Quick verdict
Neither platform wins on a spec sheet. This is a cloud operating-model decision:
- AWS-centric stack, want model optionality: Bedrock. One API over Anthropic Claude, Meta Llama, Mistral, Amazon Nova and more, with IAM/VPC security you already operate. Best for teams whose data, identity, and ops already live on AWS. Honest limit: no Gemini and no GPT — the two most-hyped model families are absent from the catalog.
- Data gravity in BigQuery, or Gemini-first: Vertex AI. Native BigQuery grounding with no data movement, Gemini long-context multimodal models, and TPU economics that favour very large batch workloads. Best for GCP-native and analytics-heavy organisations. Honest limit: pricing spans managed inference plus training compute, which makes bills harder to predict, and Claude is not in the first-party catalog.
- Genuinely multi-cloud: pick where your data and security team already are, and route the odd model you're missing through its native API behind a gateway.
Cipher's first-hand rule: the platform follows the data, not the model. Every cross-cloud AI architecture we've reviewed that chose the "better model" over the incumbent data platform paid for it in egress, duplicate governance, and two security reviews instead of one.
Last updated: August 2026. Catalogs and rates shift monthly — verify Bedrock pricing and Vertex AI pricing before committing.
Platform matrix
| Amazon Bedrock | Google Vertex AI | |
|---|---|---|
| Model catalog | Claude, Llama, Mistral, Cohere, Amazon Nova — multi-vendor by design | Gemini family first-party; Llama, Mistral and others via Model Garden |
| Agent runtime | AgentCore (Runtime, Memory, Gateway, Browser, Code Interpreter) | Vertex AI Agent Builder / Agent Engine; originated the A2A protocol |
| Grounding / RAG | Knowledge Bases (managed or customer-managed vector stores) | BigQuery-native grounding + Google Search grounding |
| Safety layer | Guardrails — six policies, works across providers | Safety filters + Model Armor equivalents; explainability tooling |
| Custom training | Fine-tuning on supported models; not a training platform | Full MLOps: AutoML, pipelines, GPU/TPU training |
| Cost shape | Serverless per-token; provisioned throughput for guarantees | Per-token inference plus compute-hour training; TPU batch advantage |
| AU/SG presence | Sydney + Singapore Regions; check per-model availability | Sydney + Singapore regions; check per-model availability |
Per-model availability in Sydney and Singapore changes quarter to quarter on both platforms. If residency is contractual, verify the specific model in the specific region before signing anything — cross-region inference profiles can quietly route traffic elsewhere.
The agent question
If agents are your 2026 roadmap, the platforms diverge more than the model catalogs suggest. Bedrock's answer is AgentCore: a production runtime with metered Memory, Gateway, Browser, and Code Interpreter components that pairs naturally with framework-level code (Strands, LangGraph — see our cutover guide). Google's answer is Agent Builder plus the A2A agent-to-agent protocol it originated, with a managed multi-agent runtime that shipped earlier than AWS's equivalent.
Our read from client work: AgentCore's component billing is easier to map to a unit cost per agent action; Vertex's agent stack is further ahead on multi-agent coordination primitives. Neither locks you out of the other's protocols — MCP support exists on both sides.
Cost: where the bills actually diverge
Headline token rates for comparable models are close. The divergence is the surrounding stack:
- Bedrock: the extras are Knowledge Bases (vector store baseline — OpenSearch Serverless is a real floor cost), Guardrails per-filter meters, and AgentCore components. See Bedrock pricing explained.
- Vertex: the extras are training/tuning compute (GPU/TPU hours), and the savings are real if your grounding corpus already sits in BigQuery — no pipeline, no egress, no second copy.
- Batch at serious scale: third-party analyses in 2026 consistently give Google's TPU serving hardware the edge above roughly ten thousand requests per hour. Below that, developer experience matters more than silicon. Treat all such thresholds as directional.
Choose Bedrock if / choose Vertex if
- Choose Bedrock if: AWS is your system of record; you want Claude plus open-weight fallbacks behind one API; your security review is IAM-shaped; agents will run next to existing AWS workloads.
- Choose Vertex if: BigQuery holds the data your AI needs; Gemini's long-context multimodal capability is a genuine requirement (not a preference); you run large batch inference where TPU economics compound; or you are betting on A2A-style multi-agent systems early.
- Run both if: you are large enough to have real data estates on each cloud. Put the gateway pattern in front (GCP vs AWS for ML covers the broader platform question) and keep prompts and evals portable.
Related: Bedrock KB vs custom RAG · Guardrails production guide · Cloud Engineering
FAQ
Is Bedrock cheaper than Vertex AI? Not universally. For 10–50M tokens/month of standard inference, serverless Bedrock often lands modestly cheaper in third-party comparisons; Vertex pulls ahead on very large batch and when BigQuery grounding removes pipeline costs. Model, region, and access path decide it — not the platform brand.
Can I get Claude on Vertex or Gemini on Bedrock? Claude is available through Vertex AI's partner catalog in some regions; Gemini is not on Bedrock. Verify current regional availability rather than assuming.
Which is better for a regulated Australian company? Both clear the bar with Sydney regions and enterprise controls. The honest answer is whichever cloud your security and data teams already govern well — a second cloud is a second compliance program.
Do I need either, or can I call model APIs directly? Direct APIs are fine for prototypes. The platforms earn their keep at production: private networking, guardrails, grounding, audit trails, and unified billing.
Who helps implement this? Cipher Projects designs and ships production AI on AWS for Australian and Singapore teams — Bedrock-first, with honest advice when Vertex is the better fit for your data estate.
Conclusion
Bedrock vs Vertex AI is a footprint decision wearing a model-comparison costume. Put the platform where your data, identity, and security investment already live; buy missing models through a gateway; and spend the energy you saved on evals and grounding, which move product quality far more than swapping clouds ever will.