Hyperscaler AI Pricing - Bedrock vs Foundry vs Vertex

What It Does

Compares the on-demand inference price of the same foundation model across the three managed AI platforms:

• AWS Bedrock (us-east-1)
• Azure AI Foundry (eastus, global deployment tier)
• Google Vertex AI (us-east5)

For each model in the catalogue, you see input + output prices on each platform, the cheapest option highlighted, and “N/A” for platforms where the model isn’t offered.

The cost calculator at the top lets you input your daily token volumes - the comparison rebuilds with monthly USD figures so you can see the actual dollar difference for your workload.

How to Use It

1. Enter your daily volumes: input tokens per day + output tokens per day. Defaults to 500K input / 100K output (a typical small production app).
2. Browse the model cards: each card shows the model’s price on Bedrock, Foundry, and Vertex. Cheapest platform gets a green ”✓ Cheapest” badge. Unavailable platforms show “N/A” with a one-line reason.
3. Filter by vendor (Anthropic, Meta, etc.) or by tier (cheapest first / most capable first).
4. Read the verification date at the bottom of each row - that’s when the price was last confirmed against the platform’s official pricing API.

How the Pricing Stays Current

Every Monday at 9 AM UTC, a GitHub Actions workflow:

1. Calls AWS Bedrock’s public Bulk Pricing JSON
2. Calls Azure’s Retail Prices API
3. Calls GCP’s Cloud Billing Catalog API
4. Diffs the responses against the canonical JSON in the repo
5. If anything changed, opens a pull request with the diff for human review

Pricing data has a last_verified field per row + a verified_via: api flag once auto-confirmed. Cost: $0 (all three platforms publish pricing for free; GitHub Actions on public repos has unlimited free minutes).

The pipeline source code lives in scripts/pricing/ in the gekro repo. The pricing JSON itself is at apps/web/src/content/data/hyperscaler-pricing.json - version-controlled, every change is a reviewable PR.

What’s In Scope

• ✅ On-demand inference pricing (input + output tokens)
• ✅ Standard region per platform (us-east-1 / eastus / us-east5)
• ✅ Standard pay-as-you-go tier
• ✅ ~12 cross-platform foundation models with active 2026-era pricing

What’s NOT In Scope (Intentionally)

• ❌ Provisioned throughput / committed use - Bedrock PT, Foundry PTU, Vertex PVM. These are negotiated, not generally publishable.
• ❌ Regional price deltas - same model can be 0-30% cheaper in different regions; tracking N regions × N models combinatorially explodes maintenance.
• ❌ Fine-tuning, embedding, image generation, storage, RAG primitives - different cost dimensions, separate decision.
• ❌ Every model on every platform - there are 73 models on Bedrock alone. The comparison tracks the top ~12 with cross-platform interest. Add requests via GitHub issue.

Why This Matters

The same model is sold at slightly different prices across the three hyperscalers, and the prices change frequently. A static comparison page goes stale within weeks. By the time you’ve spent an hour comparing manually, AWS has dropped Llama 4 by 20% or Azure has added a new GPT tier.

This tool exists because the comparison is genuinely valuable AND there’s no honest way to maintain it without automation. The combination of (a) an auto-fetcher pipeline, (b) a JSON-as-source-of-truth, (c) human-reviewed PR for every change makes the data both fresh AND accountable.

Where the Numbers Come From

Platform	Source	Auth required
AWS Bedrock	Public Bulk Pricing JSON	None
Azure Foundry	Retail Prices API	None
GCP Vertex	Cloud Billing Catalog API	API key (free)

Run discovery locally:

node scripts/pricing/update-pricing.mjs --discover

Outputs every model on every platform with its raw SKU. Useful when adding new models to the comparison.

Limitations

• Region locked to defaults: us-east-1 / eastus / us-east5. Cheaper regions exist for some models but tracking all combinations is unmaintainable.
• Standard tier only: Provisioned throughput, batch inference, and flex tiers all have different (often cheaper) pricing not covered here. Standard on-demand is the apples-to-apples comparison.
• Output may differ ±2% from the platform’s pricing page if the API hasn’t propagated yet (pricing API updates can lag the marketing page by a few hours).
• Catalogue selectivity: the comparison covers the top ~12 cross-platform models, not every model on every platform. By design - full catalogues are noisy, this surfaces the meaningful comparisons.

Related Tools

• LLM Cost Calculator - direct API pricing (Anthropic, OpenAI, Google direct, not via hyperscalers)
• Model Benchmark - capability comparison (not cost)
• GPU VRAM Calculator - local hosting alternative