• October 21, 2026
• 13 min read

Bedrock — Models, Access, Regions & the Converse API¶

How you reach a model on Bedrock (access, model IDs, inference profiles, Regions), how you call it (Converse), and what limits you (quotas, tokens, price). Model facts below come from the AWS model cards and user guide as of 2026-10-21 and change often: treat the tables as a snapshot and re-check the model card before you pin an ID.

How Model Access Works Now¶

The old "request access per model" console page is gone in commercial Regions. Access is enabled on first use, and you control it with IAM and SCPs.

# Check that a model is usable before a smoke test (control-plane client)
import boto3

bedrock = boto3.client("bedrock", region_name="us-east-1")
av = bedrock.get_foundation_model_availability(modelId="anthropic.claude-haiku-5-5")
assert av["agreementAvailability"]["status"] == "AVAILABLE", av
profiles = bedrock.list_inference_profiles(typeEquals="SYSTEM_DEFINED")["inferenceProfileSummaries"]
print([p["inferenceProfileId"] for p in profiles if p["inferenceProfileId"].startswith("global.")][:3])

Model Catalog Snapshot¶

Model cards list every ID. Rows reflect the model cards as of 2026-10-21 and what each card marks as supported on bedrock-runtime.

Anthropic's older line (Claude 4.x, 3.x) is still listed. Some models are gated (for example Claude Mythos 5 and 5.1 are marked "Gated" in the caching table); you do not get them by calling the API.

Find IDs programmatically instead of copying them:

Model lifecycle¶

Every model is Active, Legacy or End-of-Life. A Legacy model still works, but new customers cannot adopt it and existing customers "may lose access after 15 days of inactivity". After the EOL date requests fail and nothing migrates automatically. For models launched on or after 2026-09-07 the model card shows an "EOL no sooner than" date (Sonnet 5.5: 2027-09-28) and a legacy period of 6 months (a few models: 45 days). Put the model ID in configuration, not in code, and alert on LegacyModelInvocations (CloudWatch).

Model IDs vs Inference Profiles¶

The modelId parameter of Converse accepts four kinds of identifier:

flowchart TD
  C["client.converse(modelId=...)"] --> ID{"which ID?"}
  ID -->|"anthropic.claude-…"| FM["foundation model<br/>one Region"]
  ID -->|"us. / eu. / au. / jp."| GEO["geographic profile<br/>routes inside a geography"]
  ID -->|"global."| GL["global profile<br/>routes to any commercial Region"]
  ID -->|"…:application-inference-profile/…"| APP["application profile<br/>cost tags, wraps one of the above"]
  GEO --> DEST["destination Regions<br/>(depend on SOURCE Region)"]
  GL --> DEST

Rules that bite in practice:

• Destination Regions depend on the Region you call from. us.anthropic.claude-haiku-5-5 from us-west-1 can route to four US Regions; from us-east-1 to three. For data-residency claims call get_inference_profile from every source Region you use; AWS says not to rely on the geographic prefix alone.
• Cross-Region calls need IAM access to the profile and to the foundation model in every destination Region. If any destination Region is blocked by an SCP, the request fails even if others are open. Global profiles need an SCP that allows aws:RequestedRegion = unspecified.
• Price is that of the Region you call from; there is no routing surcharge. CloudTrail logs the call in the source Region with additionalEventData.inferenceRegion showing where it ran.
• Inference profiles do not support Provisioned Throughput. Prompts and outputs may be stored in opt-in Regions for abuse detection when a profile routes there.
• Some profiles lack features: for example, AWS lists structured outputs as unsupported for Claude Haiku 4.5 through the India (in.) profile. Check the model card per profile.

Converse vs InvokeModel¶

Embeddings, image generation (Nova Canvas, Stability) and video models are called with InvokeModel (or async invoke); Converse is for text-and-multimodal chat models.

Anatomy of a Converse Request¶

Fields below were read from the botocore model (* = required):

Rules: the conversation alternates user and assistant; tool results go in a user message; maxTokens caps thinking tokens plus text; unsupported fields fail with ValidationException instead of being ignored.

Anatomy of a Converse Response¶

resp = client.converse(modelId=MODEL, messages=[...])
msg = resp["output"]["message"]          # {"role": "assistant", "content": [blocks]}
resp["stopReason"]                       # why generation stopped: see table
resp["usage"]                            # inputTokens, outputTokens, totalTokens (+ cache fields)
resp["metrics"]["latencyMs"]             # server-side latency
resp.get("trace")                        # guardrail / prompt-router trace when enabled

A response with several content blocks is normal (reasoning, text, toolUse). Join text blocks instead of reading content[0]:

def reply_text(response: dict) -> str:
    return "".join(b["text"] for b in response["output"]["message"]["content"] if "text" in b)

usage with prompt caching: inputTokens counts only the non-cached part, so total input = inputTokens + cacheReadInputTokens + cacheWriteInputTokens.

Quotas and Throttling¶

Inference is limited by tokens per minute (TPM) and tokens per day (TPD) per model per Region (embedding models by requests per minute). bedrock-runtime and bedrock-mantle have separate quotas. Values are in Service Quotas; they depend on account age and usage, and increases are not automatic.

How a request is charged against the quota:

Cache reads never count. Output burndown (how many quota tokens one output token costs): 15× for Claude 4.8, 10× for Claude Opus 5.5, Sonnet 5, Opus 5 and Fable 5.1 and for GPT-5.6, 5× for Claude 4.7 and older, 1× for other models. For Claude Sonnet 5.5 and Haiku 5.5, check your quotas in Service Quotas. You are billed for actual tokens only.

def estimate_quota_tokens(input_tokens, max_tokens, output_tokens, burndown=1, cache_write=0):
    """(deducted at request start, deducted after completion)"""
    return input_tokens + cache_write + max_tokens, input_tokens + cache_write + output_tokens * burndown

# AWS docs example, 5x burndown: input 3000, cache write 1000, output 1000
assert estimate_quota_tokens(3000, 32000, 1000, burndown=5, cache_write=1000) == (36000, 9000)   # maxTokens too high
assert estimate_quota_tokens(3000, 1250, 1000, burndown=5, cache_write=1000) == (5250, 9000)     # maxTokens right-sized

The lesson: with maxTokens=32000 each request reserves 36,000 quota tokens while it runs, so fewer run in parallel. Size maxTokens to the real answer length.

Capacity options when on-demand throttles:

Pricing Dimensions¶

No prices here: they change. Know what you are billed on so tests and budgets measure the right thing:

IAM and Authentication¶

A role that may call Claude Haiku 5.5 only through the global profile in us-east-1 (pattern copied from AWS's example for another model, with this guide's IDs and the placeholder account 123456789012):

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": ["bedrock:InvokeModel*"],
      "Resource": ["arn:aws:bedrock:us-east-1:123456789012:inference-profile/global.anthropic.claude-haiku-5-5"]
    },
    {
      "Effect": "Allow",
      "Action": ["bedrock:InvokeModel*"],
      "Resource": [
        "arn:aws:bedrock:us-east-1::foundation-model/anthropic.claude-haiku-5-5",
        "arn:aws:bedrock:::foundation-model/anthropic.claude-haiku-5-5"
      ],
      "Condition": {
        "StringLike": {
          "bedrock:InferenceProfileArn": "arn:aws:bedrock:us-east-1:123456789012:inference-profile/global.anthropic.claude-haiku-5-5"
        }
      }
    }
  ]
}

For a geographic profile, list the foundation-model ARN of each destination Region instead of the Region-less ARN.

API keys¶

A Bedrock API key is a bearer token instead of SigV4. Short-term keys last up to 12 hours (or the session) and inherit the creator's permissions; AWS recommends them for production. Long-term keys create an IAM user and are "for exploration only". boto3 reads AWS_BEARER_TOKEN_BEDROCK with no other credentials and sends the header:

Authorization header: Bearer bedrock-api-key-PLACEHOLDER

Control them with bedrock:CallWithBearerToken (deny it to forbid keys) and the bedrock:bearerTokenType condition (SHORT_TERM, LONG_TERM). A long-term key works only in the Region it was created in. In CI prefer an OIDC role (Foundations) over a stored key.

Counting Tokens Before You Call¶

bedrock-runtime has count_tokens(modelId=..., input={"converse": {...}}) (or {"invokeModel": {"body": ...}}); it is free and returns inputTokens. Not every model supports it: the Claude Sonnet 5.5, Haiku 5.5 and Nova 2 Lite cards list it as unsupported on bedrock-runtime. For Claude models without it, AWS documents an HTTP count_tokens call on the bedrock-mantle endpoint (/anthropic/v1/messages/count_tokens, SigV4 or API key; no SDK method). Otherwise read usage from responses and budget from that (page 04).

Quick Rules¶

1. Resolve access before blaming code: get_foundation_model_availability, the Anthropic use-case form, Marketplace permissions, then wait up to 15 minutes.
2. Use inference profile IDs for current models; choose global. for capacity and cost, a geography prefix for residency.
3. Verify routing from each source Region with get_inference_profile before claiming data residency.
4. Branch on stopReason, not on text. max_tokens means the answer is incomplete.
5. Right-size maxTokens. It is reserved against TPM at the start and lowers your concurrency when too high.
6. Keep model IDs in config and watch the lifecycle (LEGACY, LegacyModelInvocations).
7. Scope IAM to the profile ARN and each destination Region's model ARN; add SCPs for Regions.
8. Prefer short-term credentials or short-term API keys; never commit a long-term key.

See also¶

• Amazon Bedrock
• 02 Streaming, Tools, Structured Output, Multimodal & Caching
• 03 Guardrails, Knowledge Bases, Agents & Batch
• 04 Testing & Production
• AWS Foundations: Clients, Config, Retries & Errors
• AWS Foundations: Credentials, IAM & Local Testing
• Amazon SQS
• LiteLLM — One API for 100+ LLM Providers