Bedrock models and providers

Bedrock exposes seven model families from six providers — Anthropic, Meta, Mistral AI, Cohere, AI21 Labs, Amazon (Nova + Titan), and Stability AI — plus a few smaller embedding-and-rerank-only models from Cohere and Titan. This note is the canonical lookup for model IDs, on-demand pricing (USD / MTok in us-east-1 unless noted), context windows, and regional availability. Bedrock also exposes inference profiles — both system-defined cross-region profiles (route invocations across multiple regions for higher quota and failover) and application inference profiles (user-defined wrappers for cost tagging) — which change the model ID you actually call from your code.

See also

1. Anthropic Claude (the most-used family on Bedrock)

Per docs.aws.amazon.com/bedrock/latest/userguide/model-ids.html and Anthropic’s Bedrock-specific docs.

FeatureClaude Opus 4.7Claude Sonnet 4.6Claude Haiku 4.5Claude Haiku 3.5 (legacy)
Bedrock model IDanthropic.claude-opus-4-7-20260420-v1:0anthropic.claude-sonnet-4-6-20260119-v1:0anthropic.claude-haiku-4-5-20251001-v1:0anthropic.claude-3-5-haiku-20241022-v1:0
Cross-region profile ID (US)us.anthropic.claude-opus-4-7-20260420-v1:0us.anthropic.claude-sonnet-4-6-20260119-v1:0us.anthropic.claude-haiku-4-5-20251001-v1:0us.anthropic.claude-3-5-haiku-20241022-v1:0
Input price$5 / MTok$3 / MTok$1 / MTok$0.80 / MTok
Output price$25 / MTok$15 / MTok$5 / MTok$4 / MTok
Cache-write price (5 min)$6.25 / MTok$3.75 / MTok$1.25 / MTok$1 / MTok
Cache-read price$0.50 / MTok$0.30 / MTok$0.10 / MTok$0.08 / MTok
Context window1,000,000 tokens1,000,000 tokens200,000 tokens200,000 tokens
Max output (sync)128k tokens64k tokens64k tokens8k tokens
VisionYesYesYesNo
Tool use (Converse)YesYesYesYes
Extended thinkingAdaptive onlyYesYesNo
Prompt cachingYes (cachePoint block)YesYesYes
Regionsus-east-1/2, us-west-2, eu-central-1, eu-west-1/3, ap-northeast-1, ap-southeast-1/2SameSameSame

Pricing note: Bedrock matches Anthropic’s published prices to the dollar — there is no Bedrock markup on per-token pricing for Claude models. Where Bedrock charges extra is when you turn on Guardrails (\$0.15 / 1k text units), Knowledge Base reranking (\$1 / 1k queries with Cohere Rerank 3), or Agent intelligent prompt routing (\$1 / 1k requests).

Sonnet 4.5 (still available)anthropic.claude-sonnet-4-5-20250929-v1:0 at the same pricing as Sonnet 4.6. Migrate before the inevitable deprecation.

Opus 4.5 (still available)anthropic.claude-opus-4-5-20251101-v1:0 at \$5 / \$25. Equivalent to 4.7 on most non-coding tasks.

2. Meta Llama

Per aws.amazon.com/bedrock/llama/. Llama 4 ships in three sizes — Scout (17B active / 109B total MoE), Maverick (17B active / 400B total MoE), Behemoth (preview, 288B active). Llama 3.x is still actively maintained.

ModelBedrock IDInput / Output (USD/MTok)Context
Llama 4 Maverick 17B Instructmeta.llama4-maverick-17b-instruct-v1:0$0.24 / $0.971,000,000
Llama 4 Scout 17B Instructmeta.llama4-scout-17b-instruct-v1:0$0.17 / $0.6610,000,000
Llama 3.3 70B Instructmeta.llama3-3-70b-instruct-v1:0$0.72 / $0.72128,000
Llama 3.2 90B Instruct (vision)meta.llama3-2-90b-instruct-v1:0$2 / $2128,000
Llama 3.2 11B Instruct (vision)meta.llama3-2-11b-instruct-v1:0$0.16 / $0.16128,000
Llama 3.2 3B Instructmeta.llama3-2-3b-instruct-v1:0$0.15 / $0.15128,000
Llama 3.2 1B Instructmeta.llama3-2-1b-instruct-v1:0$0.10 / $0.10128,000
Llama 3.1 405B Instructmeta.llama3-1-405b-instruct-v1:0$5.32 / $16128,000
Llama 3.1 70B Instructmeta.llama3-1-70b-instruct-v1:0$0.72 / $0.72128,000
Llama 3.1 8B Instructmeta.llama3-1-8b-instruct-v1:0$0.22 / $0.22128,000

Tool use: All Llama 3.1+ models support tool use via Converse API — the runtime injects Meta’s tool-call template behind the scenes.

Scout’s 10M context is on paper; in practice quality degrades past ~1M tokens. Use cross-region inference profiles to access Scout — it’s only available via profile ID us.meta.llama4-scout-17b-instruct-v1:0 (no direct on-demand in single regions).

3. Mistral AI

Per aws.amazon.com/bedrock/mistral/.

ModelBedrock IDInput / Output (USD/MTok)Context
Mistral Large 2 (24.11)mistral.mistral-large-2411-v1:0$2 / $6128,000
Mistral Large (24.02)mistral.mistral-large-2402-v1:0$4 / $1232,000
Mistral Small (24.02)mistral.mistral-small-2402-v1:0$1 / $332,000
Mistral 7B Instructmistral.mistral-7b-instruct-v0:2$0.15 / $0.2032,000
Mixtral 8x7B Instructmistral.mixtral-8x7b-instruct-v0:1$0.45 / $0.7032,000
Pixtral Large (image+text)mistral.pixtral-large-2502-v1:0$2 / $6128,000

Mistral Large 2 is the pick for the family — same intelligence-tier as Llama 3.3 70B but with stronger structured output and tool use.

4. Cohere

Per aws.amazon.com/bedrock/cohere/.

ModelBedrock IDInput / Output (USD/MTok)Context
Command R+ (08-2024)cohere.command-r-plus-v1:0$2.50 / $10128,000
Command R (08-2024)cohere.command-r-v1:0$0.15 / $0.60128,000
Command (legacy)cohere.command-text-v14$1.50 / $24,096
Embed English v3cohere.embed-english-v3$0.10 / MTok (input only)512
Embed Multilingual v3cohere.embed-multilingual-v3$0.10 / MTok (input only)512
Rerank 3.5 (reranker)cohere.rerank-v3-5:0$1 / 1k queries4,096 docs/query

Command R+ is the RAG-purpose-built generation model — it has native citation support in its output schema, which downstream tools can parse. Rerank 3.5 is the only non-OpenSearch reranker available inside Bedrock Knowledge Bases (per the Knowledge Bases docs).

5. AI21 Labs

ModelBedrock IDInput / Output (USD/MTok)Context
Jamba 1.5 Largeai21.jamba-1-5-large-v1:0$2 / $8256,000
Jamba 1.5 Miniai21.jamba-1-5-mini-v1:0$0.20 / $0.40256,000
Jamba Instruct (legacy)ai21.jamba-instruct-v1:0$0.50 / $0.70256,000

Jamba is a hybrid Transformer–Mamba architecture, which gives it ~3× the throughput of pure-transformer models at the same parameter count for very long contexts. Useful for “extract structured data from this 200-page contract” workloads where you need the throughput.

6. Amazon Nova

Amazon’s own foundation models, released at re:Invent 2024 and refreshed at re:Invent 2025. Per docs.aws.amazon.com/nova/.

ModelBedrock IDInput / Output (USD/MTok)ContextModalities
Nova Premieramazon.nova-premier-v1:0$2.50 / $12.501,000,000Text + image + video → text
Nova Proamazon.nova-pro-v1:0$0.80 / $3.20300,000Text + image + video → text
Nova Liteamazon.nova-lite-v1:0$0.06 / $0.24300,000Text + image + video → text
Nova Microamazon.nova-micro-v1:0$0.035 / $0.14128,000Text → text
Nova Canvas (image generation)amazon.nova-canvas-v1:0$0.04–$0.08 / image1024 prompt tokensText → image
Nova Reel (video generation)amazon.nova-reel-v1:0$0.08 / second of video1024 prompt tokensText + image → video (6s)
Nova Sonic (real-time speech)amazon.nova-sonic-v1:0$0.0034 / 1k speech tokens in, $0.0136 / 1k speech tokens outStreamingSpeech ↔ speech

Nova Micro at $0.035/MTok is the cheapest competitive LLM on Bedrock — useful for high-throughput classification, summarization, or as a draft-generator in a draft-and-review pipeline with Sonnet or Opus reviewing.

Nova Pro and Premier both handle video natively — pass an mp4 block in Converse and the model can answer questions about its content. No other Bedrock model does this (except Pixtral Large for image-only inputs).

7. Amazon Titan (legacy, embeddings-focused)

Most Titan generation models are deprecated (or pinned at low quality compared to Nova / Llama). The embedding models are still actively used.

ModelBedrock IDPriceUse
Titan Text Embeddings V2amazon.titan-embed-text-v2:0$0.02 / MTok1024-dim embeddings, default for Knowledge Bases
Titan Text Embeddings V1amazon.titan-embed-text-v1$0.10 / MTok1536-dim, legacy
Titan Multimodal Embeddings G1amazon.titan-embed-image-v1$0.06 / MTok + $0.00006 / image1024-dim, image+text
Titan Image Generator G1amazon.titan-image-generator-v1$0.008–$0.10 / imageSuperseded by Nova Canvas
Titan Text Lite / Expressamazon.titan-text-*-v1$0.15–$1.30 / MTokSuperseded by Nova

8. Stability AI

ModelBedrock IDPriceNotes
Stable Diffusion 3.5 Largestability.sd3-5-large-v1:0$0.08 / imagePremium quality
Stable Diffusion 3.5 Mediumstability.sd3-5-medium-v1:0$0.04 / imageDefault
Stable Image Ultrastability.stable-image-ultra-v1:0$0.14 / imagePhotorealistic
Stable Image Corestability.stable-image-core-v1:0$0.03 / imageCheap
Stable Diffusion XL (legacy)stability.stable-diffusion-xl-v1$0.04–$0.08 / imageLegacy

9. Regional availability matrix (top models, 2026-05)

Per docs.aws.amazon.com/bedrock/latest/userguide/models-regions.html. Most models launch in us-east-1 + us-west-2 first, with EU and APAC following 1–3 months later.

Modelus-east-1us-west-2eu-central-1eu-west-3ap-northeast-1ap-southeast-2
Claude Opus 4.7YesYesYesYesYesYes
Claude Sonnet 4.6YesYesYesYesYesYes
Claude Haiku 4.5YesYesYesYesYesYes
Llama 4 MaverickProfile-onlyProfile-onlyProfile-onlyNoNoNo
Llama 3.3 70BYesYesYesYesNoYes
Mistral Large 2YesYesYesNoNoNo
Cohere Command R+YesYesYesNoNoYes
Nova ProYesYesYesYesYesYes
Nova LiteYesYesYesYesYesYes
Nova MicroYesYesYesYesYesYes

When a model is “profile-only”, it means you must use a cross-region inference profile (e.g. us.meta.llama4-maverick-17b-instruct-v1:0) — the on-demand single-region invocation will return ValidationException.

10. Inference profiles — what changes vs raw model IDs

Per docs.aws.amazon.com/bedrock/latest/userguide/inference-profiles.html.

Two kinds:

  1. System-defined cross-region profiles (created by AWS): IDs like us.anthropic.claude-sonnet-4-6-20260119-v1:0, eu.anthropic.claude-sonnet-4-6-20260119-v1:0, apac.amazon.nova-pro-v1:0. Routes requests across multiple regions in that geography for higher throughput + automatic failover. Quota is the union of underlying regions — e.g. us. Claude profiles aggregate the per-region quotas of us-east-1, us-east-2, us-west-2, and us-east-2.

  2. Application inference profiles (user-defined): wraps a foundation-model or system-defined profile ARN with custom tags. The ARN looks like arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123. Use them for cost tagging — every invocation through the profile carries the tag, so you can break down spend by application, team, or environment in Cost Explorer.

import boto3
 
bedrock = boto3.client("bedrock", region_name="us-east-1")
runtime = boto3.client("bedrock-runtime", region_name="us-east-1")
 
# Create an application inference profile that wraps a cross-region profile
resp = bedrock.create_inference_profile(
    inferenceProfileName="prod-customer-chat",
    description="Production customer-chat workload",
    modelSource={
        "copyFrom": "arn:aws:bedrock:us-east-1::inference-profile/us.anthropic.claude-sonnet-4-6-20260119-v1:0"
    },
    tags=[
        {"key": "Environment", "value": "production"},
        {"key": "Application", "value": "customer-chat"},
        {"key": "CostCenter", "value": "CS-1234"},
    ],
)
profile_arn = resp["inferenceProfileArn"]
 
# Invoke through the profile — looks the same as invoking a model
response = runtime.converse(
    modelId=profile_arn,   # the ARN replaces the model ID
    messages=[{"role": "user", "content": [{"text": "Hello!"}]}],
)

The same pattern in TypeScript:

import {
  BedrockRuntimeClient,
  ConverseCommand,
} from "@aws-sdk/client-bedrock-runtime";
 
const client = new BedrockRuntimeClient({ region: "us-east-1" });
 
const response = await client.send(
  new ConverseCommand({
    modelId: "arn:aws:bedrock:us-east-1:123456789012:application-inference-profile/abc123",
    messages: [{ role: "user", content: [{ text: "Hello!" }] }],
  }),
);

11. Service tiers — Standard, Flex, Priority, Reserved

Per docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html.

TierPricing modifierSLA / latencyUse
Standard (default)1× baseBest-effortDefault. Single per-request invocations.
Flex0.5× base (50% discount)Best-effort, may queue minutesAsynchronous / batchable work where latency doesn’t matter.
Priority1.75× base (75% premium)Lower TTFT, guaranteed throughputLatency-sensitive production (real-time chat).
Reserved (Provisioned Throughput)Hourly fixed price per model unitGuaranteed dedicated capacityHigh-volume steady-state workloads where dedicated capacity wins on $/token.

Specify with the serviceTier field on the Converse request:

response = runtime.converse(
    modelId="anthropic.claude-sonnet-4-6-20260119-v1:0",
    messages=[...],
    serviceTier={"type": "flex"},   # or "priority", "default", "reserved"
)

Provisioned Throughput pricing for Claude Sonnet 4.6 is ~$50/hour per “model unit” (1 MU = ~5k tokens/sec sustained). The breakeven vs on-demand at 100% utilization is around ~60 GTok/month of sustained throughput. Below that, on-demand wins.

12. Selection heuristics

  • Pure coding agent — Claude Sonnet 4.6 or Opus 4.7. Nothing else on Bedrock competes on agentic coding quality.
  • High-throughput classification / extraction — Nova Micro ($0.035 / MTok) or Llama 3.2 3B Instruct ($0.15 / MTok). Use Cohere Embed V3 for embeddings.
  • Long-context document analysis — Llama 4 Scout (10M context, $0.17 / MTok input) or Claude Sonnet 4.6 (1M, $3 / MTok) depending on whether you need the deeper reasoning.
  • RAG — Cohere Command R+ + Cohere Embed V3 + Cohere Rerank 3.5 if you want a coherent stack from one provider; or Sonnet/Haiku + Titan Embeddings V2 if you want AWS-native.
  • Image generation — Nova Canvas ($0.04/image) is the default; Stability Stable Image Ultra ($0.14/image) for photorealism.
  • Real-time speech — Nova Sonic is the only Bedrock option for bidirectional speech-to-speech.

Further reading