Two weeks before this release shipped, we turned per-person context cost on for our own engineers. Within a week, gross context waste was down roughly 300,000 tokens per session, worth about $10,000 a month to a team our size. The team kept working the same way. We wrote up what happened and why we think it worked.
Revenium 2.20.0 makes AI coding cost visible in the places it was still hiding, including the spend that used to look free. Context bloat crept up on jobs that ran every night. Claude Enterprise customers saw only an organization total. Teams running Claude Code through their own Bedrock account saw no cost at all, even though AWS was billing them the whole time. This release closes each of those gaps and ties coding-assistant spend to work delivered.
Here is what shipped.
1) Spot context waste before it shows up on the bill
A new AI Insights finding flags agents and jobs that pull far more context than the work needs, and estimates what that overhead is costing. It does this without reading prompts, so sensitive content stays where it is.
Over-fed context fails quietly in production AI. A retrieval step gets loose, an agent starts stuffing history, and the token bill drifts upward without anything looking broken. No alert fires, because nothing is actually down. A nightly job that has tripled its context costs three times as much on every run.
- Catches one-off outlier runs, and jobs whose context has grown since the last period
- A cost estimate for that overhead, so it can be prioritized against other work
- Works from usage totals alone, so a security reviewer can approve it
2) See what each person spends on Claude Enterprise
Claude Enterprise customers can now see per-person spend instead of only the organization total. Seat subscription charges billed separately by Anthropic are not included.
- Per-person view on the Users tab of the Anthropic Enterprise provider overview, under User Cost Analytics
- A basis for internal cost allocation from what used to be one invoice line
- Where usage concentrates, and which teams are barely using what the company already pays for
3) Enhanced AWS Bedrock support for Claude Code
Many enterprises run Claude Code through their own AWS Bedrock account, so AWS bills them directly and the usage showed no cost in Revenium. It is now counted as real spend, priced at your own rates where you have filed them with us, and included in spend alerts alongside the rest of your AI cost. List price applies where rates are not filed.
- AI coding spend that used to look free now shows at the real number
- Your negotiated Bedrock rates apply when filed, so the number matches your invoice
- Bedrock-routed spend is covered by the same guardrails and alerts as everything else
This is an Enterprise capability. Contact us for more information on turning it on.
4) Claude purchased through Azure Marketplace is now captured
Claude purchased through Azure Marketplace and billed on your Azure bill is now pulled into Revenium automatically, alongside the rest of your AI cost. It appears in the Azure provider’s cost figures on the provider overview.
- Azure Marketplace Claude spend arrives in billing and analytics on its own
- One more source of AI spend that no longer needs a manual export
- Nothing to configure
5) Per-person context cost on the Claude Code dashboard
The Adoption tab of the Claude Code dashboard now includes a Context efficiency section showing each person’s context cost and how it compares to the team. Context is the biggest lever on what a coding assistant costs, and until now it wasn't visible per person.
- Per-person context cost, ranked against the rest of the team
- Early warning on context bloat before it grows into the bill
- Fills in once a team has enough usage to compare
This is the feature we turned on for ourselves. See what it did.
6) Cost per pull request, by vendor
You can now see what a merged pull request actually cost in AI. The Comparison tab on the provider overview carries the Cost/PR column next to vendor rate benchmarks and billed token rates, and each coding assistant’s dashboard shows cost per pull request in its user table.
Engineering leaders have been asking for a per-PR number since coding assistants became a real line item. Totals and per-user numbers only get partway there. Cost per PR ties AI spend to work delivered.
- Cost per merged pull request, per assistant, with vendor rate benchmarks alongside it
- Your invoiced rates shown next to advertised vendor pricing
- A basis for comparing assistants on what they produce rather than on what they cost
This one needs your source control connected before the numbers appear.
7) Traces now sit alongside your logs
A trace is the step-by-step record of one AI request or agent run, including its cost. Traces now get a dedicated tab on the Logs page, so getting from a log line to the trace behind it no longer takes a page change. The cost, performance, and efficiency trace views export to CSV.
- Traces in the same place as logs
- CSV export on the cost, performance, and efficiency trace views
- Fewer places to look when investigating a slow, expensive, or unexpected request
Also in this release
Unattributed spend is now visible. Spend a provider cannot tie to a person is shown rather than hidden, so a per-person view tells you how much of the total it actually accounts for.
Insights you can act on. AI Insights recommendations now weigh every connected source when forming a finding, and say clearly why a specific finding cannot be promoted into an alert.
Skills and Tools filtering. The Skills and Tools view can now be filtered by person, and those filters persist in the page URL. Filtered views are shareable and survive a refresh.
For the full technical details on everything that shipped, the changelog has the complete rundown.
Try it out
Revenium 2.20.0 is live in production.
If you already use Revenium, start with per-person Claude Enterprise spend and the new context-waste Insights finding. Those two don’t need new plumbing. If your Claude Code usage runs through Bedrock, or if you buy Claude through Azure Marketplace, the new coverage in this release brings that spend into the same view as the rest of your AI cost.
Talk to us if you want to see what your team’s context habits are costing.
