a Vertex AI deployed endpoint costs at least $1 while serving fewer than 1,000 tokens. The finding shows the capacity cost and the tokens actually served.

Each collection records, per provider, the days it covered, the dates missing from the window, and whether it succeeded, came back empty or failed. Bedrock and Vertex AI each get a freshness badge, and a banner lists missing dates.

An admin can backfill up to 90 days for a provider from that banner, once every 15 minutes.

Every Claude call GetFinOps makes for your workspace is stored with its model, input and output tokens, cache writes and reads, cost, duration, stop reason and, where Anthropic returns one, its request ID. Cache tokens are priced separately because Anthropic bills them separately from input.

Click a model to open its individual calls, with totals for the whole filter rather than just the page. Calls made inside an agent run are grouped by agent, showing runs, calls per run and cost per run, plus the share of spend that could be tied to an agent at all.

• It doesn’t collect spend your own applications send directly to Anthropic, OpenAI, Azure OpenAI or other model APIs.
• It doesn’t show individual requests, cache tokens or agents for Bedrock and Vertex AI, which arrive as daily totals.
• Vertex AI figures are the billing export’s cost column, before credits.
• AI spend views cover the whole workspace; they don’t follow the cloud account selector.
• It doesn’t change provisioned capacity, endpoints or model settings.