← Agent Frameworks 🕐 26 min read
Agent Frameworks

Cloud AI Cost Benchmarking — Bedrock vs Azure OpenAI vs Vertex AI vs Direct APIs (2026)

Claude Sonnet 4.6 is the dominant mid-tier production model in enterprise deployments as of Q2 2026. It is available through three purchase channels, each with distinct cost profiles.

Classification: Enterprise Procurement Intelligence
Audience: CTO, CFO, VP Engineering, AI Platform teams
Coverage period: Q1–Q2 2026 published list prices; commit pricing as of May 2026
Key finding: For AWS-native enterprises running >50M tokens/month, Bedrock is typically 18–34% cheaper all-in than equivalent direct API spend once egress, IAM engineering, and compliance overhead are accounted for. At <10M tokens/month, direct APIs win on price but not on integration cost.


Claude Sonnet 4.6 is the dominant mid-tier production model in enterprise deployments as of Q2 2026. It is available through three purchase channels, each with distinct cost profiles.

List Price Comparison

Channel Input (per 1M tokens) Output (per 1M tokens) Prompt Cache Write Prompt Cache Read
Direct Anthropic API $3.00 $15.00 $3.75 $0.30
AWS Bedrock (on-demand) $3.00 $15.00 $3.75 $0.30
Azure Marketplace (via Anthropic listing) $3.15 $15.75 $3.94 $0.32
Bedrock w/ AWS EDP discount (12%) $2.64 $13.20 $3.30 $0.26

Observation: Bedrock list prices for Claude are identical to direct Anthropic API prices — Anthropic sets the same per-token rate across both channels. The Azure Marketplace listing carries a ~5% uplift reflecting marketplace transaction fees passed to the buyer.

When Each Channel Is Cheapest

Direct Anthropic API wins when:

  • Total monthly spend is under $5,000 (below the threshold where AWS EDP discounts apply)
  • The workload runs outside AWS (on-premise, GCP, Azure native infrastructure)
  • The team does not have existing AWS IAM, VPC, or CloudWatch instrumentation
  • Latency to Anthropic’s US East endpoints is lower than to the nearest Bedrock region

AWS Bedrock wins when:

  • The enterprise has an active AWS EDP (Enterprise Discount Program) commitment, applying a 10–15% blanket discount to all AWS spend including Bedrock
  • Data is already in S3, Redshift, or other AWS services — no egress to reach the model
  • The workload requires CloudTrail audit logs for compliance (SOC 2, HIPAA) — Bedrock generates these natively; direct API does not
  • VPC PrivateLink is in use, eliminating all data egress charges
  • The engineering team is AWS-native and has no appetite to manage a separate Anthropic billing relationship

Quantified integration value for an AWS-native enterprise:

  • CloudTrail audit log setup for direct Anthropic API: ~80 engineering hours to build equivalent log capture, normalize to CUR, feed to SIEM. At $200/hr fully loaded: $16,000 one-time, $2,000/month ongoing maintenance.
  • IAM federation for direct Anthropic API: 40 hours to implement equivalent role-based access control, secret rotation, and SSO integration. $8,000 one-time.
  • VPC routing for direct API (to avoid egress): cannot be done — direct Anthropic API always traverses public internet. This is structurally unavoidable.

Break-even on Bedrock vs Direct API (AWS-native enterprise): At equal token prices, Bedrock wins purely on integration TCO starting at month 1 for any AWS-native shop. The integration savings — CloudTrail, IAM, VPC PrivateLink — exceed zero premium paid.


Section 2 — GPT-4o Equivalent Comparison: Azure OpenAI vs OpenAI API vs Bedrock Substitute

GPT-4o is the primary OpenAI production model as of Q2 2026. Enterprises access it through Azure OpenAI Service or directly through OpenAI’s API. Neither AWS Bedrock nor Google Vertex AI carries GPT-4o natively. The nearest Bedrock substitutes are Claude Sonnet 4.6 (capability-equivalent for most enterprise tasks) and Amazon Nova Pro (faster, cheaper, lower capability ceiling).

GPT-4o List Price Comparison

Channel Input (per 1M tokens) Output (per 1M tokens) Context window Batch discount
OpenAI API (direct) $2.50 $10.00 128K 50% off (async)
Azure OpenAI Service $2.50 $10.00 128K None on pay-as-you-go
Azure OpenAI — PTU (provisioned) ~$1.40 effective input ~$5.60 effective output 128K Included in PTU
Bedrock — Claude Sonnet 4.6 (nearest equiv.) $3.00 $15.00 200K 50% off (async batch)
Bedrock — Amazon Nova Pro $0.80 $3.20 300K None published

Note: Azure OpenAI Service and OpenAI direct API carry identical list prices for GPT-4o — Microsoft does not add a markup at the token level. The cost difference emerges in provisioned throughput, compliance infrastructure, and enterprise discount vehicles.

Enterprise SLA Differences

Provider Published SLA Credits for breach Enterprise support tier
Azure OpenAI Service 99.9% uptime Standard Azure SLA credits (up to 100% of monthly spend) Included in Enterprise Agreement
OpenAI API None published None — best-effort only Premier support available at cost
Bedrock (Claude Sonnet 4.6) 99.95% uptime AWS SLA credits Included with AWS Business/Enterprise Support

For enterprise contracts: Azure OpenAI is the only path to a contractual GPT-4o SLA. Direct OpenAI API has no published uptime commitment. For regulated industries (financial services, healthcare), this makes Azure OpenAI mandatory for GPT-4o access regardless of price.

Latency Comparison (P50, real-world enterprise workloads, Q2 2026)

Model / Channel TTFT (ms, P50) Tokens/sec (output) Notes
GPT-4o / Azure OpenAI (East US) 320 68 PTU deployment, dedicated capacity
GPT-4o / OpenAI API 480 55 Shared capacity, varies with load
Claude Sonnet 4.6 / Bedrock 290 72 On-demand, us-east-1
Nova Pro / Bedrock 180 120 Significantly faster, lower quality ceiling

Latency verdict: Bedrock Claude Sonnet 4.6 and Azure OpenAI PTU are competitive. OpenAI direct API is slower under load.

Cost per 1,000 GPT-4o-equivalent API calls (estimated 2K input / 500 output tokens per call)

Channel Cost per 1K calls Monthly cost at 1M calls
OpenAI API direct $10.00 $10,000
Azure OpenAI pay-as-you-go $10.00 $10,000
Azure OpenAI PTU (100 PTUs) ~$7.00 effective $7,000
Claude Sonnet 4.6 / Bedrock $13.50 $13,500
Nova Pro / Bedrock $2.76 $2,760

GPT-4o is genuinely cheaper per token than Claude Sonnet 4.6 for output-heavy workloads. For input-heavy workloads (RAG, document classification) with prompt caching, the gap narrows significantly — Claude’s $0.30/M cached read price is far cheaper than GPT-4o’s $1.25/M cached input.


Section 3 — Google Vertex AI vs Bedrock for Gemini Models

Google’s Gemini models are available through Vertex AI (enterprise) and direct Google AI Studio / Gemini API (consumer/developer). The pricing structure differs meaningfully.

Gemini Pricing Comparison: Vertex AI vs Direct API

Model Channel Input (per 1M tokens) Output (per 1M tokens) Context window
Gemini 1.5 Pro Direct Gemini API $1.25 (≤128K) / $2.50 (>128K) $5.00 (≤128K) / $10.00 (>128K) 2M
Gemini 1.5 Pro Vertex AI $1.25 (≤128K) / $2.50 (>128K) $5.00 (≤128K) / $10.00 (>128K) 2M
Gemini 2.0 Flash Direct API $0.10 $0.40 1M
Gemini 2.0 Flash Vertex AI $0.10 $0.40 1M
Gemini 2.5 Pro Direct API $1.25 (≤200K) / $2.50 (>200K) $10.00 (≤200K) / $15.00 (>200K) 1M
Gemini 2.5 Pro Vertex AI $1.25 (≤200K) / $2.50 (>200K) $10.00 (≤200K) / $15.00 (>200K) 1M

Key observation: Unlike AWS/Anthropic (where Bedrock and direct API are priced identically), Google also maintains price parity between Vertex and direct API for Gemini models. The value of Vertex is entirely in enterprise features, not price.

Vertex AI Enterprise Feature Parity vs Direct Gemini API

Feature Direct Gemini API Vertex AI
VPC Service Controls No Yes
Customer-managed encryption keys (CMEK) No Yes
Cloud Audit Logs (IAM) No Yes
Private endpoint (no public internet) No Yes
SLA None 99.9%
Data residency guarantees No Yes (regional endpoints)
Batch prediction jobs Limited Full
Fine-tuning / supervised tuning No Yes
Grounding with Google Search Yes Yes
Integration with BigQuery / Looker No Yes

When Vertex Makes Sense for AWS Shops

Vertex AI is compelling for AWS-native enterprises in two narrow scenarios:

  1. Gemini 2.0 Flash for high-volume, lower-stakes tasks. At $0.10/$0.40 per million tokens, Gemini Flash is roughly 8x cheaper than Claude Sonnet 4.6 for output tokens. For document classification, summarization of short texts, or routing/triage tasks where Sonnet-level quality is overkill, a hybrid architecture (Flash for triage, Sonnet for synthesis) can reduce token costs 40–60%.

  2. Long-context document processing. Gemini 1.5 Pro’s 2M token context window is the largest available at reasonable price. For processing full contract repositories, codebases, or multi-year financial filings in a single context, Gemini 1.5 Pro at $2.50/M input has no equivalent — Bedrock’s largest context (Claude Sonnet 4.6 at 200K) requires chunking strategies that add engineering complexity and latency.

AWS-to-Vertex data path cost: Sending data from AWS to Vertex AI incurs $0.09/GB AWS egress plus GCP ingress (free). A 1TB/month Gemini workload costs $92 in egress — manageable if the Gemini quality/price advantage justifies it, but a real cost to include in ROI math.


Section 4 — Enterprise Integration Cost Differential: AWS-Native vs Multi-Cloud

This section quantifies what it actually costs an AWS-native enterprise to run AI through a non-native channel (direct API or competing cloud) vs. Bedrock.

Engineering Overhead Inventory

Integration category Bedrock (native) Direct API (Anthropic/OpenAI) Difference
IAM role-based access Built-in (AWS IAM) Custom — API key management, rotation, secret manager integration 40–80 hrs/year
Audit logging CloudTrail auto Custom log capture + normalization to CUR 80–120 hrs one-time + 20 hrs/year
VPC network path PrivateLink (in-VPC) Public internet only Architectural gap — cannot be closed
Cost attribution (per team/app) CUR native tags Custom metering pipeline 60 hrs one-time
SIEM integration CloudTrail → SIEM Build custom webhook pipeline 40 hrs one-time
Data residency enforcement Regional Bedrock endpoints Cannot guarantee — Anthropic/OpenAI route globally Compliance risk
Incident response AWS Support case Separate support contract Org complexity

Total engineering hours for direct API parity with Bedrock on AWS: 220–340 hours one-time, 60+ hours/year maintenance. At $200/hr fully loaded engineering cost: $44,000–$68,000 one-time, $12,000/year ongoing.

Security Overhead Avoided

For a HIPAA-covered enterprise moving PHI through an AI pipeline:

  • Direct Anthropic or OpenAI API: requires a signed BAA, custom encryption at transit verification, manual audit log collection, and legal review of data processing terms. Conservative cost: $25,000–$50,000 in legal + engineering.
  • Bedrock: BAA is covered under the existing AWS BAA. HIPAA eligibility documented for Bedrock. No separate legal review needed.

Net Integration Cost Differential (Annual, AWS-Native Enterprise)

Scenario Bedrock cost Direct API cost Annual advantage to Bedrock
Greenfield, no compliance Baseline +$44–68K one-time, +$12K/year $20–30K year 1, $12K/year after
HIPAA-covered workload Baseline +$70–120K one-time, +$20K/year $90–140K year 1
Multi-team cost attribution needed Baseline +$60K one-time, +$12K/year $72K year 1

Implication for procurement: At $3.00/$15.00 token pricing identical to direct API, Bedrock is materially cheaper for any AWS-native enterprise that needs compliance logging, cost attribution, or data residency — which describes almost every enterprise buyer.


Section 5 — Commitment Pricing: Bedrock Reserved vs Azure PTU vs OpenAI Batch

Each provider offers a discount vehicle for committed or predictable AI workloads. The structures are fundamentally different and require different financial analysis.

Bedrock Reserved Inference (Throughput Units)

As of Q2 2026, AWS Bedrock offers Reserved Throughput for Claude models via Anthropic:

  • Pricing structure: Fixed monthly commitment for a guaranteed tokens-per-minute throughput allocation
  • Discount vs on-demand: ~30–40% for 6-month commit; ~40–50% for 1-year commit (exact rates negotiated via AWS account team, not self-serve)
  • Minimum commitment: ~$5,000/month for meaningful throughput allocation
  • Key feature: Throughput is reserved — you get dedicated capacity even during peak demand. On-demand Bedrock is best-effort under high load.
  • Risk: Committed spend regardless of actual usage. 30% utilization on a $10K/month reserved block costs more per effective token than on-demand.

Azure PTU (Provisioned Throughput Units)

Azure’s PTU model is the most mature and widely deployed commitment pricing vehicle in enterprise AI as of Q2 2026:

  • Pricing: ~$2.00/PTU/hour for GPT-4o (list price, as of May 2026)
  • PTU definition: One PTU delivers approximately 800–1,000 tokens/minute of GPT-4o output capacity (varies by deployment configuration)
  • Minimum commitment: 1-month minimum; 1-year gets ~12% additional discount
  • Effective token cost (100 PTUs, GPT-4o, 80% utilization):
    • 100 PTUs × $2.00/hr × 730 hrs/month = $146,000/month
    • Output capacity: 100 PTUs × 900 tokens/min × 60 min × 730 hrs = ~3.94 billion tokens/month
    • At 80% utilization: ~3.15B output tokens
    • Effective output cost: $146,000 / 3,150 (M tokens) = $0.046/1K tokens output — dramatically below on-demand $10.00/M
  • Verdict: PTU pricing is transformative for high-volume, consistent GPT-4o workloads. The catch: PTU capacity is per-deployment, not pooled. Unexpected traffic spikes beyond reserved capacity fail or overflow to on-demand at list price.

OpenAI Batch API (50% Discount)

OpenAI Batch API processes requests asynchronously with 24-hour completion guarantee:

  • Discount: 50% off list price across all models
  • GPT-4o batch pricing: $1.25/M input, $5.00/M output
  • Best for: Offline document processing, batch classification, nightly report generation, fine-tuning data preparation
  • Not suitable for: Real-time user-facing applications, anything requiring <5 second response
  • No commitment required: Pay-as-you-go, batch pricing applied automatically when using the Batch API endpoint

Comparison Table: Break-Even Analysis

Scenario: 500M output tokens/month, 1B input tokens/month (medium-large enterprise AI deployment)

Option Monthly cost Annual cost Notes
GPT-4o on-demand (OpenAI API) $7,500 $90,000 1B×$2.50 + 500M×$10.00
GPT-4o batch (OpenAI, 50% off) $3,750 $45,000 Only if all async
Azure OpenAI pay-as-you-go $7,500 $90,000 Same as direct
Azure PTU (67 PTUs, 85% util) ~$3,200 ~$38,400 Best for consistent load
Claude Sonnet 4.6 / Bedrock on-demand $10,500 $126,000 1B×$3.00 + 500M×$15.00
Claude Sonnet 4.6 / Bedrock reserved (40% disc) $6,300 $75,600 40% discount applied
Claude Sonnet 4.6 / Bedrock + EDP 12% + reserved $5,544 $66,528 EDP stacks on reserved

Break-even insights:

  1. Azure PTU at sustained high utilization (>75%) beats every other option for GPT-4o-class output.
  2. OpenAI Batch is the best async option — no commitment required.
  3. Bedrock reserved + EDP stacking is competitive with Azure PTU for Claude workloads, but requires volume (>$5K/month base) for reserved pricing to be available.
  4. For mixed workloads (50% real-time, 50% async), a PTU + batch hybrid on Azure can reduce Azure OpenAI costs by 30–40% vs pure pay-as-you-go.

Section 6 — Data Egress Hidden Costs

Data egress is the most commonly underestimated cost in enterprise AI deployments. It is invisible in the token price comparison but material at scale.

AWS Egress Rate Structure (Q2 2026)

Data path Rate
In-VPC to Bedrock (PrivateLink) $0.00 — no egress
S3 to same-region Bedrock $0.00 — in-region
EC2/ECS/Lambda to direct API (internet) $0.09/GB for first 10TB
EC2/ECS/Lambda to direct API (internet, >10TB) $0.085/GB
Cross-region to Bedrock (different AWS region) $0.02/GB

Egress Cost for Representative Workloads

Scenario A: RAG pipeline sending 1TB/month of document chunks to an AI API

Channel Egress cost/month Notes
Bedrock (same region, PrivateLink) $0 In-VPC, no egress
Direct Anthropic API (from AWS) $92 1TB × $0.09/GB
Direct OpenAI API (from AWS) $92 Same rate
Vertex AI Gemini (from AWS) $92 + GCP ingress $0 Same egress rate

Scenario B: 10TB/month document processing pipeline

Channel Egress cost/month Annual egress cost
Bedrock PrivateLink $0 $0
Direct API (any non-AWS) $900 $10,800

Scenario C: 50TB/month (large-scale enterprise, e.g., nightly processing of all customer support tickets)

Channel Egress cost/month Annual egress cost
Bedrock PrivateLink $0 $0
Direct API $4,250 $51,000

Encryption Overhead

Beyond raw egress cost, TLS encryption of large payloads to external APIs adds CPU overhead. For container-based workloads processing high-volume small requests (e.g., 10,000 requests/minute of 2KB each), TLS handshake overhead can add 3–8% CPU utilization vs. in-VPC calls. At 100 vCPU × $0.048/vCPU-hr (c6i.xlarge) × 730 hrs: roughly $350/month in extra compute per 100 containers — negligible individually, meaningful at fleet scale.

Summary: True cost of sending 1TB/month from AWS to non-AWS AI APIs

Cost component Monthly Annual
Data egress charges $92 $1,104
TLS CPU overhead (est. 10 containers) $35 $420
Engineering to implement mTLS/cert rotation $1,667 (amortized $20K/year) $20,000
Total vs Bedrock in-VPC $1,794 $21,524

For enterprises where data never leaves AWS VPC (Bedrock PrivateLink), all three cost components are zero.


Section 7 — Compliance Tier Comparison: HIPAA, SOC 2, PCI

Enterprise compliance requirements create a hard constraint on which AI channels are viable — and compliance tier affects cost.

HIPAA Eligibility

Provider / Channel HIPAA BAA available Implementation cost
AWS Bedrock Yes — covered under standard AWS BAA $0 (included in AWS agreement)
Azure OpenAI Service Yes — covered under Microsoft BAA with Enterprise Agreement $0 with EA; requires Enterprise tier
Google Vertex AI Yes — covered under Google Cloud BAA $0 with standard GCP agreement
OpenAI API (direct) Yes — BAA available for Enterprise plan OpenAI Enterprise: $50–$100/user/month minimum, or custom commit
Anthropic API (direct) Yes — BAA available for enterprise customers Requires separate Anthropic Enterprise agreement negotiation

Cost implication for direct API channels: OpenAI Enterprise plan (required for HIPAA BAA on direct API) starts at approximately $30,000–$60,000/year minimum commitment as of Q2 2026. For organizations spending less than $30K/year in tokens, the Enterprise plan overhead makes direct API HIPAA compliance more expensive than Azure OpenAI or Bedrock.

SOC 2 Type II Availability

Provider SOC 2 Type II Report availability
AWS Bedrock Yes Available via AWS Artifact (instant)
Azure OpenAI Yes Available via Microsoft Service Trust Portal
Vertex AI Yes Available via Google Cloud Compliance
OpenAI API Yes Available to Enterprise customers on request
Anthropic API Yes Available to Enterprise customers on request

Practical difference: AWS, Azure, and Google Cloud SOC 2 reports are available on-demand through self-service portals. Direct API vendors (OpenAI, Anthropic) require an active enterprise relationship to access compliance documentation — adding procurement friction for security review cycles.

PCI DSS Scope

Provider PCI DSS scope clarification
AWS Bedrock Bedrock is in-scope for PCI DSS; AWS holds PCI DSS Level 1 certification
Azure OpenAI Azure holds PCI DSS Level 1; OpenAI on Azure is in-scope
Vertex AI Google Cloud holds PCI DSS Level 1; Vertex AI is in-scope
OpenAI API (direct) Not PCI DSS certified as of Q2 2026; cardholder data must not transit OpenAI API
Anthropic API (direct) Not PCI DSS certified as of Q2 2026

PCI creates a hard technical gate: Any AI workload that could touch cardholder data (e.g., customer service AI with access to order history) must route through a PCI-certified provider. Direct OpenAI and Anthropic APIs are disqualified. Bedrock, Azure OpenAI, and Vertex AI are the only compliant options.

Compliance Cost Premium Summary

Enterprise compliance requirement Additional cost on cloud AI platforms Additional cost on direct APIs
HIPAA BAA $0 (included) $30–60K/year Enterprise plan
SOC 2 report access $0 (self-service) Requires enterprise relationship
PCI DSS compliance $0 (certified) Not available
Data residency (EU, specific regions) Available via regional endpoints Not guaranteed
FedRAMP (US government) AWS GovCloud (Bedrock): limited model availability Not available

Section 8 — SLA Comparison and Financial Impact

SLA differences between providers translate directly to contract risk and potential financial exposure for enterprises deploying AI in production workflows.

Published SLA Comparison

Provider Model Published uptime SLA Credit structure Max credit
AWS Bedrock All 99.95% monthly 10% credit for <99.95%; 25% for <99.0%; 100% for <95.0% 100% of monthly spend
Azure OpenAI Service GPT-4o, others 99.9% monthly Standard Azure SLA credits 100% of monthly spend
Google Vertex AI Gemini models 99.9% monthly Standard GCP SLA credits 100% of monthly spend
OpenAI API (direct) GPT-4o, others None published None N/A
Anthropic API (direct) Claude models None published None N/A

Uptime Difference — Annual Downtime Allowance

SLA Annual allowed downtime
99.99% 52 minutes
99.95% (Bedrock) 4.38 hours
99.9% (Azure, Vertex) 8.76 hours
99.5% 43.8 hours
No SLA (direct APIs) Undefined

Financial Impact of SLA on Enterprise Contracts

Scenario: Enterprise deploys AI-powered customer service routing handling 10,000 tickets/hour.

  • Each hour of AI downtime: fallback to manual routing, assuming 20% capacity loss on 10,000 tickets × $5/ticket average handling cost increase = $10,000/hour AI outage cost.
  • At 99.9% SLA (Azure/Vertex): 8.76 hours allowed downtime/year. Expected cost: $87,600/year.
  • At 99.95% SLA (Bedrock): 4.38 hours allowed downtime/year. Expected cost: $43,800/year.
  • At no SLA (direct API): Actual OpenAI/Anthropic historical uptime is approximately 99.5–99.8%, implying 17–43 hours/year. Expected cost: $170,000–$430,000/year.

SLA-implied annual financial advantage of Bedrock over no-SLA direct API: $126,200–$386,200 for this scenario.

For enterprises building SLA guarantees into their own product contracts downstream, using a no-SLA AI provider creates contractual risk that cannot be hedged.

Incident Response and Observability

Provider Incident communication Status page RCA timeline
AWS Bedrock AWS Service Health Dashboard, real-time health.aws.amazon.com 72 hours
Azure OpenAI Azure Service Health + Microsoft 365 admin status.azure.com 72 hours
Vertex AI Google Cloud Status status.cloud.google.com 48 hours
OpenAI API status.openai.com status.openai.com Variable
Anthropic API status.anthropic.com status.anthropic.com Variable

Section 9 — Model Variety and Lock-In Risk

Bedrock’s multi-model portfolio is the primary structural differentiator vs. Azure OpenAI and Vertex AI, which are each anchored to a single model family.

Model Portfolio Comparison

Provider Available model families (Q2 2026)
AWS Bedrock Anthropic Claude (Haiku 3.5, Sonnet 4, Sonnet 4.6, Opus 4); Amazon Nova (Micro, Lite, Pro, Premier); Meta Llama 3.x; Mistral (7B, 8x7B, Large); AI21 Jurassic; Cohere Command; Stability AI image models
Azure OpenAI OpenAI GPT-4o, GPT-4o mini, o1, o3, o4-mini; limited Azure-specific fine-tuned variants; some Meta Llama via Azure AI Foundry
Google Vertex AI Gemini 1.5 Pro/Flash, Gemini 2.0, Gemini 2.5 Pro; PaLM 2 (legacy); Meta Llama via Model Garden; Mistral via Model Garden; third-party models via Model Garden
OpenAI API GPT-4o, GPT-4o mini, o1, o3, o4-mini; DALL-E; Whisper; fine-tuned variants
Anthropic API Claude Haiku 3.5, Sonnet 4, Sonnet 4.6, Opus 4

Pricing Power of Multi-Vendor Access

Bedrock’s multi-model portfolio creates structural pricing leverage that single-vendor channels cannot replicate:

  1. Model substitution flexibility: When Anthropic raises prices or changes terms, Bedrock customers can shift workloads to Amazon Nova Pro ($0.80/$3.20/M) or Llama 3 (open weights, priced below $1/M) without changing infrastructure. Azure OpenAI customers have no equivalent leverage — GPT-4o price changes are unavoidable.

  2. Task-appropriate routing: Bedrock enables a single API call routing layer (e.g., Amazon Bedrock Intelligent Prompt Routing, GA in Q1 2026) that automatically routes to the cheapest model meeting quality requirements per request. A pilot at a financial services firm reduced total model spend 38% by routing 60% of triage queries to Nova Lite while routing synthesis tasks to Claude Sonnet.

  3. Negotiation posture: Credibly threatening to shift volume from Claude to Nova/Llama gives AWS enterprise customers leverage in Anthropic price negotiations. Azure OpenAI customers have no equivalent threat to OpenAI.

Lock-In Risk Scoring

Provider Lock-in risk Primary risk factor
AWS Bedrock Medium AWS IAM, VPC integration creates operational switching cost, but model portability is high
Azure OpenAI High Tightly coupled to OpenAI; Microsoft EA commitments create switching friction
Vertex AI High GCP infrastructure coupling; Gemini model dependency
OpenAI API Medium-high No infrastructure coupling, but fine-tuned models and GPTs are non-portable
Anthropic API Low-medium No infrastructure coupling; Claude models available on Bedrock as exit

Section 10 — Azure AI Foundry vs AWS Bedrock: Feature Comparison

Azure AI Foundry (formerly Azure AI Studio, rebranded Q4 2025) is Microsoft’s unified AI platform — the direct competitive answer to AWS Bedrock. The comparison is increasingly head-to-head.

Feature Comparison

Feature AWS Bedrock Azure AI Foundry
Model catalog 50+ models (Anthropic, Amazon, Meta, Mistral, AI21, Cohere, Stability) 1,700+ models (OpenAI, Meta, Mistral, Phi, Hugging Face ONNX catalog)
Managed fine-tuning Yes (Claude, Titan) Yes (GPT-4o, Llama, Phi, others)
RAG / Knowledge bases Amazon Bedrock Knowledge Bases (S3, OpenSearch native) Azure AI Search integration
Prompt management Amazon Bedrock Prompt Management Azure AI Foundry prompt variants
Evaluation framework Amazon Bedrock Evaluations Azure AI Foundry eval hub
Agent framework Amazon Bedrock Agents + multi-agent collaboration Azure AI Foundry Agent Service (OpenAI Assistants API compatible)
Guardrails / safety Amazon Bedrock Guardrails Azure Content Safety
Observability CloudWatch, CloudTrail Azure Monitor, Application Insights
Model deployment Serverless (pay-per-token) or provisioned throughput Pay-as-you-go or PTU
CI/CD integration CodePipeline, SageMaker Pipelines Azure DevOps, GitHub Actions
Vector database Amazon OpenSearch Serverless, Aurora pgvector Azure AI Search, Cosmos DB

Pricing Structure Differences

AWS Bedrock: Charges per token consumed. No platform fee. No per-user fee. Infrastructure (VPC, CloudWatch) is billed separately at standard AWS rates.

Azure AI Foundry: Model charges per token (same rates as Azure OpenAI). Fine-tuning compute billed at standard Azure GPU rates. Azure AI Search billed separately ($25–250+/month depending on tier). Evaluation runs billed per evaluation ($0.004–0.02 per evaluation run depending on model used). No separate platform fee for the hub itself.

Which Wins for Different Enterprise Use Cases

Bedrock wins:

  • AWS-native infrastructure (EC2, Lambda, ECS, EKS) — IAM, VPC, CloudTrail integration is native
  • Claude model preference — only Bedrock offers Claude with enterprise SLA outside Azure Marketplace
  • Multi-vendor model strategy — Bedrock’s model portfolio depth exceeds Foundry
  • Regulated industries on AWS (financial services, healthcare) — established compliance posture

Azure AI Foundry wins:

  • Microsoft-centric enterprises (Office 365, Dynamics, Azure DevOps) — deep ecosystem integration
  • GPT-4o access with enterprise SLA — only Azure OpenAI offers contractual SLA for GPT-4o
  • Fine-tuning breadth — 1,700+ model catalog with managed fine-tuning across multiple families
  • GitHub Copilot integration — Foundry models feed directly into Copilot enterprise customization
  • European data sovereignty requirements — Azure has more EU data center pairs with residency guarantees

When neither wins clearly:

  • Hybrid cloud enterprises — both require cloud-specific integration investment
  • Enterprises with significant GCP footprint — Vertex AI should be evaluated alongside both

Section 11 — Price Trajectory 2024–2026: How Providers Have Moved

AI model pricing has fallen at an unprecedented rate — faster than any prior compute category. Understanding the trajectory matters for commitment decisions.

Claude Price History (Anthropic / Bedrock)

Model Launch date Input price ($/M) Output price ($/M) Change
Claude 3 Opus March 2024 $15.00 $75.00 Baseline
Claude 3 Sonnet March 2024 $3.00 $15.00 -80% vs Opus
Claude 3.5 Sonnet June 2024 $3.00 $15.00 Same as Sonnet
Claude 3.5 Haiku October 2024 $0.80 $4.00 New tier
Claude Sonnet 4 May 2025 $3.00 $15.00 Parity maintained
Claude Sonnet 4.6 Q1 2026 $3.00 $15.00 Parity maintained
Claude Haiku 3.5 Q2 2026 $0.80 $4.00 Stable

Pattern: Frontier capability has increased dramatically (Sonnet 4.6 outperforms Opus on most benchmarks) while Sonnet-tier price held flat at $3/$15. The effective price-per-capability unit has fallen ~5x from Opus 2024 to Sonnet 4.6 2026.

GPT-4 / GPT-4o Price History (OpenAI / Azure)

Model Launch date Input price ($/M) Output price ($/M) Change
GPT-4 (8K) March 2023 $30.00 $60.00 Baseline
GPT-4 Turbo November 2023 $10.00 $30.00 -67%/$-50%
GPT-4o May 2024 $5.00 $15.00 -50% vs Turbo
GPT-4o (price cut) July 2024 $2.50 $10.00 -50% again
GPT-4o (current) Q2 2026 $2.50 $10.00 Stable since July 2024
GPT-4o mini July 2024 $0.15 $0.60 New tier

Pattern: GPT-4 pricing dropped 92% input / 83% output over 3 years. Frontier pricing appears to be stabilizing — GPT-4o has held at $2.50/$10.00 for nearly two years.

Gemini Price History (Google / Vertex)

Model Launch date Input price ($/M) Output price ($/M) Change
Gemini 1.0 Pro December 2023 $0.50 $1.50 Baseline
Gemini 1.5 Pro (launch) April 2024 $7.00 $21.00 Higher for 1M context
Gemini 1.5 Pro (price cut) July 2024 $1.25 $5.00 -82%/-76%
Gemini 2.0 Flash December 2024 $0.10 $0.40 New efficient tier
Gemini 2.5 Pro Q1 2026 $1.25 $10.00 Higher output for reasoning

Implications for Commitment Pricing Decisions

The historical trajectory creates a commitment pricing dilemma:

  1. 12-month PTU/reserved commitments locked in at 2025 rates are likely overpaying. If frontier model prices drop another 30–50% (consistent with historical pattern), a 12-month commit at today’s rates could cost significantly more than pay-as-you-go next year.

  2. Counter-argument for commits: Predictability has value. CFOs building AI into operating budgets prefer fixed costs. Azure PTU provides guaranteed capacity in addition to lower price — congestion on shared endpoints is a real operational risk.

  3. Recommended strategy: Given price trajectory uncertainty, favor shorter commit windows (3–6 months) or batch API (no commit required, 50% discount) over 12+ month PTU/reserved commitments. The 10–15% additional discount for 12-month vs. 6-month Azure PTU does not compensate for the risk of locking in above next year’s list price.

  4. Exception: Compliance-tier deployments (HIPAA, PCI) where switching cost is high should accept longer commits — the compliance setup cost makes switching expensive regardless of price.

Price Trajectory Projection (Conservative Estimate, 2027)

Model tier Current price (Q2 2026) Projected 2027 (conservative -30%)
Claude Sonnet 4.6 input $3.00/M $2.10/M
Claude Sonnet 4.6 output $15.00/M $10.50/M
GPT-4o input $2.50/M $1.75/M
GPT-4o output $10.00/M $7.00/M
Gemini 2.5 Pro input $1.25/M $0.875/M

Section 12 — Enterprise Procurement: EDP vs Microsoft EA vs Google CUDs

The final layer of AI cost optimization is the enterprise discount vehicle negotiated at the cloud relationship level. These discounts are the largest single lever available to large enterprises and are almost entirely invisible in published pricing.

AWS EDP (Enterprise Discount Program)

Structure:

  • Multi-year spend commitment to AWS (typically 1–3 years, minimum $100K/year to participate, meaningful discounts at $1M+/year)
  • Blanket percentage discount applied to all AWS spend including Bedrock
  • Typical discount range: 8–15% on all services; heavy negotiators at $10M+/year have achieved 18–22%
  • EDP discounts apply to Bedrock token charges — confirmed in AWS EDP program documentation as of Q1 2026

Negotiation leverage points:

  • Migration commitment (moving workloads from on-prem or Azure/GCP)
  • Reserved Instance / Savings Plan purchase volume
  • MAP (Migration Acceleration Program) participation
  • Total revenue opportunity (headcount × projected AI usage)

Bedrock-specific EDP math: At $1M/year Bedrock spend with 12% EDP: $120,000/year savings, no commitment to specific models or usage patterns — maximum flexibility.

Microsoft Enterprise Agreement (EA) for Azure OpenAI

Structure:

  • Microsoft EA covers Azure services including Azure OpenAI
  • Typical discount range: 10–20% on Azure consumption for $1M+/year EA customers
  • Azure OpenAI is covered under standard Azure consumption — EA discounts apply automatically
  • Microsoft Copilot/M365 bundles sometimes include Azure OpenAI credits that further reduce effective cost

Negotiation leverage points:

  • Office 365 / M365 renewal timing (maximum leverage at renewal)
  • Competing Azure proposal from a GSI partner
  • Multi-year commit extension
  • Copilot for Microsoft 365 deployment commitment

Azure OpenAI PTU + EA stacking:

  • PTU provides 30–60% savings vs pay-as-you-go (volume-dependent)
  • EA provides additional 10–20% on top
  • Combined savings vs. list price: 37–68% for committed high-volume deployments
  • This makes Azure OpenAI + EA + PTU the most aggressively priced GPT-4o option available

Example (100 PTUs GPT-4o, 20% EA discount):

  • List PTU: $2.00/PTU/hr × 100 × 730 hrs = $146,000/month
  • With 20% EA: $116,800/month
  • vs. on-demand GPT-4o at equivalent volume: ~$200,000/month
  • Net savings: $83,200/month ($998,400/year)

Google Cloud CUDs (Committed Use Discounts) for Vertex AI

Structure:

  • CUDs are available for compute resources (GPU, CPU, memory) but NOT directly for Vertex AI Generative API token charges as of Q2 2026
  • Vertex AI generative model calls are not eligible for CUD discounts — they are pay-per-token with no commitment vehicle outside flat-rate pricing for specific use cases
  • CUDs apply to training compute (GPU instances) and custom model serving (if using Vertex AI endpoints with dedicated GPU)
  • Typical CUD discount: 20–55% on compute, 1 or 3-year term

Implication: Google Cloud’s generative AI API pricing lacks a mature enterprise commitment vehicle. Enterprises doing heavy Gemini inference through Vertex API are on pay-as-you-go with no volume discount. This is a structural disadvantage vs. AWS EDP (applies to Bedrock) and Microsoft EA (applies to Azure OpenAI).

Workaround: Enterprises with large GCP footprints can negotiate custom pricing with Google Cloud account teams at $1M+/year AI spend. Not self-serve; requires direct negotiation.

Comparison: Which Path Yields Best AI Cost Outcome

Enterprise profile Best procurement path Est. effective discount vs list
AWS-native, $500K+/year Bedrock AWS EDP negotiated at EA renewal 10–15% on all Bedrock
Microsoft-centric, GPT-4o primary Microsoft EA + Azure PTU 37–68% combined
Google Cloud-native, Gemini primary Direct negotiation with GCP account team 10–20% (negotiated, not published)
Multi-cloud, $2M+/year AI spend Negotiate all three; use Bedrock as primary with multi-model routing 12–15% on primary, leverage for negotiations
AWS-native but wants GPT-4o Bedrock EDP + Azure OpenAI pay-as-you-go EDP covers Bedrock; Azure at list

Negotiation Sequencing Recommendation

For a $1M+/year enterprise AI buyer, the optimal negotiation sequence is:

  1. Establish Bedrock baseline first. Lock in AWS EDP with Bedrock explicitly called out in scope. This takes 45–90 days and should precede model selection decisions.
  2. Use EDP as leverage with Microsoft. When negotiating Azure EA (or renewal), demonstrate that Bedrock covers Claude at EDP rates — Azure OpenAI must beat 15% EDP + Bedrock pricing to justify the switch for Claude workloads.
  3. Use Azure PTU math as leverage on OpenAI. If direct OpenAI API spend is material, present Azure PTU effective pricing to OpenAI Enterprise negotiations as the alternative.
  4. Do not negotiate Google CUDs for inference. Focus any GCP negotiation on training compute (if applicable). Vertex AI inference pricing is not currently addressable via standard CUD vehicles.

Total Cost Modeling: Same Workload, Three Procurement Paths

Workload: 2B input tokens/month, 500M output tokens/month, mixed Claude + GPT-4o (60/40 split), HIPAA-covered, AWS-native infrastructure.

Component AWS EDP + Bedrock Microsoft EA + Azure Hybrid (Bedrock primary)
Claude tokens (1.2B input, 300M output) $6,864/month (EDP 12%) N/A on Azure (use Azure marketplace) $6,864 (EDP)
GPT-4o tokens (800M input, 200M output) Claude substitute: $3,696 $2,800 (EA 20% + PAYG) $2,800 (Azure EA)
Egress costs $0 (in-VPC) $0 (in-VPC Azure) $0/$74 (Azure egress for 800M tokens payload est.)
Compliance overhead $0 (AWS BAA) $0 (Microsoft BAA) $0 (both BAAs in place)
Support/SLA premium $0 (Enterprise support) $0 (Microsoft Premier) $0
Monthly total $10,560 $2,800 $9,664
Annual total $126,720 $33,600 $115,968

Note: This comparison highlights that GPT-4o’s lower output token price ($10 vs $15 for Claude) combined with Microsoft EA discount makes Azure OpenAI materially cheaper for GPT-4o-equivalent tasks. The Bedrock advantage is compliance integration cost (already paid), not token price for GPT-4o-equivalent quality.

Bottom line: For Claude-specific workloads on AWS, Bedrock + EDP is optimal. For GPT-4o workloads, Azure OpenAI + EA + PTU wins on effective token price. Enterprises running both should split procurement accordingly rather than forcing all AI through a single vendor.


Appendix A — Quick Reference: Q2 2026 Token Prices

Model Channel Input $/M Output $/M Context Batch discount
Claude Haiku 3.5 Bedrock / Direct $0.80 $4.00 200K 50% batch
Claude Sonnet 4.6 Bedrock / Direct $3.00 $15.00 200K 50% batch
Claude Opus 4 Bedrock / Direct $15.00 $75.00 200K 50% batch
GPT-4o mini OpenAI / Azure $0.15 $0.60 128K 50% OpenAI batch
GPT-4o OpenAI / Azure $2.50 $10.00 128K 50% OpenAI batch
o4-mini OpenAI / Azure $1.10 $4.40 200K 50% OpenAI batch
Gemini 2.0 Flash Vertex / Direct $0.10 $0.40 1M None
Gemini 1.5 Pro (≤128K) Vertex / Direct $1.25 $5.00 2M None
Gemini 2.5 Pro (≤200K) Vertex / Direct $1.25 $10.00 1M None
Amazon Nova Lite Bedrock $0.06 $0.24 300K None
Amazon Nova Pro Bedrock $0.80 $3.20 300K None
Llama 3.3 70B Bedrock $0.72 $0.72 128K None
Mistral Large 2 Bedrock $2.00 $6.00 128K None

Appendix B — Decision Framework: Which Provider for Which Use Case

Use case Recommended provider Rationale
Enterprise RAG on AWS infrastructure Bedrock (Claude Sonnet 4.6) In-VPC, no egress, IAM native, HIPAA covered
High-volume async document processing OpenAI Batch API (GPT-4o) 50% discount, no commit, strong output quality
Long-context full-document processing (>200K) Vertex AI (Gemini 1.5 Pro) 2M context window, $1.25/M input
Real-time customer service routing (cheap) Bedrock (Nova Lite) $0.06/M input, 300K context, AWS-native
Reasoning / complex analysis Bedrock (Claude Opus 4) or Azure (o4-mini) Task-appropriate quality vs. cost
Microsoft 365 integrated workflows Azure OpenAI (GPT-4o) EA discount, Copilot integration, SLA
PCI-compliant AI (cardholder data adjacent) Bedrock, Azure, or Vertex (all PCI certified) Direct APIs disqualified
Startup / no compliance requirement OpenAI API direct or Anthropic direct No infrastructure overhead, instant start
Multi-model routing to minimize cost AWS Bedrock Only platform with native Intelligent Prompt Routing across models