AWS Bedrock vs OpenRouter vs Direct API: LLM Pricing Compared
AWS Bedrock vs OpenRouter vs a direct API key is mostly a question of fees and data location, not token prices. On November 16, 2026, Claude Sonnet 5.5 costs $2 per million input tokens and $10 per million output tokens on all three routes. What changes the bill is a 10% premium for regional endpoints on Bedrock (and for US-only inference on the Anthropic API), OpenRouter's 5.5% fee on credit purchases, and whether you use caching and batch at all.
What does AWS Bedrock vs OpenRouter vs direct API pricing come down to?
All three routes charge the vendor's list price per token for the current Claude and GPT models. The differences are in the extras: Bedrock adds 10% for regional (in-geography) endpoints, OpenRouter adds a 5.5% fee when you buy credits (8% on its Business plan), and the direct APIs add 10% only when you pin data residency.
| Route | Token price vs list | Extra fees | Data residency option |
|---|---|---|---|
| Anthropic API / OpenAI API | List price | None | +10% (US-only inference or regional processing) |
| AWS Bedrock, global endpoint | List price | None | No |
| AWS Bedrock, regional endpoint | +10% | None | Yes, built in |
| OpenRouter Standard | List price | 5.5% on credit purchases ($0.80 minimum) | Business plan only (8% fee) |
All prices below are per million tokens (MTok), USD, checked on November 16, 2026. Sources: Anthropic pricing docs, OpenAI pricing docs, the AWS Price List feed behind the Amazon Bedrock pricing page (us-east-1, published November 9, 2026), and OpenRouter's pricing page and public models API.
Claude pricing on Anthropic API vs Bedrock vs OpenRouter
Anthropic's docs state that, from the Claude 4.5 generation on, Bedrock's "regional and multi-region endpoints include a 10 percent premium over global endpoints." The AWS price list confirms it line by line.
| Model | Anthropic API (input / output) | Bedrock global | Bedrock regional | OpenRouter |
|---|---|---|---|---|
| Claude Fable 5.2 | $8 / $40 | $8 / $40 | $8.80 / $44 | $8 / $40 |
| Claude Opus 5.5 | $4 / $20 | $4 / $20 | $4.40 / $22 | $4 / $20 |
| Claude Sonnet 5.5 | $2 / $10 | $2 / $10 | $2.20 / $11 | $2 / $10 |
| Claude Haiku 5.5 (prompts up to 100K tokens) | $0.10 / $0.50 | $0.10 / $0.50 | $0.11 / $0.55 | $0.10 / $0.50 |
Caching follows the same pattern. Sonnet 5.5 cache hits cost $0.10, 5-minute writes $2.50 and 1-hour writes $4 on the Anthropic API and Bedrock global; Bedrock regional adds 10%. OpenRouter lists the same $0.10 read and $2.50 write.
- Haiku 5.5 has a long-prompt tier. Above 100,000 prompt tokens, it jumps to $0.50 input and $2.50 output on both the Anthropic API and Bedrock global. Cache reads and writes count toward that threshold.
-
The direct API has its own residency premium. Setting
inference_geo: "us"on Claude 4.6 and later applies a 1.1x multiplier to every token category. So Bedrock regional only costs more than the direct API's global default, not more than a US-pinned direct setup.
For the direct numbers in depth, see Claude API cost.
GPT pricing on OpenAI API vs Bedrock vs OpenRouter
OpenAI models run on Bedrock too, with the same structure: global at list price, regional at 10% more. OpenAI's own API charges a matching 10% uplift for regional processing on models released on or after March 5, 2026.
| Model | OpenAI API (input / cached / output) | Bedrock global | Bedrock regional | OpenRouter |
|---|---|---|---|---|
| GPT-6 Astra | $10 / $1 / $50 | $10 / $1 / $50 | $11 / $1.10 / $55 | $10 / $1 / $50 |
| GPT-6.1 Sol | $2 / $0.10 / $10 | $2 / $0.10 / $10 | $2.20 / $0.11 / $11 | $2 / $0.10 / $10 |
| GPT-6.1 Terra | $1 / $0.05 / $5 | $1 / $0.05 / $5 | $1.10 / $0.055 / $5.50 | $1 / $0.05 / $5 |
| GPT-6 Luna | $0.10 / $0.01 / $0.50 | $0.10 / $0.01 / $0.50 | $0.11 / $0.011 / $0.55 | $0.10 / $0.01 / $0.50 |
One line differs. OpenAI's pricing page lists no cache-write rate, but Bedrock's price list carries a 30-minute cache-write line for GPT models at 1.25x input ($2.50 for GPT-6.1 Sol, global), and OpenRouter's listing shows the same figure. If you rewrite large cached prefixes often, test real traffic on each route first. These are short-context rates (up to 272K input tokens on the OpenAI API).
The fees on top: OpenRouter, Bedrock and residency premiums
Token prices tie, so the fees decide the ranking.
OpenRouter. Its FAQ says it passes "through the pricing of the underlying model providers without any markup." You pay instead when you buy credits:
- Standard plan: 5.5% fee, $0.80 minimum per purchase.
- Business plan: 8% fee, adds EU and US in-region routing.
- Crypto payments: 5% fee.
- Bring your own key (BYOK): the first $25,000 of list-price inference per month carries no fee, then 5% of what the same request would cost on OpenRouter.
Bedrock. No platform fee on on-demand tokens; the 10% premium applies only to regional endpoints. Its Reserved tier sells capacity by tokens-per-minute (Claude Sonnet 4.6: $0.18 per 1K input TPM per hour, 1-month, global), which pays off only at steady high volume.
Direct APIs. No fees. You pay 10% more for US-only Claude inference or OpenAI regional processing, and OpenAI's Fast mode (renamed from Priority on July 30, 2026) costs 2x Standard.
On a $900 monthly Sonnet 5.5 workload, the cheapest and most expensive routes differ by $90.
Caching and batch discounts matter more than the route
Prompt caching and batch processing cut the bill by 50% or more. The route choice moves it by 10% at most.
Take the same workload and assume 80% of input tokens come from cache:
- Fresh input: 60M x $2 = $120
- Cache reads: 240M x $0.10 = $24
- Output: 30M x $10 = $300
- Total: $444 a month, before cache writes, versus $900 uncached
Batch works the same way across vendors:
| Discount | Anthropic API | OpenAI API | Bedrock | OpenRouter |
|---|---|---|---|---|
| Batch | 50% off input and output | 50% off (Batch and Flex share rates) | 50% off for select models |
:batch model variants listed at 50% |
| Stacks with caching | Yes | Cached input has its own batch rate | Per model | Per listing |
Bedrock's price list shows 50% batch lines for models such as Claude Sonnet 4.6 ($1.50 / $7.50 global) but none yet for the 5.5 generation, so check your model in the console. For the math, read Claude Opus 5.5 prompt caching cost math and the Claude Batch API guide.
Rate limits on each route
-
OpenRouter paid models have "no platform-level request cap," per its docs; limits pass through from the provider. Free
:freemodels cap at 20 requests per minute and 50 per day (1,000 after $10 in credits). - Anthropic and OpenAI direct use usage tiers that grow with spend. Our GPT API rate limits guide covers OpenAI's tiers.
- Bedrock uses per-account, per-region quotas; global endpoints route across regions, and Reserved tier buys fixed tokens-per-minute.
Which route should you pick?
Match the route to the constraint you actually have:
- Pick the direct API if you use one vendor and don't need AWS billing. List price, no fees.
- Pick Bedrock global if your company already pays AWS and wants one invoice and IAM auth. Same token price as direct.
- Pick Bedrock regional if data must stay in a geography. You pay 10% more, the same premium the direct APIs charge for residency.
- Pick OpenRouter if you switch between Claude, GPT and open models often, want one key and one bill for all of them, and accept 5.5% on top. Use BYOK to drop the fee on your first $25,000 a month. Our guides to using Claude Code with OpenRouter and OpenRouter alternatives go further.
Let your coding agent compare routes with Apidog CLI
Price tables don't show whether two routes return the same token counts and latency for your prompts. Let your coding agent run the same request on each route with Apidog CLI.
npm install -g apidog-cli@latest
apidog login --with-token <YOUR_TOKEN>
npx -y skills add https://apidog.com # pick your agent when prompted
Then give your agent one prompt:
In Apidog project <id>, create one environment per LLM route: Anthropic API
and OpenRouter. Create a test case that sends the same Claude Sonnet 5.5
prompt, asserts a 200 status and a non-empty usage object, and extracts
input and output token counts. Run it against each environment and give me
a table of tokens, latency and cost at $2 / $10 per million tokens.
The agent reads the schema, validates its payload, writes the case and runs it per environment:
apidog environment create anthropic-direct --project <id> --base-url https://api.anthropic.com
apidog environment create openrouter --project <id> --base-url https://openrouter.ai/api
apidog cli-schema get test-case-create
apidog cli-schema validate test-case-create --file case.json
apidog test-case create --project <id> --file case.json
apidog run --test-case <caseId> --project <id> -e <envId> -r cli,json
Add Bedrock the same way once your AWS credentials are set up. The same apidog run ... -r junit call drops into CI, so a response change on any route fails the check. The Apidog CLI in Claude Code guide walks through the agent setup.
FAQ
Is Claude cheaper on AWS Bedrock than on the Anthropic API? No. On Bedrock's global endpoint, Claude costs the same as the Anthropic API: Sonnet 5.5 is $2 input and $10 output per million tokens on both. Bedrock's regional endpoints cost 10% more.
Does OpenRouter mark up Claude and GPT prices? Not per token. OpenRouter passes through list prices and charges a 5.5% fee when you buy credits on its Standard plan, or 8% on Business. With BYOK, the first $25,000 a month of list-price usage has no fee.
What is the 10% premium on Bedrock? It's the price of regional and multi-region endpoints, which keep inference inside a geography. Global cross-region endpoints carry no premium. The direct Claude API charges a matching 1.1x for US-only inference.
Can I use prompt caching and batch on all three routes? Caching, yes. Batch is 50% off on the Anthropic and OpenAI APIs and appears as :batch variants on OpenRouter; on Bedrock, check per model.
Which route is cheapest for GPT models? The OpenAI API and Bedrock's global endpoint tie at list price, for example $2 input and $10 output for GPT-6.1 Sol. OpenRouter adds its credit fee, and regional endpoints on either OpenAI or Bedrock add 10%.


Top comments (0)