The phrase "free tier" usually comes from two places: consumer chat products with free plans, and cloud providers' general free-credit programs. Neither is a free tier for Claude API usage. Across the Claude API and all four third-party platforms, model usage is pay-as-you-go per token from the first billable request. What differs per platform is the gate you pass through before that request can happen — and that gate is where surprise invoices are born or prevented.
Where the meter starts, platform by platform
Claude API (first-party). Billing is per token at list prices from your first request. What you do get is a built-in ceiling: organizations sit on usage tiers with monthly spend caps — Start $500, Build $1,000, Scale $200,000 — and once you hit your tier's cap, API usage pauses until the next month unless you request a higher limit. You can also set your own spend limit below the cap in the Console (Settings > Limits). For a new team, that Start-tier cap is the difference between a bad week and a bad quarter.
Amazon Bedrock. The subtle one. Access to foundation models is enabled by default given correct AWS Marketplace permissions, and the first invocation of a third-party model auto-initiates the Marketplace subscription (setup can take up to 15 minutes). There is no separate purchase step to catch — the first successful call is both the subscription and the first billable event, and first use implies agreeing to the model's EULA. On the legacy surface, Anthropic models additionally require a one-time First Time Use form per account or AWS Organization; that requirement does not apply on the bedrock-mantle endpoint. To keep experimentation from becoming spend, control it with IAM — denying Marketplace subscribe permissions alone does not block invocation; deny bedrock:InvokeModel itself.
Google Vertex AI. The gate is explicit: you need a Google Cloud project with billing enabled and the Vertex AI API turned on, then you enable each Claude model from its Model Garden card and accept the Terms of Service. Because billing enablement is a prerequisite, the first prompt after enablement is billable — pay-as-you-go per 1M tokens at the published rates.
Microsoft Foundry. The strictest gate. Claude requires a paid Azure subscription: free trial, student, and credit-only subscriptions are not supported, and Free Trial subscriptions get quota of zero across the board. You accept Azure Marketplace terms at deployment time; usage is then metered in Claude Consumption Units (CCUs) hourly and invoiced monthly in arrears — pay-as-you-go, with no prepaid CCU credits or balances.
Claude Platform on AWS. Same CCU model through AWS Marketplace ($0.01 per CCU; 100 CCU = $1.00 of fees at standard rates), metered hourly, postpaid arrears. One warning specific to this platform: Anthropic spend limits are not available here — budget guardrails live on the AWS billing side.
What actually changes at "production"
Technically, nothing — there is no separate production SKU, and list prices are the same across platforms. What changes is exposure: volume multiplies, and the arrears billing model on the marketplace platforms means a whole month of unnoticed usage can accumulate before the first invoice arrives. Three habits prevent the surprise:
| Habit | Why |
|---|---|
| Set the ceiling before the traffic | Console spend limits (1P), workspace limits per team, or cloud-side budget controls on the marketplace platforms |
| Watch usage weekly from day one | Usage/Cost API or Console on 1P; Foundry Monitoring tab; your own request logs everywhere |
| Price the pilot honestly | Multiply expected tokens by list prices — and note Sonnet 5's intro $2/$10 pricing ends August 31, 2026, so a pilot priced in August costs more in September |
Where to go next
Read free tier realities for the credits question in depth, pricing explained for the rate card, and the quickstart to make that first (billable) call deliberately.