Claude Code 2.1.239: reconcile the US-only inference premium before trusting the budget cap

Author: Ahab · Date: August 22, 2026

Quick answer

Claude Code 2.1.239, released on August 21, 2026, changes three local cost surfaces—/cost, the status line, and --max-budget-usd—so they include Anthropic's 1.1× US-only-inference premium for data-residency workspaces. This closes an estimate gap; it does not change the underlying premium, prove where a request ran, or make a local budget estimate equal to the final invoice.

Upgrade the canary lane, run one bounded task, and reconcile four records: effective inference geography, Claude Code's displayed estimate, --max-budget-usd behavior, and provider usage or billing. Keep a margin until all four agree. Do not turn off a residency control merely to save 10%; residency is a policy decision, not a model-routing optimization.

Who this is for

This guide is for teams using Claude Code on a usage-based Enterprise organization or a Claude API workspace that requires US-only inference. It also helps platform owners who enforce a dollar ceiling on unattended coding tasks.

It is not a general Claude pricing comparison. If the immediate problem is containing agent spend, start with the Claude Code spend-limit checklist. If the risk is a runaway local build rather than inference cost, use the Linux Bash memory-limit canary.

What changed—and what did not

Surface	Confirmed behavior	Boundary
Claude Code 2.1.239	`/cost`, the status line, and `--max-budget-usd` include the 1.1× premium for data-residency workspaces	The release note does not say the premium began in 2.1.239; it says estimates now include it
Claude organization apps	US-only inference applies across Claude apps, including Claude Code, on eligible usage-based Enterprise organizations	Seat fees do not receive the 1.1× multiplier; usage does
Claude API	`inference_geo: "us"` is priced at 1.1× for Claude 4.6 and later across input, output, cache writes, and cache reads	`"global"` uses standard pricing; older unsupported models reject the parameter
Priority Tier	US-only tokens draw down committed capacity at 1.1 tokens per token	Priority capacity and ordinary rate limits still both apply
Partner platforms	Anthropic documents separate regional pricing behavior for Bedrock and Google Cloud	Do not copy first-party Claude API math onto a partner invoice

The important correction is semantic: a display can now estimate the known premium, but only the response usage field and the provider's usage record establish what was served and billed.

The four-surface reconciliation

1. Freeze the policy before measuring cost

Record the organization, workspace, model, provider, and required geography. For first-party Claude API calls, capture the workspace's allowed_inference_geos and default_inference_geo; for an organization-wide Claude app setting, record whether US-only inference is enabled.

Do not change these controls during the canary. A mixed global/US sample makes a 10% difference look like noise. Use a synthetic repository and a task with no customer data so the measurement itself does not expand the compliance scope.

2. Calculate the expected premium by token category

The safe estimate multiplies every applicable token-price category, not only uncached input and output:

standard_cost =
  input_tokens       × input_rate
+ output_tokens      × output_rate
+ cache_write_tokens × cache_write_rate
+ cache_read_tokens  × cache_read_rate

expected_us_cost = standard_cost × 1.10

Use the model's current official rates and the provider that actually serves the request. Batch discounts, Priority Tier commitments, negotiated pricing, taxes, and partner-provider regional rates are separate terms; a universal × 1.10 invoice calculator would be wrong.

3. Run one bounded Claude Code canary

Pin one launcher to 2.1.239 or newer and record claude --version. Use a small deterministic task, such as reading two synthetic files, changing one line, and running one local test. Run it once without a cap to observe the estimate, then repeat an equivalent fixture with --max-budget-usd set above the expected US-only cost but below a deliberately wasteful retry loop.

Capture only non-sensitive evidence:

the /cost result;
the final status-line estimate;
the configured --max-budget-usd value and terminal outcome;
model and provider identity;
input, output, cache-write, and cache-read usage;
the response usage.inference_geo when the API surface exposes it.

Pass only if the task completes below the cap, both Claude Code displays agree within a documented rounding tolerance, and the effective geography matches policy. The cap stopping an oversized negative control is useful, but do not create a costly loop just to test it; lower the cap against a disposable task instead.

4. Reconcile with the authoritative usage record

Compare the local estimate with Anthropic Console usage, the Usage and Cost API, or the actual partner-provider bill. Normalize the same time window, workspace, model, and token categories before comparing.

Use this order of authority:

policy configuration
  -> response usage and inference_geo
     -> provider usage/cost record
        -> Claude Code local estimate

If the local estimate differs, do not immediately raise the cap. First check model aliases, cache categories, retries, parallel subagents, currency or tax, partner routing, and whether the task crossed the reporting window.

Budget decision tree

Does policy require US-only inference?
  no  -> use the approved global policy and standard-rate baseline
  yes -> is the serving route first-party Claude API / eligible Claude Enterprise?
           no  -> use the partner's regional price and usage record
           yes -> does version report 2.1.239 or newer?
                    no  -> do not trust local premium-aware estimates
                    yes -> do usage.inference_geo and workspace policy agree?
                             no  -> stop; fix the policy or routing mismatch
                             yes -> do /cost, status line, cap, and provider usage reconcile?
                                      no  -> keep the lane canary-only
                                      yes -> add measured retry headroom and promote gradually

Common mistakes

Treating 2.1.239 as a price increase rather than an estimate correction.
Applying 1.1× only to output while ignoring cache reads and writes.
Assuming a US workspace location automatically proves every inference request ran in the US.
Using a Claude API formula for Bedrock or Vertex without checking that provider's regional price.
Setting --max-budget-usd exactly at the median successful task cost, leaving no room for legitimate variance.
Comparing a local one-task estimate with an organization invoice that includes other users, retries, or taxes.
Disabling a residency requirement because the global route is cheaper.

Copyable acceptance record

date / owner / launcher / claude_version:
organization / workspace / provider / model:
required_geo / allowed_geos / default_geo / observed_inference_geo:
fixture / expected_result / actual_result:
input / output / cache_write / cache_read:
standard_formula / expected_us_cost / rounding_tolerance:
/cost / status_line / max_budget_usd / terminal_outcome:
provider_usage_record / reporting_window / variance_reason:
decision: hold | investigate | raise-margin | limited-rollout | promote

FAQ

Did Claude Code 2.1.239 introduce the 1.1× premium?

No. Anthropic already documented the premium. Version 2.1.239 makes Claude Code's three cost-estimate surfaces include it for data-residency workspaces.

Does the premium apply to Claude subscription seat fees?

Anthropic's Enterprise help article says seat fees do not change; the multiplier applies to usage. Contract-specific terms can differ, so the organization bill remains authoritative.

Can `--max-budget-usd` replace an organization spend limit?

No. It is a task-level guard based on the client-visible estimate. Keep provider-side spend limits, alerts, access controls, and task concurrency limits as separate controls.

Should I switch to global inference to avoid the premium?

Only if the data owner and compliance policy explicitly allow it. Cost cannot override a residency requirement. If the requirement is unclear, hold the route and get a policy decision before changing geography.