From AWS Bedrock to ClaudeAPI.cheap.
Same Claude models. 70-80% off the bill.
Bedrock packages Claude with the full AWS billing apparatus — retail pricing plus region multipliers, IAM, VPC, CloudWatch charges, and the boto3 SDK with its model-ID-versioning quirks. We sit in front of the same Opus, Sonnet, and Haiku models at the Pro discount. The migration swaps boto3 for the standard Anthropic SDK — fewer lines of code, simpler errors.
Real pricing math
Per 1M tokens, USD. Bedrock prices track Anthropic list in most regions; some Asia-Pacific and EU regions add 5-15% on top via cross-region inference fees.
Doesn't include Bedrock's CloudWatch logging, VPC endpoint hourly, or KMS encryption surcharges — those compound the savings further. Pro tier is $19 lifetime, paid once.
Why teams leave AWS Bedrock for a proxy
AWS resells Claude at the same headline rates Anthropic publishes — sometimes with a regional uplift. Bedrock's pricing page reads ~identical to api.anthropic.com per 1M tokens. The AWS value-add (IAM, VPC, multi-region) costs what it costs; the Claude math itself isn't discounted.
Bedrock model IDs encode the AWS region inference profile (us.anthropic.claude-..., apac.anthropic.claude-...), the model version timestamp, and the suffix :0. Hard-coded model strings break when AWS rotates inference profiles. Standard Anthropic SDK IDs (claude-opus-4-7) are stable.
Bedrock catalog is broader than Claude only, but the Bedrock shape forces a per-vendor body schema. ClaudeAPI.cheap dispatches Opus/Sonnet/Haiku, GPT-5 family, and Gemini 3 Pro/Flash from one sk-cc- key — Anthropic-format request body for Claude, OpenAI-format request body for GPT and Gemini. One auth, one balance, one invoice.
AWS bills card-on-file end-of-month plus all the ancillary CloudWatch / KMS / data-transfer line items. We bill against a prepaid USDT/BTC/ETH balance with zero hidden line items — a runaway agent hits the balance ceiling, not your AWS root account credit limit. Pricing per model is posted publicly.
Bedrock typically lags Anthropic direct by 1-3 weeks per model release while AWS provisions regional inference and signs the carve-out. We track the Anthropic catalog directly — Opus 4.7 was available here within days of its release. Catalog at /models.
Migration in 3 steps
Get an API key (30 seconds)
Sign up at claudeapi.cheap/signup — email + crypto top-up, no card. Generate an sk-cc-... key from the dashboard.
Swap boto3 for the Anthropic SDK
Bedrock's invoke_model takes a JSON-stringified body wrapped in a specific anthropic_version envelope. The Anthropic Python SDK uses the cleaner messages.create shape. Side-by-side:
Same response shape — Anthropic SDK exposes .content[0].text directly. No JSON marshal/unmarshal, no region argument, no IAM role. Roll back by switching the import back.
Verify with one curl
Confirm the proxy answers from any environment — no AWS credentials, no signed-request boilerplate.
The first response logs to /dashboard/usage with exact token count and cost. After that, your existing Anthropic-SDK code is in production.
Model ID mapping
Bedrock IDs encode region inference profile + timestamp + version suffix. ClaudeAPI.cheap uses the canonical Anthropic IDs — same names you'd use against api.anthropic.com directly.
Cross-region profiles (apac., eu.) map to the same canonical ID — there's no per-region duplication on our side. Full catalog including GPT-5 and Gemini 3 at /models.
FAQ
If Bedrock and Anthropic Direct cost the same, why is one of them "more expensive"?
Headline per-token rate is roughly identical. Bedrock's "tax" is the surrounding AWS bill: CloudWatch log ingestion (~$0.50/GB), VPC endpoint hourly (~$0.01/h per endpoint per AZ), KMS key usage if you encrypt request bodies, and cross-region inference fees if you route from a region that doesn't host the model directly. For high- throughput agent workloads these accumulate to 10-25% on top of token cost. ClaudeAPI.cheap has zero ancillary line items.
Does prompt caching work after migration?
Yes. The cache_control blocks you used in the Bedrock request body work identically through the Anthropic SDK shape — write at 1.25× input, read at 0.1× input. Cache metadata in the response is intermittent (Anthropic-side, not our choice) — plan for 80-90% passthrough.
What about IAM, VPC endpoints, and AWS PrivateLink?
Not applicable — ClaudeAPI.cheap is a public HTTPS endpoint authenticated by a bearer token. For tenants that need IP-locked egress, the proxy is reachable from any HTTPS-capable network; you can whitelist claudeapi.cheap in your egress firewall. For VPC-internal traffic only, you would need to keep Bedrock or self-host LiteLLM. Most teams who migrate accept the trade for the discount + simpler ops.
Does the Bedrock Converse API map to anything?
Converse was AWS's vendor-neutral wrapper to call Claude, Llama, Mistral, etc. with one shape. ClaudeAPI.cheap solves the same multi-vendor problem at a different layer: onesk-cc- key, Anthropic-format for Claude, OpenAI-format for GPT and Gemini. Migration from Converse is straightforward — pick the appropriate request shape per vendor instead of the unified Converse envelope. The savings are large enough that most teams accept the change.
Data residency — HIPAA, EU GDPR, SOC 2?
We do not log request or response content. We do not have BAA (HIPAA) or SOC 2 attestation today — if those are required by your compliance program, Bedrock's AWS-native attestations may still be the right choice. See /trust and /privacy for the exact policies. We'll be straight about what we do and don't cover.
What if I'm not satisfied after migrating?
Email support@claudeapi.cheap within 7 days for a full refund of the remaining balance and the $19 Pro upgrade. Crypto refunds return to the originating wallet within 24 hours. Reverting to Bedrock is git revert on the SDK swap commit.
No card, crypto only. Basic free forever; $19 lifetime Pro. Read our SLA →