Bedrock or the Anthropic API?

Almost everyone arrives at this question expecting a price difference. There isn't one: for the 6 current Claude models available on both, the published per-token rates are identical, and the cache and batch multipliers match as well.

So the decision is not about cost per token. It is about what each platform can do — where 9 capabilities below are first-party only — and about who governs the deployment.

The price check

Bedrock figures come from this site's snapshot, in us-east-1 — the cheapest region for essentially every model, so the comparison never flatters it. Anthropic figures are hand-verified from the published pricing page.

Model	Bedrock in / out	Anthropic in / out	Verdict
Claude Fable 5 claude-fable-5	$10.00 / $50.00	$10.00 / $50.00	identical
Claude Opus 5 claude-opus-5	$5.00 / $25.00	$5.00 / $25.00	identical
Claude Opus 4.8 claude-opus-4-8	$5.00 / $25.00	$5.00 / $25.00	identical
Claude Sonnet 5 claude-sonnet-5	$3.00 / $15.00	$3.00 / $15.00	identical
Claude Sonnet 4.6 claude-sonnet-4-6	$3.00 / $15.00	$3.00 / $15.00	identical
Claude Haiku 4.5 claude-haiku-4-5	$1.00 / $5.00	$1.00 / $5.00	identical

curated Anthropic figures verified Aug 5, 2026 against the published pricing. Bedrock figures rebuilt from the AWS APIs, snapshot Aug 27, 2026. Promotional $2 / $10 through 2026-08-31 on the first-party API; the headline figure is the standard rate.

The multipliers are the same too

These are properties of the product rather than of the reseller, so they apply on both platforms — and they move a bill far more than any platform choice does.

Cache read	0.1×	Roughly a tenth of the input rate. The single largest lever on an agentic bill.
Cache write, 5-minute TTL	1.25×	Two reads break even against paying full price twice.
Batch	0.5×	Half rate for work that tolerates asynchronous completion.

Which is the practical point: an agentic workload with caching switched on costs dramatically less than the same workload without it, on either platform. Choosing between Bedrock and the first-party API for cost reasons is optimising the wrong number — the calculator shows the one that matters.

Inference surface

Capability	Where	What it means
Messages API, streaming, tool use	Both	The core surface is the same on both, which is what makes a migration mostly a client swap.
Adaptive thinking and effort levels	Both	Available on Bedrock, including the effort parameter.
1M-token context, PDF input, vision, citations	Both	All present on Bedrock.
Prompt caching (5-minute and 1-hour)	Both	Explicit cache breakpoints work on both platforms at the same multipliers.
Automatic prompt caching	Anthropic API only	The convenience form that caches the last cacheable block for you is first-party only. On Bedrock you place cache breakpoints explicitly — the same saving, more code.
Mid-conversation system messages	Anthropic API only	Injecting an operator instruction mid-session without invalidating the cached prefix is not available on Bedrock. The workaround costs you the cache on every injection.
Model IDs	Differs	Bedrock prefixes every Anthropic model with `anthropic.` — anthropic.claude-opus-5 against claude-opus-5. Passing a first-party ID to Bedrock is a 400, and it is the single most common migration error.

Server-side tools

Capability	Where	What it means
Web search and web fetch	Anthropic API only	The Anthropic-hosted search tools are not offered on Bedrock. Anything needing live web data has to be a tool you build and run yourself.
Code execution	Anthropic API only	The Anthropic-hosted sandbox is not on Bedrock. The AWS-native answer is a container or Lambda you operate.
Bash, text editor and memory tools	Both	These execute on your side on either platform, so they carry over unchanged.
Tool search	Both	Available on Bedrock through the InvokeModel API, though not through Converse.
Programmatic tool calling	Anthropic API only	Composing several tool calls inside one script, so intermediate results never enter the context window, is first-party only. On Bedrock each call is a round trip.
MCP connector	Anthropic API only	Calling a remote MCP server from the model request is not available on Bedrock. You run the MCP client yourself.

Agent platform

Capability	Where	What it means
Managed agent runtime	Differs	Anthropic runs Managed Agents; AWS runs Bedrock AgentCore. Neither is available on the other, and they are different products with different pricing models — not two names for one thing.
Agent Skills	Anthropic API only	The packaged document skills (pptx, xlsx, docx, pdf) are not offered on Bedrock.

Operations and governance

Capability	Where	What it means
Batch inference	Differs	Both offer batch at half rate, through different endpoints: Anthropic has the Message Batches API, AWS has Bedrock batch inference. The saving is the same; the integration is not portable.
Region and data residency control	Differs	This is the reason most Bedrock deployments exist. You pick the AWS region, and it is governed by IAM, VPC endpoints and your existing AWS agreements — where the first-party API gives you an inference-geo hint rather than a region.
Authentication and billing	Differs	Bedrock is IAM and your AWS bill, with no separate vendor contract. The first-party API is an Anthropic API key and an Anthropic invoice. For a large enterprise this is frequently the whole decision.
Models API	Anthropic API only	Programmatic capability discovery — context window, feature support per model — is first-party only. Bedrock has ListFoundationModels, which reports less. That gap is why this site exists.
Fast mode and task budgets	Anthropic API only	Higher-throughput output and agent-loop token budgets are first-party only.
Feature lead time	Differs	New capabilities land on the first-party API first and reach Bedrock later, if at all — the rows above are the accumulated result. Budget for the lag if your roadmap depends on a feature the week it ships.

How to actually choose

Choose Bedrock when governance is the constraint

You need the workload in a named AWS region, access governed by IAM and VPC endpoints, and spend on an AWS invoice under agreements that already exist. That is the entire value proposition, and for a regulated deployment it usually settles the question before any feature comparison starts. Check where the model actually runs before treating the region as settled, though — availability is not residency.

Choose the first-party API when the feature surface is the constraint

You want hosted web search, the code-execution sandbox, the MCP connector or Agent Skills, and you would rather not rebuild each of them. New capabilities also land there first — if your roadmap depends on a feature the week it ships, Bedrock will be behind.

The migration itself is small — until it isn't

The core request shape is identical, so moving is mostly a client swap plus the anthropic. model-ID prefix. What is not portable is everything built on a first-party-only capability: each of those is a rewrite, not a configuration change. Audit against the tables above before estimating the work.

Questions

Is Amazon Bedrock more expensive than the Anthropic API?

No. For the 6 current Claude models available on both, the published per-token list prices are the same figure on each platform — and the cache-read, cache-write and batch multipliers match too. Price is not the reason to choose one over the other; the feature surface and the governance model are.

What can the Anthropic API do that Bedrock cannot?

9 capabilities in the comparison on this page are first-party only, including the hosted web search, web fetch and code execution tools, the MCP connector, programmatic tool calling, Agent Skills, automatic prompt caching and the Models API. Bedrock has AWS-native equivalents for some of them, but not as the same API.

Why would I use Bedrock at all then?

Region choice and data residency, IAM instead of a separate API key, VPC endpoints, and consolidated AWS billing under agreements you already have. For a regulated or enterprise deployment that list usually outweighs the missing features, which is exactly why Bedrock exists.

Are the model IDs the same?

No, and this is the most common migration error. Bedrock prefixes every Anthropic model with anthropic. — anthropic.claude-opus-5 against claude-opus-5 on the first-party API. Passing a first-party ID to Bedrock returns a 400.

Does prompt caching work the same on both?

The explicit form does, at the same multipliers: cache reads at roughly a tenth of the input rate, writes at 1.25× for the five-minute TTL. What Bedrock does not have is automatic caching, the convenience form that places the breakpoint for you — there you place breakpoints explicitly and get the same saving for more code.

Bedrock availability and pricing across 31 regions is rebuilt daily from the AWS APIs. The first-party column is hand-maintained and dated above — this site has no way to query Anthropic's pricing programmatically, so it is stated as curated rather than implied to be live. Methodology.