Fix us-gov-{east,west}-1 Bedrock Sonnet 4.5 pricing by Cyberfilo · Pull Request #28573 · BerriAI/litellm

Cyberfilo · 2026-05-22T09:59:10Z

Relevant issues

Pre-Submission checklist

Added testing in `tests/test_litellm/` — `tests/test_litellm/test_bedrock_usgov_pricing.py` (6 cases, all pass).
My PR passes all unit tests on `make test-unit`.
My PR's scope is as isolated as possible — touches only the five us-gov sonnet-4-5 entries.
Greptile review pending — will comment `@greptileai` after open.

Type

🐛 Bug Fix

Changes

AWS GovCloud Bedrock pricing carries a +20% premium over the global Anthropic rates, not the +10% premium that applies to commercial US regions. Litellm's five us-gov Claude Sonnet 4.5 entries mirror the commercial-US prices, undercharging customers by ~9%.

Per the AWS Bedrock pricing page:

Field	Old (commercial US value)	New (AWS GovCloud value)
`input_cost_per_token`	3.3e-06	3.6e-06 ($3.60/MTok)
`output_cost_per_token`	1.65e-05	1.8e-05 ($18.00/MTok)
`cache_creation_input_token_cost`	4.125e-06	4.5e-06 ($4.50/MTok)
`cache_read_input_token_cost`	3.3e-07	3.6e-07 ($0.36/MTok)
`cache_creation_input_token_cost_above_1hr`	(missing)	7.2e-06 ($7.20/MTok)

Files

`model_prices_and_context_window.json` — corrects five us-gov sonnet-4-5 entries.
`litellm/model_prices_and_context_window_backup.json` — same change, 1:1 mirror.
`tests/test_litellm/test_bedrock_usgov_pricing.py` — new regression test with six parametrized cases. Includes a 1.2× global→us-gov ratio invariant matching the style of the existing `test_bedrock_anthropic_1hr_cache_pricing.py` and `test_vertex_anthropic_1hr_cache_pricing.py` tests.

Entries corrected

`bedrock/us-gov-east-1/anthropic.claude-sonnet-4-5-20250929-v1:0`
`bedrock/us-gov-west-1/anthropic.claude-sonnet-4-5-20250929-v1:0`
`bedrock/us-gov-east-1/claude-sonnet-4-5-20250929-v1:0` (alias without the `anthropic.` prefix)
`bedrock/us-gov-west-1/claude-sonnet-4-5-20250929-v1:0` (alias without the `anthropic.` prefix)
`us-gov.anthropic.claude-sonnet-4-5-20250929-v1:0` (cross-region inference profile)

Test plan

```sh
$ python -m pytest tests/test_litellm/test_bedrock_usgov_pricing.py -v
...
============================== 6 passed in 0.32s ===============================
```

Black + ruff clean. Diff is +50 lines, -40 lines (each corrected entry has one fewer line because the 1-hour cache field is now inserted but the 5m field already had a trailing comma — net is balanced; the 10 added lines are the 1-hour field across the 5 entries × 2 files).

Note on scope

The same +10% / +20% confusion likely affects us-gov entries for Haiku 4.5 and other Bedrock-hosted Anthropic models in the GovCloud regions. I limited this PR to the model specifically named in #27120 to keep scope tight; happy to extend coverage in a follow-up.

greptile-apps · 2026-05-22T10:01:29Z

Greptile Summary

This PR corrects the AWS GovCloud (us-gov-east-1, us-gov-west-1) Bedrock pricing for Claude Sonnet 4.5 from the commercial-US +10% premium to the correct GovCloud +20% premium over global Anthropic rates, and adds a 1-hour cache write tier that was previously absent. All five affected entries have been verified as exactly 1.2× the global rates.

Pricing data (model_prices_and_context_window.json + backup): 4 bedrock/us-gov-*/ entries and the us-gov.anthropic.* cross-region inference profile updated; the cross-region profile also has its _above_200k_tokens tier fields corrected and cache_creation_input_token_cost_above_1hr_above_200k_tokens added.
Tests (test_bedrock_usgov_pricing.py): New parametrized regression suite covering all 5 model keys, all base pricing fields, the 5 _above_200k_tokens fields on the cross-region profile, and the 1.2× ratio invariant against the global entry.

Confidence Score: 5/5

Safe to merge — purely a JSON pricing data correction with no logic changes and a comprehensive new test suite that locks in the correct values.

All updated pricing values verify exactly as 1.2× the global rates. The new tests cover every field that was previously wrong, including the _above_200k_tokens tier on the cross-region profile that was flagged in the prior review round. Both JSON files are kept in sync. No code paths are modified.

No files require special attention.

Important Files Changed

Filename	Overview
model_prices_and_context_window.json	Corrects 5 us-gov Sonnet 4.5 entries to the +20% GovCloud premium; all updated values verify as exactly 1.2× the global rates
litellm/model_prices_and_context_window_backup.json	1:1 mirror of main JSON — same 5 entries corrected with identical values
tests/test_litellm/test_bedrock_usgov_pricing.py	New regression tests covering all 5 us-gov model keys, base pricing fields, 200k-tier fields on the cross-region profile, and the 1.2× ratio invariant

_{Reviews (4): Last reviewed commit: "Apply us-gov 200k-tier GovCloud premium ..." | Re-trigger Greptile}

codecov · 2026-05-22T10:03:52Z

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

oss-pr-review-agent-shin · 2026-05-22T10:09:38Z

🤖 litellm-agent: This PR is currently BLOCKED from merge.

Score: 2/5 ❌

Why blocked:

1 PR-related CI failure (Greptile gate: score 3/5 below required 4/5 — request a Greptile review (@greptileai) and resolve its comments before maintainer review.) (pr_related_failures, -2 pts)
Greptile 3/5 (greptile_low, -1 pts)

Details: Score docked for: 1 PR-related CI failure (Greptile gate: score 3/5 below required 4/5 — request a Greptile review (@greptileai) and resolve its comments before maintainer review.); Greptile 3/5.

Fix the issues above and push an update — the bot will re-review automatically.

Note: This bot is still in beta and might not always work as expected. Please share any feedback via Slack.

…riAI#28598) Squash-merged by litellm-agent from Terrajlz's PR.

Squash-merged by litellm-agent from devauxbr's PR.

Cyberfilo · 2026-05-25T07:20:43Z

Proof of fix

This is a pure data / pricing-JSON change with no visible runtime output — adding the screenshot-exempt justification by surfacing the regression-test run instead.

$ python -m pytest tests/test_litellm/test_*bedrock_usgov_pricing.py -v
... 6 passed in 0.32s ...

JSON validity is also enforced by the existing test_aaamodel_prices_and_context_window_json_is_valid test. Black + ruff clean.

Requesting screenshot-exempt triage per the gate's own instructions for data-only PRs.

Cyberfilo · 2026-05-25T08:13:43Z

@greptileai

) AWS GovCloud Bedrock pricing carries a +20% premium over the global Anthropic rates, not the +10% premium that applies to commercial US regions. Litellm's five us-gov Claude Sonnet 4.5 entries mirror the commercial-US prices (3.3e-06 input, 1.65e-05 output, 4.125e-06 cache write), undercharging customers by ~9%. Corrects all five affected entries in both model_prices_and_context_window.json and the bundled model_prices_and_context_window_backup.json to match AWS's published GovCloud rates (per https://aws.amazon.com/bedrock/pricing/): - input_cost_per_token 3.3e-06 -> 3.6e-06 ($3.60/MTok) - output_cost_per_token 1.65e-05 -> 1.8e-05 ($18.00/MTok) - cache_creation_input_token_cost 4.125e-06 -> 4.5e-06 ($4.50/MTok) - cache_read_input_token_cost 3.3e-07 -> 3.6e-07 ($0.36/MTok) - cache_creation_input_token_cost_above_1hr (added) ($7.20/MTok) Affected entries: - bedrock/us-gov-east-1/anthropic.claude-sonnet-4-5-20250929-v1:0 - bedrock/us-gov-west-1/anthropic.claude-sonnet-4-5-20250929-v1:0 - bedrock/us-gov-east-1/claude-sonnet-4-5-20250929-v1:0 - bedrock/us-gov-west-1/claude-sonnet-4-5-20250929-v1:0 - us-gov.anthropic.claude-sonnet-4-5-20250929-v1:0 New regression test tests/test_litellm/test_bedrock_usgov_pricing.py pins the corrected rates across all five entries and asserts the 1.2x global-to-us-gov ratio invariant, matching the pattern used by the existing Vertex and EU/AU/JP Bedrock pricing tests.

veria-ai · 2026-05-25T08:29:12Z

        {{- end }}
        {{- with .Values.podAnnotations }}
-        {{- toYaml . | nindent 8 }}
+        {{- tpl (toYaml .) $ | nindent 8 }}


Medium: Helm annotation templates can expose render-time data

tpl evaluates annotation values as Helm templates, so a user who can set podAnnotations can add expressions like {{ .Values.postgresql.auth.password }} or {{ lookup "v1" "Secret" "ns" "name" }} and have Helm write the result into Deployment/Pod metadata. Keep annotations as literal values, or move templated annotations behind a separate opt-in value with a clear warning that it executes templates with Helm's rendering privileges.

Suggested change

{{- tpl (toYaml .) $ | nindent 8 }}

{{- toYaml . | nindent 8 }}

veria-ai · 2026-05-25T08:30:00Z

PR overview

One security issue remains open in the Helm deployment template: annotation values are evaluated as templates. A user who can control podAnnotations could cause Helm to render sensitive values or cluster data into pod metadata, creating a concrete risk of secret exposure. No issues have been addressed yet, so the PR’s security posture still depends on making annotations literal by default or clearly isolating templated annotations behind an opt-in path.

Open issues (1)

deploy/charts/litellm-helm/templates/deployment.yaml:33 — Medium: Helm annotation templates can expose render-time data tpl evaluates annotation values as Helm templates, so a user who can set

Fixed/addressed: 0 · PR risk: 7/10

Cyberfilo · 2026-05-25T10:59:13Z

@greptileai please review commit e1c947a16b — addresses the P2 finding on us-gov.anthropic.* 200k-tier rates. Four existing fields updated from the +10% commercial-US rate to the +20% GovCloud premium, and the missing cache_creation_input_token_cost_above_1hr_above_200k_tokens field added for parity with the us. cross-region profile. Test parametrizes all 5 fields and asserts the 1.2x ratio invariant against anthropic.claude-sonnet-4-5-*.

The us-gov.anthropic.claude-sonnet-4-5-20250929-v1:0 cross-region inference profile carried _above_200k_tokens fields at the +10% commercial-US rate while the base tier was correctly at +20%. AWS GovCloud pricing applies the same +20% uplift across all tiers, so long-context requests through this profile were undercharging by ~9%. Updates four existing fields (input/output/cache_creation/cache_read _above_200k_tokens) and adds cache_creation_input_token_cost_above_1hr _above_200k_tokens for parity with the us. cross-region entry. Extends test_bedrock_usgov_pricing.py with five parametrized field assertions plus a 1.2x ratio invariant across all 200k-tier fields, so any future regression on either dimension is caught.

Cyberfilo · 2026-05-25T11:29:42Z

@greptileai please review commit 7a255ba896 — re-pushed to invalidate the prior review (which is now showing as edited in shin-agent's verifier). The P2 finding on us-gov.anthropic.* 200k-tier rates is addressed: 4 fields updated to +20% GovCloud premium and a 5th (_above_1hr_above_200k_tokens) added for parity with the us. cross-region profile. Test parametrizes all 5 fields + 1.2× ratio invariant.

…5-pricing

greptile-apps Bot reviewed May 22, 2026

View reviewed changes

Comment thread tests/test_litellm/test_bedrock_usgov_pricing.py

Terrajlz and others added 2 commits May 22, 2026 13:45

fix(thinking): handle None thinking param in is_thinking_enabled (Ber…

a5939c8

…riAI#28598) Squash-merged by litellm-agent from Terrajlz's PR.

feat(helm): support tpl rendering in podAnnotations (BerriAI#28609)

2c144fc

Squash-merged by litellm-agent from devauxbr's PR.

Cyberfilo changed the base branch from litellm_internal_staging to shin_agent_oss_staging_05_22_2026 May 25, 2026 08:17

Cyberfilo force-pushed the fix/27120-us-gov-bedrock-sonnet-4-5-pricing branch from ff3ed48 to 6cc1dad Compare May 25, 2026 08:17

chore: trigger shin-agent re-eval on retargeted staging base

6e55e6f

veria-ai Bot reviewed May 25, 2026

View reviewed changes

Cyberfilo force-pushed the fix/27120-us-gov-bedrock-sonnet-4-5-pricing branch from e1c947a to 7a255ba Compare May 25, 2026 11:28

chore: trigger shin-agent re-eval against updated Greptile state

6cb95ad

Sameerlite changed the base branch from shin_agent_oss_staging_05_22_2026 to litellm_010626 June 1, 2026 11:07

Merge branch 'litellm_010626' into fix/27120-us-gov-bedrock-sonnet-4-…

4e2f811

…5-pricing

Sameerlite merged commit b26b015 into BerriAI:litellm_010626 Jun 1, 2026
32 checks passed

Cyberfilo mentioned this pull request Jun 9, 2026

Preserve original PR authorship when squashing staging branches into main #30012

Open

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Fix us-gov-{east,west}-1 Bedrock Sonnet 4.5 pricing#28573

Fix us-gov-{east,west}-1 Bedrock Sonnet 4.5 pricing#28573
Sameerlite merged 7 commits into
BerriAI:litellm_010626from
Cyberfilo:fix/27120-us-gov-bedrock-sonnet-4-5-pricing

Cyberfilo commented May 22, 2026

Uh oh!

greptile-apps Bot commented May 22, 2026 •

edited

Loading

Important Files Changed

Uh oh!

Uh oh!

codecov Bot commented May 22, 2026

Uh oh!

oss-pr-review-agent-shin Bot commented May 22, 2026

Uh oh!

Cyberfilo commented May 25, 2026

Uh oh!

Cyberfilo commented May 25, 2026

Uh oh!

veria-ai Bot May 25, 2026

Uh oh!

veria-ai Bot commented May 25, 2026

Uh oh!

Cyberfilo commented May 25, 2026

Uh oh!

Cyberfilo commented May 25, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	{{- tpl (toYaml .) $ \| nindent 8 }}
	{{- toYaml . \| nindent 8 }}

Uh oh!

Conversation

Cyberfilo commented May 22, 2026

Relevant issues

Pre-Submission checklist

Type

Changes

Files

Entries corrected

Test plan

Note on scope

Uh oh!

greptile-apps Bot commented May 22, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Greptile Summary

Confidence Score: 5/5

Important Files Changed

Uh oh!

Uh oh!

codecov Bot commented May 22, 2026

Codecov Report

Uh oh!

oss-pr-review-agent-shin Bot commented May 22, 2026

Uh oh!

Cyberfilo commented May 25, 2026

Proof of fix

Uh oh!

Cyberfilo commented May 25, 2026

Uh oh!

veria-ai Bot May 25, 2026

Choose a reason for hiding this comment

Uh oh!

veria-ai Bot commented May 25, 2026

PR overview

Open issues (1)

Uh oh!

Cyberfilo commented May 25, 2026

Uh oh!

Cyberfilo commented May 25, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

greptile-apps Bot commented May 22, 2026 •

edited

Loading