Every Amazon Bedrock Model ID in One Place (October 2026): 136 Models, Their Inference Profiles, and the Half Already Past Their Guaranteed Date

Amazon Bedrock has 136 models from 19 providers as of October 10, 2026, and every one of them has a model ID such as anthropic.claude-opus-5-5, openai.gpt-6-sol or meta.llama4-maverick-17b-instruct-v1:0. Many of the newest models, every current Claude model among them, are not called by that bare ID but through an inference profile with a prefix such as global., us. or eu., and some IDs differ between Bedrock's two endpoints. This page lists all of them, provider by provider, with their profiles and status; to check one in your own account, run aws bedrock list-foundation-models in your Region or copy the ID from the model's page in the Bedrock console. The finding that matters most: 70 of the 136 models, 51%, are already past the date until which AWS guaranteed them, so AWS may now schedule their retirement, and 11 model cards are still published for models that are Legacy or already switched off, including Claude Sonnet 4.5, which went Legacy two days ago, and Claude Sonnet 4, which ends on October 14.

Jake runs a phone repair shop. Ethan, who looks after AWS for a few local businesses, spent a Friday afternoon auditing a client's AI features before a renewal meeting. Their code base called nine different Bedrock model IDs, written by three developers over two years. Two were Legacy models with end-of-life dates already set, one was an ID that only works on a different endpoint, and one had been switched off for a month without anyone noticing, because the feature that used it ran once a quarter. Jake's question was the one every owner asks: "How do I know which of these will stop working?" This page is the list Ethan wished he had that afternoon: every model ID on Bedrock, how the IDs work, how to get them yourself, what the lifecycle dates mean, and which models are on borrowed time.

If Bedrock itself is new to you, start with our plain-English guide to what Amazon Bedrock is; this page assumes only that.

How Amazon Bedrock model IDs work

Every model on Bedrock has a base model ID made of the provider's name, a dot, and the model's own name and version: anthropic.claude-sonnet-5-5, amazon.nova-2-lite-v1:0, qwen.qwen3-coder-next. There is no single format after the dot. Older models carry a date and a version suffix, such as anthropic.claude-opus-4-1-20250805-v1:0; newer ones use plain names. Some IDs end in :0 or :1, which is a version of that model on Bedrock, so stability.stable-image-core-v1:1 is not the same thing as v1:0.

Many current models are not called by the base ID at all. They are called through a system-defined inference profile: the same ID with a geography prefix in front, which tells Bedrock where it may run the request.

Three rules explain most errors. A geo profile works only when you call it from a Region inside that geography: eu. from an EU Region, jp. from Tokyo or Osaka. Many current models, Claude among them, have no in-Region option on bedrock-runtime in most Regions, so calling the bare ID fails with the "on-demand throughput isn't supported" error, which our on-demand throughput guide explains. And for Anthropic's models, geo and Regional routing cost 10% more than global.

There is a third kind of ID you create yourself: an application inference profile, whose ID is an ARN such as arn:aws:bedrock:us-east-1:111122223333:application-inference-profile/abc123. It points at a model or a system profile, and its job is cost tracking: tag it with a team or project, call it instead of the model ID, and the bill splits itself.

Jake: "So one model can have five different names?"

Ethan: "Like a part with a maker's number and a different number in each supplier's catalog. The base ID is the maker's number. The prefixes say which warehouse ships it. Order the wrong warehouse for your area and the order bounces."

Which Bedrock model ID should you use? A short guide by job

A list of 136 IDs is only useful once you know which row you need. Here is how you narrow it down, by the job you are doing, using models that are Active today:

• You want the strongest general model for reasoning and code: global.anthropic.claude-opus-5-5. At launch it scored above Claude Fable 5.1 on every benchmark compared, at $4 and $20 per million tokens.
• You want a strong everyday model at half that price: global.anthropic.claude-sonnet-5-5, at $2 and $10.
• You run millions of small calls: global.anthropic.claude-haiku-5-5, at $0.10 and $0.50 for prompts up to 100,000 tokens.
• You prefer OpenAI's models: openai.gpt-6-sol, openai.gpt-6-luna, openai.gpt-6-astra or openai.gpt-6.1-sol, each with global. and us. profiles; our GPT-6 on Bedrock guide compares them.
• You want a model whose weights you could also run yourself: gpt-oss, Qwen3, Llama, Gemma, DeepSeek and Kimi are all on Bedrock, so you can prototype on AWS and later move to your own hardware, or the reverse.
• You need embeddings for search: amazon.titan-embed-text-v2:0, cohere.embed-v4:0 with its 128K-token input, or amazon.nova-2-multimodal-embeddings-v1:0 if your documents include images; TwelveLabs Marengo Embed 3.0 for video.
• You need images: the Stability AI models, such as stability.sd3-5-large-v1:0 and stability.stable-image-ultra-v1:1. Amazon's Nova Canvas reached end of life on September 30, 2026, so remove it if your code still calls it.
• You need to understand video: TwelveLabs Pegasus 1.5, global.twelvelabs.pegasus-1-5-v1:0, added October 5, 2026.
• You need real-time speech: amazon.nova-2-sonic-v1:0; the first Nova Sonic ended on September 14, 2026.
• Your data must stay in India: pick from the six models with an in. profile, listed in the FAQ below.

Whatever you pick, check its status column in the tables below before you build on it, and prefer a model whose "no EOL before" date is still ahead of you.

How to get a Bedrock model ID yourself

This page is a snapshot. The source of truth is your own account in your own Region, and you have three ways to read it. Use whichever you already have open; they all return the same IDs.

In the console

1. Open the Amazon Bedrock console in the Region your code will run in.
2. Go to the Model catalog and choose the model you want.
3. Copy the model ID from its model card, where every inference profile ID sits next to it.
4. Open the model in a playground and send one prompt, so you know your account can call it before you write any code.

With the AWS CLI

List the models in a Region, with their lifecycle status. The --by-provider filter narrows the list; there is no filter for status, so the query pulls it out of the results:

aws bedrock list-foundation-models \
  --region us-east-1 \
  --by-provider anthropic \
  --query "modelSummaries[].{id:modelId,status:modelLifecycle.status}" \
  --output table

The modelLifecycle block also carries legacyTime, publicExtendedAccessTime and endOfLifeTime when they are set, so a Legacy model's dates come back with it. List the system inference profiles, which is where the global. and us. IDs live:

aws bedrock list-inference-profiles \
  --region us-east-1 \
  --type-equals SYSTEM_DEFINED \
  --query "inferenceProfileSummaries[].{id:inferenceProfileId,name:inferenceProfileName}" \
  --output table

And check whether your account can actually call a third-party model, which is a separate question from whether it exists:

aws bedrock get-foundation-model-availability \
  --region us-east-1 \
  --model-id anthropic.claude-sonnet-5-5

You want AVAILABLE and AUTHORIZED in every field. These access commands need AWS CLI 2.27.42 or later.

On the Mantle endpoint

bedrock-mantle has its own model list in the OpenAI style, and it reports per-model status for your account, including models that are unavailable because of your data retention setting:

curl https://bedrock-mantle.us-east-1.api.aws/v1/models \
  -H "x-api-key: $BEDROCK_API_KEY"

Legacy and end-of-life: the dates that matter right now

Every Bedrock model moves through three states: Active, Legacy and End-of-Life. The rules depend on when the model launched. For models launched before September 7, 2026, AWS promises that a model "will remain on Amazon Bedrock for at least 12 months" after launch, that it "will be in the Legacy state for at least 6 months before the EOL date," and, for models with end-of-life dates after February 1, 2026, that after at least three months in Legacy it enters "public extended access," during which "you should expect higher pricing, which will be set by the model provider." For models launched on or after September 7, 2026, each model card shows an "EOL no sooner than" date and a Legacy period of either six months or 45 days; GPT-6.1 Sol is the exception and follows OpenAI's own deprecation terms.

Two Legacy rules catch people out. "New customers can't adopt the model," so if your team never used a Legacy model, you cannot start now. And "existing customers may lose access after 15 days of inactivity," which is how your quarterly job, like the one in Ethan's audit, can wake up to find its model gone. After the end-of-life date, "requests made to it will fail," and "migration will not happen automatically."

The models that are Legacy or already ended, as of October 10, 2026:

Lifecycle can differ by Region, and for any Region you use, list-foundation-models there is the final word. Claude Sonnet 4.5 matters beyond its own users: it is what Claude Code's unpinned sonnet alias and background tasks still use on Bedrock, which our Claude Code on Bedrock guide shows how to pin away from.

The 70 models already past their minimum date

Here is the number that made Ethan's client sit up. Each model card shows an "EOL no sooner than" date. For 70 of the 136 models, that date has already passed. Six of those are already Legacy and five have ended, but 59 are still listed as Active. Passing the date does not mean a model stops working; it means AWS's minimum promise has been met, and the model can be moved to Legacy whenever the provider retires it, with at least six months of notice for models launched before September 7, 2026.

The Active models in that group, by provider:

• Meta: Llama 3 8B and 70B, all three Llama 3.1 models, all four Llama 3.2 models, Llama 3.3 70B, and both Llama 4 models, Scout and Maverick.
• Amazon: Nova Pro, Nova Lite and Nova Micro, Titan Text Embeddings V2, Titan Embeddings G1 Text and Titan Multimodal Embeddings G1.
• Mistral AI: Mistral 7B, Mixtral 8x7B, Mistral Large (2402), Pixtral Large, Devstral 2 123B, Voxtral Mini 3B and Magistral Small 2509.
• Qwen: Qwen3 235B A22B 2507, Qwen3 32B, Qwen3 Coder 480B, Qwen3 Coder 30B A3B, Qwen3 Next 80B and Qwen3 VL 235B.
• OpenAI: gpt-oss-120b and gpt-oss-20b, both past August 5, 2026.
• DeepSeek: DeepSeek-R1 and DeepSeek-V3.1.
• Google: all three Gemma 3 models.
• Others: Cohere's Embed English, Embed Multilingual, Embed v4 and Rerank 3.5; Writer's Palmyra X4; Luma Ray2; TwelveLabs Pegasus 1.2; NVIDIA Nemotron Nano 9B v2; and thirteen of Stability AI's image tools.

You do not need to panic, but you do need to watch. If a feature you ship depends on one of these, subscribe to AWS Health notifications for Bedrock, re-run your list-foundation-models query monthly, and write down which newer model you would move to. You will not be stuck for choice: every provider on this list already has a newer model on Bedrock.

Jake: "So half the shelf is past its use-by date, but still on the shelf?"

Ethan: "Past its guarantee date, not its use-by date. Like a phone model that is out of the maker's promised update window: it still works today, and you would be silly to build your whole shop around it."

Every Amazon Bedrock model ID, by provider (checked October 10, 2026)

Each table lists the bedrock-runtime ID, the inference profile prefixes the model offers on bedrock-runtime (combine the prefix with the model ID, as in global.anthropic.claude-opus-5-5), the bedrock-mantle ID where the model is on Mantle, the launch date, and its status. "Profiles only" means the model has no in-Region option on bedrock-runtime, so you must use a profile. "Minimum date passed" means the model's "EOL no sooner than" date is behind us. Providers are ordered roughly by how many people search for them.

Anthropic (Claude): 19 models

OpenAI: 15 models

Amazon (Nova and Titan): 12 models

Meta (Llama): 12 models

Mistral AI: 13 models

Qwen (Alibaba): 7 models

DeepSeek: 3 models

Google (Gemma): 6 models

Moonshot AI (Kimi): 3 models

MiniMax: 3 models

Z.AI (GLM): 4 models

xAI (Grok): 3 models

NVIDIA (Nemotron): 4 models

Cohere: 6 models

AI21 Labs: 2 models

Writer (Palmyra): 3 models

TwelveLabs: 4 models

Stability AI: 16 models

Luma AI: 1 model

Same model, different ID: bedrock-runtime vs bedrock-mantle

Bedrock has two endpoints. bedrock-runtime is the original, with the Converse and InvokeModel APIs and Bedrock's platform features such as Guardrails, Knowledge Bases and Agents. bedrock-mantle is newer and speaks the OpenAI-style APIs and Anthropic's Messages API natively. Most models with a Mantle entry use the same ID on both. Some do not, and if you copy an ID from one endpoint's example into code for the other, you will earn an "invalid model identifier" error:

The rest of the Qwen3 line has no single rule. Qwen3 32B and Qwen3 Coder 30B follow the -v1:0 pattern above; Qwen3 Next 80B and Qwen3 VL 235B use a plain ID on bedrock-runtime and add -instruct on Mantle; only Qwen3 Coder Next, qwen.qwen3-coder-next, is identical on both. The Kimi K2 Thinking row is the subtle one: the provider prefix itself differs, moonshot. against moonshotai..

Nine models exist only on bedrock-mantle and have no bedrock-runtime ID at all: Claude Mythos 5; Gemma 4 26B-A4B, Gemma 4 31B and Gemma 4 E2B; GPT-5.4 and GPT-5.5; the two Daybreak cybersecurity models, Daybreak Red (GPT-5.6 Cyber) and Daybreak Blue (GPT-5.6 Sol); and Grok 4.3. If you call any of them through Converse or InvokeModel on bedrock-runtime, the request fails no matter how the ID is spelled. Our Mantle vs bedrock-runtime guide shows how to call Mantle instead.

Which models offer Priority, Flex, Reserved, Batch or Ultrafast

Bedrock now sells several service tiers alongside Standard pay-per-token: Priority for faster responses at a premium, Flex for cheaper non-urgent work, Reserved for committed throughput, and a few model-specific extras. The model cards on October 10 show:

• Priority and Flex: 50 models, including Nova 2 Lite and Nova Pro, DeepSeek V3.2, every Gemma 3 and Gemma 4 model, every MiniMax model, most of Mistral's line, all three Kimi models, the NVIDIA Nemotron 3 models and gpt-oss. None of the Claude, GPT-5.6 or GPT-6 models are on that list.
• Reserved: five older Claude models: Haiku 4.5, Opus 4.5, Opus 4.6, Sonnet 4.5 and Sonnet 4.6.
• Batch: listed in the service tier table of the Claude Opus 5 card and no other.
• Ultrafast: GPT-6 Astra and GPT-6.1 Sol.

Everything else is Standard only, which includes Claude Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1. If you send a service_tier value your model does not support, the request is rejected, so check your model's card before you set one. If you want Flex pricing for overnight batch-style work, you will find it among the open-weight and Amazon models, not the Claude ones.

Using a Bedrock model ID in code: one call, three kinds of ID

In code, the model ID is a single string, and switching between a profile and a bare ID is a one-line change. With boto3 and the Converse API, which works across most text models on bedrock-runtime:

import boto3

client = boto3.client("bedrock-runtime", region_name="eu-west-1")

MODEL = "global.anthropic.claude-sonnet-5-5"   # global profile: base price, any Region
# MODEL = "eu.anthropic.claude-sonnet-5-5"     # EU profile: stays in EU Regions, +10%

response = client.converse(
    modelId=MODEL,
    messages=[{"role": "user", "content": [{"text": "Reply with the single word OK."}]}],
)
print(response["output"]["message"]["content"][0]["text"])

A model that has in-Region access in the Region you call from can also take its bare ID in the same place; the Region tables on its model card tell you where that works. If you use an application inference profile, you pass its full ARN as modelId in exactly the same place. If you call bedrock-mantle instead, you use the Mantle ID from the tables and Mantle's own base URL, https://bedrock-mantle.<region>.api.aws, with the OpenAI or Anthropic SDK. A good habit for your own code: keep every model ID in one configuration file rather than scattered through it, so that when a model goes Legacy you change one line and your audit takes minutes instead of an afternoon.

"The provided model identifier is invalid": the five usual causes

Most model ID errors on Bedrock come from five mistakes, in rough order of frequency:

1. The bare ID where a profile is needed. Current Claude and GPT models, among others, need global. or a geo prefix on bedrock-runtime in most Regions.
2. A geo profile from the wrong geography. us. from Frankfurt or eu. from Virginia fails; use global. or the profile for where your code runs.
3. The other endpoint's ID. A Mantle-style ID such as openai.gpt-oss-120b sent to bedrock-runtime, or the reverse; see the table above.
4. A model that is not in that Region. Check the model card's Region table, or run list-foundation-models in that Region.
5. A typo or a stale version. v1:0 against v1:1, a missing date stamp, or a model that has reached end of life. An ID from an old blog post or tutorial is the usual source.

The ID must also match your IAM policy. A foundation-model ARN has no account number, arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-sonnet-5-5, while an inference profile ARN has one, arn:aws:bedrock:us-east-1:111122223333:inference-profile/global.anthropic.claude-sonnet-5-5. Profiles route across Regions, so policies for them usually use a wildcard Region on the foundation-model ARN. Our plain-English IAM guide explains how those statements are read.

Ethan's audit ended with a one-page table: nine IDs, nine replacements, two dates circled in red, and one quarterly feature quietly moved to a model that will still be there next year. The meeting took twenty minutes, most of it spent on the question Jake asked first.

A list of 136 model IDs looks like a reference page, and it is, but the useful part is the dates beside them. Models on Bedrock now come and go on a published schedule, with a notice period, a price rise near the end and a hard stop. None of that is a problem if you look once a month. Ethan now puts a calendar reminder next to every model ID he ships, and Jake, who will never call a model directly, has started asking his software suppliers which ones their tools depend on. Copy the IDs you need, check the status column, and write down the date.

Revision note. Written October 10, 2026, from all 136 Bedrock model cards. "A stitch in time saves nine," says the old proverb; one monthly look at the lifecycle column saves a quarterly job from failing in silence.