Product

AWS Bedrock

AWS's multi-model token API platform — Anthropic Claude's primary venue, with Llama / Mistral as side menu; bundles token API to enterprises via AWS distribution + IAM / VPC.

1. Core Product / Service

Bedrock is AWS's unified token API entry point at L3b, GA'd in 2023, with continuous model menu expansion through 2024–2026:

  • Model menu: Anthropic Claude Sonnet 4 / Claude Opus 4 / Claude Haiku 4.5 (Claude is on Bedrock as its core exclusive/semi-exclusive distribution channel); Meta Llama 3 / 3.1 / 3.3; Mistral Large / Mixtral; Cohere Command R+; Amazon Nova (first-party series); Stability SDXL; DeepSeek R1 (added 2025).
  • Inference modes: On-Demand (per-token billing), Provisioned Throughput (reserved capacity, hourly), Batch (asynchronous, ~50% discount), Cross-Region Inference (auto cross-region routing).
  • Enterprise features: VPC PrivateLink, KMS encryption, IAM policy, AWS Marketplace billing, HIPAA / FedRAMP / SOC 2 compliance, CloudWatch / CloudTrail audit — these are why enterprises choose Bedrock instead of connecting to Anthropic directly.
  • Bedrock Agents / Knowledge Bases / Guardrails: upper-layer RAG / agent frameworks; decoupled from token API, primarily GTM tools to stick token usage.
  • Amazon Nova series: Nova Micro / Lite / Pro / Premier — Amazon's first-party trained models, priced as low-cost alternatives positioned against Claude Haiku / Sonnet.

2. Target Users & Pain Points

  • AWS large customers: have already signed annual AWS contracts, data flows / data in S3, unwilling to sign separate Anthropic contract + send data out of AWS.
  • Regulated industries: finance, healthcare, government — Bedrock's compliance suite + IAM integration is the core sales pitch.
  • Cross-model hedging teams: developers who want to switch between Claude / Llama / Mistral with a single SDK.
  • Pain points: direct Anthropic API calls require data egress from AWS, slowing enterprise compliance approval; Bedrock stays inside the VPC → the only acceptable path for some enterprises.

3. Competitive Landscape

Competitor Positioning Vs. Bedrock
azure-openai-service Microsoft's 1P/3P entry, OpenAI exclusive Azure locks GPT, Bedrock locks Claude — mirror positions
gcp-vertex Google's 1P/3P entry, Gemini + Claude + Mistral Vertex has the most varied menu; Bedrock leans Claude + in-house Nova
Anthropic API direct 1P direct Direct is cheaper (no take rate), but enterprise compliance usually pushes Bedrock
together-ai / fireworks-ai 3P open-source token API Lower prices; but enterprise IAM / VPC integration is weak
AWS SageMaker JumpStart Self-hosted models JumpStart is flexible but requires endpoint management; Bedrock is fully managed

Differentiation: Anthropic Claude flagship distribution + AWS sales muscle + enterprise compliance suite + Nova in-house cost weapon.

4. Unique Observations

  • Per-token pricing (On-Demand, 2026-05):
    • Claude Sonnet 4 (Bedrock): $3/M input + $15/M output — complete parity with Anthropic direct API (Anthropic direct same price) [1]
    • Claude Haiku 4.5: $1/M input + $5/M output — same parity
    • Claude Opus 4: $15/M input + $75/M output — same parity
    • Llama 3.3 70B: ~$0.72/M input + $0.72/M output (blended ~$0.72/M), slightly more expensive than Together / Fireworks' ~$0.88/M (Bedrock actually cheaper or on par in the Llama tier)
    • Mistral Large 2: ~$2/M input + $6/M output
    • Amazon Nova Pro: $0.80/M input + $3.20/M output — significantly cheaper than Claude Sonnet 4, AWS's in-house weapon
  • vs 1P price gap (take rate): Claude on Bedrock vs Anthropic direct supply has price parity — this is the key fact. AWS doesn't directly take a cut, but recoups value via (1) locking customers into the AWS ecosystem, (2) selling other AWS services (S3 / EC2 / private endpoints), (3) through the Anthropic investment relationship (Amazon has invested $8B cumulatively in Anthropic) in the form of "investment return + platform traffic" [4]. In short: take rate is not in token sale price, hidden in investment returns + ecosystem lock-in.
  • vs 3P peers: Llama 3.3 70B on Bedrock is on par with or slightly lower than Together / Fireworks — reflecting AWS using Anthropic profits to subsidize open-source model pricing and pull customers back from Together.
  • Inference engine: not publicly disclosed. Based on AWS Trainium / Inferentia2 in-house chips + partial H100 mix; Anthropic models have been disclosed multiple times in 2024-2025 to be partially served on Trainium2, with AWS using in-house chips to amortize costs as another aspect of Bedrock's economics.
  • Compute sourcing: 100% AWS first-party data centers, mix of in-house Trainium / Inferentia / NVIDIA. The only 3P platform integrating L1+L2+L3b end-to-end (same as Azure OpenAI / Vertex, but with better cost structure for compute than standalone 3P players).
  • Strategy: Bedrock is AWS's answer to "what if we lose IaaS distribution in the AI wave?" Anthropic strategic investment + Bedrock semi-exclusivity = turning Anthropic traffic into AWS traffic. If Anthropic continues to lead in models (Claude 4 / Opus 4), Bedrock is one of AWS's most profitable products.
  • Risk: Anthropic direct API + Vertex Claude dual supply means Bedrock's Claude is no longer physically exclusive; OpenAI / Google models aren't on Bedrock, so cross-model hedging customers can only go to Vertex; whether Amazon Nova can long-term close the gap with Claude / GPT determines whether Bedrock can shake off its "Claude reseller" status.

5. Financials / Business Scale

  • GA date: 2023-09
  • Bedrock revenue not separately disclosed; AWS overall Q4 2025 AI business reported "multi-billion dollar annual run-rate" — including Bedrock + SageMaker + Trainium chip rev.
  • Anthropic strategic investment: Amazon has invested $4B (2023) + $4B (2024) = $8B cumulatively in Anthropic, one of the largest AI strategic stakes globally [4]. Bedrock is the distribution vehicle for this investment.
  • Customers: AWS self-reports "tens of thousands of enterprise customers," including Pfizer, Cisco, Siemens, GoDaddy, NetSmart, KONE.

6. People & Relationships

  • Parent: Amazon Web Services (AWS) — Adam Selipsky (former CEO, departed 2024) → Matt Garman (current CEO).
  • Bedrock business lead: Atul Deo (VP of Bedrock, frequent conference speaker).
  • Strategic partners: Anthropic (deep, including investment + Claude exclusive / semi-exclusive), Meta (Llama sync release), Mistral, Cohere, Stability AI, AI21 Labs.
  • Competes with: azure-openai-service, gcp-vertex, together-ai, fireworks-ai, Anthropic direct API.
  • Hosts models from: Anthropic, Meta, Mistral, Cohere, AI21, Stability AI, deepseek (added 2025), Amazon (Nova first-party).

Sources

Last compiled: 2026-05-10