OpenRouter vs Amazon Bedrock vs Google Vertex AI (2026): Router, Cloud, or Gateway?
OpenRouter is a hosted router, Bedrock and Vertex AI are cloud model platforms, and LiteLLM is a gateway you run yourself. Here is what each costs on the same Claude workload, and which one your cloud contract probably picks for you.
The short answer
Use OpenRouter for one API key across hundreds of models when you have no cloud spend commitment; it passes provider prices through and charges 5.5% on credit purchases (8% on Business). Use Amazon Bedrock or Google Vertex AI if you have an AWS or Google commitment or need a HIPAA-eligible service. Use LiteLLM to self-host routing. Claude Sonnet 5.5 lists at $2/$10 per million tokens on all three global endpoints, before OpenRouter's fee; pinning a region adds 10%.
Three different products that answer the same search
"Amazon Bedrock vs OpenRouter" compares products at different layers:
- Hosted router (OpenRouter). One API and one bill. Requests go out to many upstream providers, Anthropic, Google, Amazon Bedrock and Azure among them. You buy OpenRouter credits, and OpenRouter pays the provider.
- Cloud model platform (Amazon Bedrock, Google Vertex AI, Azure OpenAI Service). The cloud vendor hosts the models, bills them on your cloud invoice, and wraps them in its IAM, networking and compliance programs. Google's docs now call Vertex AI "Gemini Enterprise Agent Platform"; this post uses both names.
- Self-hosted or open-source gateway (LiteLLM, Portkey). Software that sits between your apps and the providers you already pay. It adds routing, fallbacks, budgets and logging, but no model access of its own.
Prices and plan names are as published on each vendor's own pricing page; last verified 10 November 2026.
| Vendor | Billing unit | Published entry price | What it's for |
|---|---|---|---|
| OpenRouter | Provider's per-token price, passed through, plus a fee when you buy credits | Free plan (free models only, 50 requests/day); pay-as-you-go fee of 5.5% on credit purchases (8% on Business) | One key for hundreds of models, provider fallback, fast model switching |
| Amazon Bedrock | Per 1M tokens (Standard, Flex, Priority, batch); Reserved tier per hour per 1K tokens-per-minute (1- or 3-month commitment) | Claude Sonnet 5.5: $2 input / $10 output (global cross-region); $2.20 / $11 geo or in-region | AWS-native apps, AWS spend commitments, IAM/VPC controls |
| Google Vertex AI (Agent Platform) | Per 1M tokens (standard, batch), Provisioned Throughput per GSU | Claude Sonnet 5.5: $2 / $10 global, $2.20 / $11 multi-region (US or EU), $1 / $5 batch | Google Cloud apps, Gemini first-party, Google commitments |
| Azure OpenAI Service | Per 1M tokens (Global, Data Zone, Regional), Provisioned (PTU), Batch | Per-token figures did not render on the public page on 10 Nov 2026; batch is "a 50% discount on Global Standard Pricing" | Azure-native GPT deployments |
| LiteLLM | Open-source software (MIT outside its enterprise/ folder); Enterprise is custom-priced | $0 for the software; you pay providers plus your own hosting | Self-hosted gateway in your own VPC |
| Portkey | Recorded logs per month | Developer: free, 10k logs/month; Production: $49/month for 100k logs | Hosted or self-hosted gateway with observability |
How the billing actually works
OpenRouter. OpenRouter says it passes "through the pricing of the underlying model providers without any markup", and that its fee "is charged when you buy credits, 5.5% on Standard and 8% on Business, never on individual requests". According to the FAQ, the only billing difference between Standard and Business is that fee. Business does add EU and US in-region routing, more workspaces, and a signed DPA on request. If you bring your own provider keys (BYOK), the first $25,000 of list-price inference each month carries no OpenRouter fee. After that, 5% is deducted from your OpenRouter credits. OpenRouter's model list also shows :batch variants. Claude Sonnet 5.5 batch, for example, is listed at $1 / $5 per million tokens.
Amazon Bedrock. On-demand pricing is per million tokens. Bedrock offers Standard, Flex, Priority and Reserved tiers. AWS states that Priority is "at 75% premium to Standard tier pricing", and that Flex and batch are each "at 50% discount" to Standard. Two details matter: geo and in-region endpoints cost 10% more than the global cross-region endpoint ($2.20 / $11 vs $2 / $10 for Sonnet 5.5), and the pricing table shows N/A for batch on Claude Sonnet 5.5, so the 50% batch discount does not apply to that model. AWS publishes no Provisioned Throughput rates for Anthropic models ("please reach out to your account team").
Google Vertex AI. Claude is billed per million tokens on Google's partner-model pricing page. Sonnet 5.5 costs $2 / $10 on the global endpoint, $2.20 / $11 on US or EU multi-region endpoints, and $1 / $5 through batch. Gemini is priced on the main Agent Platform page. Batch for Gemini models is at a 50% discount. Provisioned Throughput is sold in GSUs (generative AI scale units). Google's own worked example puts a 1-year commitment at $2,000 per GSU per month. Gemini 4 Argon's promotional $2 / $10 (global, via 50% credits back) runs through 30 November 2026; from 1 December it is $4 / $20.
Anthropic direct, for reference: $2 / $10 for Sonnet 5.5, 50% off via its Batch API. Its docs confirm that regional and multi-region cloud endpoints "include a 10 percent premium over global endpoints".
Which models each one can actually serve
- Claude: OpenRouter, Bedrock and Vertex. OpenRouter's Claude Sonnet 5.5 page lists Anthropic, Google Vertex, Amazon Bedrock, Azure and Claude Platform on AWS as providers. Azure OpenAI Service's own pricing page lists no Claude models.
- Gemini: Google (Vertex/Agent Platform) and OpenRouter. Under "Google", Bedrock's pricing page lists only open-weight Gemma models, not Gemini.
- OpenAI GPT: Bedrock lists GPT-6 Astra, Sol and Luna, GPT-6.1 Sol and Terra, and earlier GPT-5.x models. AWS announced GPT-6 Astra as generally available on 8 September 2026. Azure OpenAI Service is the long-standing GPT home. Google's OpenAI model page listed only gpt-oss 120B on 10 November 2026, though its data-governance page already covers OpenAI's flagship families; check your project's Model Garden. OpenRouter lists the GPT-6 and GPT-6.1 families (Astra, Sol, Terra, Luna).
- Open-weight models: all carry a long tail; OpenRouter's model API listed 465 models on 10 November 2026.
Data handling, retention and compliance
OpenRouter logs request metadata (timestamps, model, token counts). Its FAQ says "Prompt and completion are not logged by default". You can opt in to prompt logging for a 1% discount. "Zero Data Retention routing is available on every plan", and the pricing page says "We do not train on your data." Contractual SLAs are Enterprise-only, and the pricing page does not mention HIPAA or a BAA.
Amazon Bedrock runs third-party models in AWS-operated deployment accounts. AWS says model providers "don't have access to Amazon Bedrock logs or to customer prompts and completions". Retention is now an explicit per-Region setting. data_retention_mode: none gives zero retention. The exception is models whose provider requires human review, currently Claude Fable 5, 5.1 and 5.2. These need the aws_review mode, which keeps prompts and completions "within the AWS boundary for up to 30 days". ZDR for those models goes through an eligibility program. Amazon Bedrock is on AWS's HIPAA-eligible services list (updated 15 October 2026).
Google Vertex AI "won't use your data to train or fine-tune any AI/ML models without your prior permission". For zero retention, note that Gemini implicit caching is on by default (up to 24 hours) and can be disabled per project. OpenAI flagship models on the platform log prompts and responses for up to 30 days by default, and ZDR for them requires approval. Google's HIPAA page lists "Generative AI on Agent Platform" as a covered service. It says nothing either way about partner models such as Claude, so confirm that with your Google account team.
Get one new AI workflow every week
Practical playbooks, prompts, and tool stacks — for the SMB operators and enterprise teams actually shipping AI.
Why spend commitments usually decide it for enterprises
Bedrock usage lands on the AWS bill. Vertex usage lands on the Google Cloud bill; Google says its spend-based Flexible Savings Plans apply to "eligible products", so confirm whether partner models such as Claude qualify before counting them against a commitment. OpenRouter's pricing page names no cloud-marketplace purchase route, so treat its credits as a separate invoice that does not draw down an AWS or Google commitment unless OpenRouter and your account team confirm otherwise. If finance has already committed a year of spend to AWS, Bedrock tokens draw down spend the company has already committed, and a router has to win on something other than price.
The common compromise: OpenRouter or LiteLLM in front, with BYOK keys pointing at Bedrock or Vertex, so inference still bills to your cloud account (OpenRouter adds its 5% BYOK fee only above $25,000 of list-price inference a month).
Worked example: one Claude workload, routed three ways
Assumptions: Claude Sonnet 5.5, 100 million input and 20 million output tokens a month, interactive (no batch), no prompt caching, list prices as of 10 November 2026.
- Base token cost at $2 / $10: 100 × $2 = $200 input, plus 20 × $10 = $200 output, so $400/month.
- OpenRouter, Standard plan: $400 of pass-through inference, plus a 5.5% credit-purchase fee ($22), comes to about $422/month, assuming the fee sits on top of the credit amount. On Business with EU in-region routing, the providers charge $2.20 / $11 ($440), plus an 8% fee ($35.20), for $475.20.
- OpenRouter with BYOK: $400 of list-price inference is far below the $25,000 monthly allowance, so OpenRouter charges $0. Your provider bills you $400.
- Amazon Bedrock: $400 on the global cross-region endpoint, or $440 if you pin inference to a geography. Batch is listed as N/A for Sonnet 5.5.
- Google Vertex AI: $400 global or $440 multi-region (US or EU). If the job can wait for batch, $1 / $5 drops it to $200.
- LiteLLM in your VPC: the upstream price ($400 or $440) plus your own hosting cost, which LiteLLM does not publish.
Scale it 150× to $60,000 a month of list-price inference and OpenRouter's 5.5% becomes $3,300/month. With BYOK, the fee on the $35,000 above the allowance is 5%, or $1,750/month. Whether that justifies running your own gateway depends on what the proxy costs you to host and operate.
When not to use each one
When not to use OpenRouter: your spend is committed to AWS or Google; you need a signed BAA (its pricing page mentions none); or you need a contractual SLA without signing an Enterprise agreement. The Free plan is limited to free models at 50 requests a day (20 a minute).
When not to use Amazon Bedrock: you need Gemini, since Bedrock lists only Gemma; you were counting on half-price batch for Claude Sonnet 5.5, which the pricing table shows as N/A; or your policy forbids any retention but you want Claude Fable models, which require aws_review unless you qualify for ZDR.
When not to use Google Vertex AI: you need OpenAI flagship models in production today and they are not yet in your project; or you are budgeting Gemini 4 Argon on the November promotional rate, which doubles on 1 December 2026.
When not to use LiteLLM: nobody on the team will own a proxy in production. Enterprise features (SSO, audit logs, SLAs) are "Custom pricing" behind a sales form.
When not to use Portkey unchecked: its pricing page now reads "Portkey is now PRISMA AIRS AI Gateway"; confirm the contracting entity and roadmap.
Frequently asked questions
Is OpenRouter cheaper than calling Anthropic or OpenAI directly?
No. OpenRouter passes the provider's token price through without markup, then charges 5.5% when you buy credits (8% on Business). With BYOK you pay the provider directly and OpenRouter charges nothing on the first $25,000 of list-price inference a month, so cost matches going direct; above that it deducts 5% from your OpenRouter credits.
What is the difference between OpenRouter and Amazon Bedrock?
OpenRouter is a hosted router that sends requests to many providers, Bedrock among them, and bills through prepaid credits. Bedrock is AWS's own model platform, billed on your AWS invoice, with IAM, Region controls and HIPAA eligibility.
Does OpenRouter offer zero data retention?
Yes. OpenRouter says Zero Data Retention routing is available on every plan, and prompts and completions are not logged by default. It still logs request metadata such as timestamps, model and token counts.
Is Amazon Bedrock HIPAA eligible?
Yes. Amazon Bedrock is on AWS's HIPAA-eligible services list, last updated 15 October 2026. You still need a business associate agreement (BAA) with AWS before you process PHI, and you should check each model's data-retention mode against your policy.
Can I use Gemini models on Amazon Bedrock?
No. Bedrock's pricing page lists Google's open-weight Gemma models but not Gemini. Gemini is available through Google Vertex AI (Agent Platform) and through OpenRouter.
What is a self-hosted alternative to OpenRouter?
LiteLLM is an open-source option: the proxy is MIT-licensed outside its enterprise folder and runs in your own infrastructure with your own provider keys.
Next step
Compare the two most-searched pairs side by side in Amazon Bedrock vs OpenRouter and Google Vertex AI vs OpenRouter. If you are choosing the platform your agents will run on, the Enterprise AI Agent Platform Buyer's Guide covers the governance questions to settle first. Every gateway and model platform we track is in the AI infrastructure category.
Related Workflows
Reputation Monitoring & Alerts
Get instant alerts when someone mentions your business online and respond to reviews across platforms from one place.
View workflowAI Review Response Agent
Monitors Google, Yelp, and Facebook reviews; drafts brand-voice responses grounded in past replies; queues for one-click human approval.
View workflowAI Contract Review & Redlining for Small Firms
Redline agreements against your own playbook inside Word, so a two-lawyer firm reviews contracts at the pace of a twenty-lawyer one.
View workflowKeep Reading
Baseten vs Modal vs Together AI (2026): When to Leave a Token API for Your Own GPUs
Baseten runs dedicated model deployments, Modal sells serverless GPU seconds, and Together AI and Groq sell hosted open-model tokens. Here is what each costs and the volume at which your own GPU starts to pay.
Composio vs Arcade vs MCP in 2026: Agent Tool Integration Pricing, Auth and Security
The managed tool layers for AI agents price the same action four different ways. Here is what Composio, Arcade.dev and Pipedream Connect publish, what running your own MCP servers costs you instead, and a worked monthly bill.
Bedrock AgentCore vs Databricks Mosaic AI vs Google Agent Platform: 2026 Pricing
Where to build, host and govern production agents on AWS, Databricks or Google Cloud, with every published per-unit rate and a worked customer-service agent bill.
Found this helpful?
Get weekly AI workflow ideas in your inbox.