AI · Amazon Bedrock

Stop buying one model. Build a model buffet.

Bedrock turns AWS into a provider-neutral AI control plane: OpenAI-compatible APIs, Anthropic-native Messages, AWS-native Converse, 100+ models, IAM and guardrails behind one account. The useful move is not 'migrate to Bedrock'. It is 'make the model replaceable'.

ABC diagram · migration architecture

Apps stay simple. Backoffice routes. Context stays yours.

ABC keeps the architecture legible: applications ask for capabilities; the backoffice chooses and governs the model; context remains portable and independent.

A · APPS
  • Chat / copilots
  • Agents + MCP
  • FastAPI services
  • VidAI / site fleet
B · BACKOFFICE
  • Bedrock Router
  • Mantle · Converse · Invoke
  • IAM · Guardrails · budgets
  • evals · logging · fallbacks
  • OpenAI · Claude · Nova · DeepSeek · Llama · Mistral · Qwen…
C · CONTEXT
  • RAG + vectors
  • S3 / DB / APIs
  • prompts + tools
  • memory + eval sets
  • brand / policy packs
The model buffet lives in B. The application does not need to know whether today’s request is served by OpenAI, Claude, Nova or a specialist model.
Migration

OpenAI and Anthropic can move without a rewrite.

Bedrock now supports OpenAI-compatible Responses / Chat Completions and an Anthropic-native Messages API. That changes the migration conversation from “replace the SDK” to “move the boundary”.

Current
OpenAI SDK
Target
Bedrock Mantle · Responses / Chat Completions
1. Keep the SDK. Change base URL, credentials and model ID.

Bedrock exposes OpenAI-compatible Responses and Chat Completions APIs. For many applications the migration boundary becomes configuration rather than a rewrite.

Current
Anthropic SDK
Target
Bedrock Mantle · Anthropic Messages API
2. Keep the Messages mental model. Point Claude traffic at Bedrock.

Bedrock supports Anthropic-native Messages semantics as well as AWS-native runtime APIs, so Claude can move behind AWS without redesigning the whole application.

Current
Mixed providers
Target
Bedrock Converse / Mantle + external adapters
3. Put one router between the app and every model.

Normalise requests, tools, streaming, metrics and fallbacks once. Let policy choose the model by quality, latency, geography, cost or availability.

OpenAI → Bedrock
from openai import OpenAI

client = OpenAI(
  base_url="https://bedrock-mantle.eu-west-2.api.aws/v1",
  api_key=BEDROCK_API_KEY,
)

response = client.responses.create(
  model=MODEL_ID,
  input="Build the next production slice"
)
Provider-neutral contract
result = ai.run(
  capability="production_agent",
  messages=messages,
  tools=tools,
  max_cost_usd=0.20,
  latency="interactive",
  fallback=True,
)

# Provider selected by policy, not application code.
The buffet

Choose the cheapest model that can do the job well.

Bedrock’s catalogue changes quickly, so the architecture should care about capability classes rather than permanent model names.

Frontier reasoning
OpenAI GPT-5.6 Sol · Claude Opus 4.8 · Claude Sonnet 5
Hard reasoning, coding, autonomous agents, architecture and deep research.
Fast / economical
GPT-5.6 Terra / Luna · Amazon Nova · smaller provider models
High-volume extraction, classification, summaries, routing and background agent work.
Open / specialist
DeepSeek · Meta Llama · Mistral · Qwen · Moonshot/Kimi · MiniMax · Writer
Specialised workloads, cost optimisation, model diversity and second opinions.
Retrieval
Amazon embeddings · multimodal embeddings · provider embedding models
RAG, semantic search, cross-modal retrieval and enterprise knowledge systems.
Image
Stability AI Image Services · Titan Image Generator · Nova Canvas (legacy)
Generation, edit, inpaint, outpaint, upscale, background removal and style workflows.
Video / media
Luma Ray 2 · Nova Reel (legacy) · Nova Sonic / Nova 2 Sonic · TwelveLabs
Video generation, speech-to-speech, media understanding and multimodal search.
2026 lifecycle warning

Do not build a new media platform around Nova Reel or Nova Canvas. Both are legacy in Bedrock and scheduled for end-of-life on 30 September 2026. VidAI should treat every media generator as a replaceable adapter. Luma Ray 2 and Stability services can sit behind the AWS lane today; external providers remain first-class fallbacks.

Routing policy

Ask for a capability. Let B choose the model.

hard_reasoning
frontier
GPT-5.6 Sol / Claude Opus
production_agent
balanced
Claude Sonnet / OpenAI / Nova
cheap_background
fast
Luna / Nova / smaller open model
image_edit
media
Stability Image Services
video
media
Luma Ray 2 + external video adapters
speech
realtime
Nova 2 Sonic / external TTS
ABC diagram · VidAI

A Runware-style router, but owned by VidAI.

Bedrock can be the AWS-native lane, not the prison. VidAI exposes one job API and routes each media request to Bedrock, Runware, Veo, Sora, local GPU or whatever wins on cost, quality and availability that day.

A · APPS
  • VidAI UI / API
  • ABC / BOAI
  • 100-app fleet
  • YouTube / TikTok publishers
B · BACKOFFICE
  • VidAI Media Router
  • queue · retry · cost · SLA
  • Bedrock · Runware · Veo · Sora · local GPU
  • S3/R2 asset pipeline
  • QA + provenance + publish
C · CONTEXT
  • Siema / Gorai / Selena packs
  • prompt templates
  • brand rules
  • asset registry
  • quality / eval history

Text brain

Claude, OpenAI, Nova, DeepSeek and other Bedrock models generate scripts, shot lists, prompts, metadata and QA decisions.

Image lane

Stability Image Services and Titan sit behind the same job contract; external image providers remain pluggable.

Video lane

Luma Ray 2 is an AWS-native video option. Keep Runware, Veo, Sora and local generation as adapters rather than hard dependencies.

Audio lane

Nova Sonic / Nova 2 Sonic can cover realtime speech while specialist TTS/STT providers remain available through the router.

Production rules

Bedrock is most useful when it reduces coupling.

Model IDs are configuration

Business logic should request a capability tier, not hard-code a provider model throughout the codebase.

AWS owns the control plane

IAM, projects/inference profiles, Guardrails, CloudWatch, budgets, logging and regional controls sit outside the model itself.

Context stays yours

RAG sources, vectors, prompts, evaluations, brand packs and agent memory live behind stable interfaces so models can be replaced.

Route by economics

Use frontier models only where they change the result. Cheap models handle the boring 80%; fallbacks protect production availability.

The migration goal is not AWS lock-in. It is model freedom.

Move provider-specific calls behind a stable routing contract. Put security, observability, RAG and economics in the backoffice. Then change models as fast as the market changes.

Roll the dice