Thursday · June 26, 2025

Anthropic Claude in Amazon Bedrock Cleared for FedRAMP High and DoD IL-4/5 in AWS GovCloud

confirmed Importance 8/10 ~4 min read 2 sources 5 actors

Opening

On June 26, 2025, Amazon Web Services and Anthropic announced that Claude 3.5 Sonnet v1 and Claude 3 Haiku in Amazon Bedrock had been authorized for FedRAMP High and Department of Defense Impact Level 4 / Impact Level 5 workloads in AWS GovCloud (US). AWS publicly described itself as the first cloud provider to clear these authorization tiers for foundation models — completing the federal AI-substrate gating event of the second-administration buildout. By the close of June 2025, both Azure OpenAI Service (FedRAMP High since September 2024; Top Secret since January 2025) and AWS GovCloud Claude were FedRAMP-High and IL-5 authorized for federal customers.

What Happened

The June 26 authorization documented:

  1. Claude 3.5 Sonnet v1 — Anthropic’s strongest publicly-available model class as of mid-2025, available through Amazon Bedrock’s inference API in AWS GovCloud (US-West and US-East regions).
  2. Claude 3 Haiku — Anthropic’s faster, less-expensive model class available through the same channel.
  3. Meta’s Llama models — also cleared for FedRAMP High and IL-4/5 in the same AWS announcement.

The Anthropic federal posture is deployed through AWS GovCloud and through Amazon Bedrock, not through a direct Anthropic-operated federal endpoint. Federal customers can call Claude inference via Bedrock APIs from inside their AWS GovCloud tenancy; the underlying model weights are operated by AWS as the host vendor under AWS’s existing FedRAMP and IL-4/5 authorization regime.

This is categorically different infrastructure from the OpenAI federal posture: OpenAI deploys via Microsoft Azure OpenAI Service (Microsoft-operated) inside Azure Government, with separate authorization documents, separate inference-routing topologies, and a separate host-vendor compliance regime. Reporting on “AI for the federal government” that elides this distinction is structurally misleading.

Why This Event Matters

By end of June 2025 — five months into the second administration — both major hyperscaler regions (Azure Government via OpenAI; AWS GovCloud via Anthropic) had cleared FedRAMP High and DoD IL-5 authorization for frontier-model LLM inference. This is the engineering-substrate gating event documented in Layer 3 of tech-stack-of-surveillance-inc-engineering-substrate-2010-2026.

The structural significances are:

  1. Federal customers can now run RAG-based surveillance applications at IL-5. Agency case files, interview transcripts, FOIA responses, social-media archives, and broker-acquired data can be vector-embedded and made available as context to an LLM at inference time, all inside FedRAMP-High and IL-5 boundaries. The architectural pattern (retrieval-augmented surveillance) is now contractually supported on both major hyperscalers.

  2. Two-vendor frontier-AI federal substrate locked in. The OpenAI-on-Azure-Government and Anthropic-on-AWS-GovCloud deployments are the two options for federal customers wanting commercial-grade frontier-model inference at classified-tier authorization. The substrate is gated to two vendors.

  3. The June 2025 cluster context. This authorization sits in the same month as:

    • June 1, 2025 — Palantir wins $30M ICE ImmigrationOS contract
    • June 5, 2025 — Anduril Series G with Founders Fund $1B check (largest in fund history)
    • June 17, 2025 — Wyden-Goldman-Velázquez letter to Karp on ICE surveillance
    • June 26, 2025 — this event

    The cluster makes June 2025 the single most consequential month of the trump-2-surveillance-procurement-chronology-2025-2026 consolidation.

Broader Context

The June 26 authorization was followed five months later by the November 12, 2025 addition of Claude Sonnet 4.5 to Amazon Bedrock in AWS GovCloud (US-West and US-East) via U.S.-Gov Cross-Region Inference — extending federal-customer access to the strongest Claude model class as of late 2025.

The subsequent Anthropic-Pentagon dispute (February-March 2026, per 2026-02-24–hegseth-threatens-anthropic-blacklist-military-ai-access through 2026-03-09–anthropic-files-dual-lawsuits-pentagon-supply-chain-designation) was triggered by the Pentagon’s attempt to overcome Anthropic’s “red lines” around use cases — but the underlying Anthropic federal substrate authorization (this entry) was not formally rescinded. The dispute is over use-case scoping, not authorization-tier gating.

Per the tech-stack-of-surveillance-inc-engineering-substrate-2010-2026 theme: the categorically important observation is not that AWS, Azure, and GCP coordinate. They compete; AWS, Azure, and GCP fight bitterly over individual JWCC task orders. The structural fact is that each holds enough federal contract volume to be a critical-infrastructure dependency, and the federal AI-inference layer is gated to two foundation-model vendors (OpenAI / Anthropic) with Meta Llama as a third on the open-weights track.

Research Gaps

  • Specific federal customer adoption — which DoD components, intelligence-community clients, and civilian-agency teams have actually deployed Claude-via-Bedrock in AWS GovCloud since June 2025?
  • Total Anthropic federal contract volume — pre-Pentagon-dispute (Feb 2026), what was the dollar value of Anthropic federal customer commitments through Bedrock?
  • Anthropic Use Case Policy enforcement at the federal-deployment level — how does Anthropic enforce its use-case policy when models are operated by AWS as host? The boundary is the legal substrate of the February 2026 Pentagon dispute.
  • The classified-tier AWS regions (Secret Region, Top Secret Region) — Claude availability through Amazon SageMaker in those regions is referenced in some documentation; primary-source dates and authorization-tier specifics deserve research.

Sources & Citations

[1] Claude in Amazon Bedrock: Approved for Use in FedRAMP High and DoD IL4/5 Workloads · Jun 26, 2025 Tier 1
[2] AWS GovCloud gets high-level security approvals for Anthropic and Meta AI models · Jun 26, 2025 Tier 2
Tiers Tier 1 court records & gov docs · Tier 2 established outlets · Tier 3 regional & specialty press · Tier 4 opinion or single-source. Methodology →
Cite this entry
The Cascade Ledger. “Anthropic Claude in Amazon Bedrock Cleared for FedRAMP High and DoD IL-4/5 in AWS GovCloud.” The Capture Cascade Timeline, June 26, 2025. https://capturecascade.org/event/2025-06-26--anthropic-claude-bedrock-fedramp-high-aws-govcloud-il-4-5/