Claude Sonnet 4.5 Bedrock end of life: 8 April 2027, Opus 4.1 on 8 January

Summary. Amazon Bedrock retires Claude Opus 4.1 on 8 January 2027 in us-east-1, us-east-2 and us-west-2; it has been Legacy since 8 July 2026 and in public extended access since 8 October 2026. Claude Sonnet 4.5 went Legacy on 8 October 2026 in 31 Regions, including Mumbai and Hyderabad. It enters public extended access on 8 January 2027, when AWS warns the provider can raise prices, and reaches end of life on 8 April 2027, after which calls fail and nothing migrates automatically. From 1 December 2026 Bedrock is the only major platform still serving Sonnet 4.5, because the Claude API, Microsoft Foundry and Google Cloud retire it on 30 November. The targets cost less per token: Sonnet 5.5 at $2 input and $10 output per million tokens against $3 and $15, and Opus 5.5 at $4 and $20 against $15 and $75. The work is in code and contracts: thinking runs by default, prefill and temperature return 400 errors, tokens rise about 30 percent, and no replacement offers the Reserved tier or Bedrock batch today.

About this article. For engineering and platform teams calling Claude Opus 4.1 or Sonnet 4.5 on Amazon Bedrock, including from the Indian Regions. On 7 November 2026 we checked every date, model ID, Region, tier and price against AWS's lifecycle pages, model cards and price data, Anthropic's deprecation, Bedrock, pricing and migration documentation, and the Microsoft and Google schedules. This is general information, not financial or legal advice; vendors change dates and prices often, and no vendor paid for inclusion. Researched with AI assistance and reviewed by our editorial team.

The two Bedrock dates, side by side

These are AWS dates, which can differ from Anthropic's and by Region. Sources: AWS's lifecycle table and model cards, Anthropic's deprecations page, and the Microsoft and Google schedules.

Both AWS cards mark the model Legacy in all Regions; after EOL, requests fail with a ResourceNotFoundException. Sonnet 4.5's Bedrock floor was 30 September 2026, so AWS kept it about six months past its minimum.

One gap: the Sonnet 4.5 card lists GovCloud, me-central-1 and me-south-1, but the EOL table does not, and AWS says a missing Region is no promise of availability. Check those Regions with GetFoundationModel. On Claude Platform on AWS, Anthropic's dates apply, so Sonnet 4.5 ends there on 30 November 2026. For every Claude date by platform, see our 2027 Claude retirement calendar.

What Legacy and extended access change before EOL

On Bedrock, several controls tighten months before EOL:

• New customers are blocked. AWS does not define a new customer, so plan as if a newly created account, such as a fresh staging or subsidiary account, counts as one.

• Idle access lapses. Existing customers may lose access after 15 days without use, which catches monthly jobs, quarterly reports and disaster-recovery accounts that only wake up in a drill.

• No new Provisioned Throughput. Endpoints created before the Legacy date keep working until EOL.

• Extended-access pricing. For models with EOL dates after 1 February 2026, public extended access starts after at least three months in Legacy. Active users keep access until EOL but should expect higher prices set by the provider, and AWS notifies accounts of the date and any price change.

• No automatic migration. After EOL, requests fail unless you hold a private arrangement with the provider.

On 7 November 2026 the Bedrock pricing page shows no extended-access rate for either model, and AWS's price data still lists Opus 4.1 at $15 input and $75 output per million tokens, a month after its extended-access start. Treat the January price for Sonnet 4.5 as unknown until your notice arrives.

Model IDs, endpoints and the SDK trap

Sonnet 4.5 and Opus 4.1 run only on the bedrock-runtime endpoint, through InvokeModel and Converse. The replacements also take the Messages API, on bedrock-runtime and on the newer bedrock-mantle endpoint, and AWS recommends bedrock-runtime for new applications. Anthropic documents the ARN-versioned models, including both models here, on a separate page from the Messages API path. The table below comes from each AWS model card.

Swap the whole ID: global.anthropic.claude-sonnet-4-5-20250929-v1:0 becomes global.anthropic.claude-sonnet-5-5, and us.anthropic.claude-opus-4-1-20250805-v1:0 becomes us.anthropic.claude-opus-5-5. The new IDs carry no date and no -v1:0 suffix, so regexes in gateways, cost-allocation rules or IAM policies that expect one stop matching. Sonnet 5.5 also has no jp. profile and no longer takes New Zealand traffic on au., so callers in Tokyo, Osaka and New Zealand get Global routing only.

Anthropic flags a trap that can break both old models before EOL: its newest SDKs send API version 2026-10-21 by default, including through Bedrock, and the ARN-versioned models can answer it with 400 errors. Until you migrate, pin the SDK with Anthropic's install line for these models; Boto3 code that sets anthropic_version in the request body is not affected.

            pip install -U "anthropic[bedrock]<2"
          

The second common error comes from passing a base model ID. AWS serves these models through cross-Region inference, not on-demand throughput, so a call with anthropic.claude-sonnet-4-5-20250929-v1:0 returns a 400 asking for an inference profile ID or ARN. Among the targets, only Opus 5.5 accepts its bare ID, and only In-Region in London.

Cross-Region routing and data residency

From Sonnet 4.5 onwards, Bedrock's global profiles carry no premium and geographic or Regional routing costs 10 percent more; Opus 4.1 predates the scheme and keeps its old pricing. AWS's price data shows Sonnet 4.5 at $3 and $15 per million tokens on the global profile against $3.30 and $16.50 on Geo, and Sonnet 5.5 at $2 and $10 against $2.20 and $11.

India. AWS's cards for Sonnet 4.5, Sonnet 5.5, Opus 5.5, Opus 4.8 and Haiku 5.5 mark Mumbai and Hyderabad as Global only, with no In-Region or Geo option. A Mumbai application on Sonnet 4.5 already sends prompts anywhere in the global profile, across the Americas, EMEA and Asia Pacific, and moving to those models changes nothing.

The exception is Sonnet 5, whose AWS card lists an in. profile that routes only between Mumbai and Hyderabad. Anthropic's Bedrock Region table shows the Indian Regions as Global only and names only US, EU, JP and AU profiles, so the vendors disagree. AWS's advice for residency is to call GetInferenceProfile from each source Region and check every destination returned; do that in ap-south-1 before relying on in.anthropic.claude-sonnet-5. If it holds, Sonnet 5 is the only option here that keeps Claude processing in India, at the 10 percent Geo premium and with a Bedrock floor of 30 June 2027, so it buys time rather than a long-term home. Record the routing choice in your DPDP processing records; our DPDP engineering playbook covers the controls.

EU. The eu. profile routes among Frankfurt, Stockholm, Milan, Spain, Ireland and Paris, adding London or Zurich only for traffic that starts there. Opus 5.5 and Sonnet 5 also accept In-Region calls in London.

US and GovCloud. The us. profile routes among N. Virginia, Ohio and Oregon, plus the caller's own Region for N. California and Canada. Sonnet 4.5 reaches GovCloud through a Geo route; among the replacements only Opus 4.8 lists a us-gov. profile, while the others show GovCloud access In-Region on bedrock-mantle in us-gov-west-1.

Capacity: Reserved tier, Provisioned Throughput, batch and quotas

1. Reserved tier. Sonnet 4.5 offers Reserved, a term commitment your AWS account team enables. No replacement card lists Reserved, and the Reserved table on the Bedrock pricing page covers only Opus 4.6, Opus 4.5, Haiku 4.5, Sonnet 4.6 and Sonnet 4.5. Sonnet 4.6 keeps Reserved, but it is an older model with a Bedrock floor of 17 February 2027. Ask your account team now whether Reserved will reach Sonnet 5.5 before 8 April, and size Standard quotas as the fallback.

1. Provisioned Throughput. PT created before the Legacy date runs until EOL, then stops. AWS publishes no PT price for Anthropic models and refers you to your account team.

1. Batch. The pricing page lists Bedrock batch rates for Sonnet 4.5 but marks batch unavailable for Sonnet 5.5, Sonnet 5, Opus 5.5, Opus 4.8 and Haiku 5.5. Compare a nightly batch job against your current batch rate, because it moves to on-demand billing.

1. Quotas. On the Messages API path, Anthropic lists a default of 2 million input tokens per minute and says you can request up to 5 million input and 500,000 output tokens per minute without extra Anthropic approval; AWS sets request limits separately. With 30 percent more tokens, a workload at 80 percent of its token quota lands near 104 percent. Raise quotas before the canary.

Replacement models and their Bedrock floors

Prices are Anthropic list prices per million tokens. AWS's price data shows the same input and output rates on Bedrock for Sonnet 4.5, Sonnet 5.5, Opus 4.1, Opus 4.8 and Opus 5.5, and Geo or Regional profiles add 10 percent. Floors are the AWS cards' dates before which the model will not reach EOL.

Anthropic named claude-opus-4-8 as Opus 4.1's replacement and Microsoft's Foundry schedule names claude-opus-5; both name claude-sonnet-5-5 for Sonnet 4.5. The difference is runway. Opus 4.8 follows Bedrock's older policy, with a floor of 28 May 2027 and at least six months in Legacy. Opus 5.5 and Sonnet 5.5 launched after 7 September 2026, so they follow the newer policy: a set six-month Legacy period and floors in late September 2027.

Opus 4.8 and Sonnet 5 are open to all Bedrock customers; Opus 5.5, Sonnet 5.5 and Haiku 5.5 have per-model access criteria in the AWS console. Request access in every account, including idle disaster-recovery ones.

What breaks when Sonnet 4.5 code moves to Sonnet 5.5

Anthropic's Sonnet 5.5 migration guide lists these changes for code written against Sonnet 4.5:

• Thinking runs by default. Without a thinking field, Sonnet 5.5 uses adaptive thinking, billed as output and counted toward max_tokens. To skip up-front thinking, set the thinking type to between_tools at low, medium or high effort; disabled returns a 400 error.

• Thinking budgets fail. The enabled type with budget_tokens, how Sonnet 4.5 did all thinking, returns a 400. Set output_config.effort and test two or three levels.

• Effort defaults to high. Set it on every call; Anthropic suggests medium as the start for well-specified agentic tasks.

• Prefill fails. Replace JSON-forcing assistant prefills with structured outputs.

• Sampling parameters fail. Remove temperature, top_p and top_k.

• Forced tool choice fails. A tool_choice of any or tool is rejected. Use auto with strict tools, at most 20 per request.

• Responses can start with thinking blocks. Code that reads content[0].text breaks. Read blocks by type and pass thinking blocks back unchanged in tool loops.

• More tokens. The same text produces about 30 percent more tokens, and a 2000 by 1500 pixel image about 2.5 times as many.

• More refusals. Declines arrive with a refusal stop reason and categories such as cyber, bio and general_harms.

Two Bedrock specifics. Sonnet 5.5 thinking blocks are bound to the AWS account that produced them, so a failover to a second account should start a fresh conversation. Sonnet 5.5 accepts the computer_20251124 tool with its beta header on Bedrock, but no platform accepts Sonnet 4.5's computer_20250124. The minimum cacheable prompt drops from 1,024 tokens to 512.

Opus 4.1 to Opus 5.5 or Opus 4.8

On Opus 5.5, adaptive thinking is always on, default effort is medium, and tool_choice accepts only auto and none. Opus 4.8 keeps thinking off by default and accepts disabled, but rejects manual budget_tokens, so Opus 4.1 thinking budgets fail on both. Non-default sampling parameters return 400 errors on Claude 4.7 and later, and API version 2026-10-21 removes them for every model.

Decision table

Teams also on Opus 4.6 or Sonnet 4.6 face a separate timetable, covered in our Opus 4.6 and Sonnet 4.6 retirement guide.

Worked cost example

Illustrative arithmetic from Anthropic list prices on the global profile; eCorpIT did not benchmark these models. A Sonnet 4.5 assistant uses, per month, 700 million cache-read tokens, 50 million five-minute cache writes, 250 million uncached input tokens and 100 million output tokens.

Sonnet 5.5 saves 36 percent on identical token counts, about 17 percent after the tokenizer, and costs about 8 percent more if default thinking adds half again to output. That uplift is our assumption; measure output tokens per task at each effort level.

Staying on Sonnet 4.5 costs too: with no extended-access rate published, each 10 percent uplift on the $2,647.50 baseline adds $264.75 a month. On a Geo profile, add 10 percent to every line. Rerun the numbers with your traffic in our LLM token counter.

For Opus 4.1, 200 million input and 20 million output tokens a month cost $3,000 + $1,500 = $4,500. The same tokens cost $1,000 + $500 = $1,500 on Opus 4.8 and $800 + $400 = $1,200 on Opus 5.5 through the global profile. Two adjustments apply. Both newer models use the tokenizer that produces about 30 percent more tokens, taking Opus 5.5 to about $1,560. And because Opus 4.1 runs only on the us. profile without a premium, keeping US-only routing on Opus 5.5 costs $880 + $440 = $1,320 before the tokenizer effect.

A test plan that fits both deadlines

If you are past a row's date, compress the remaining rows rather than skipping them. The Sonnet column ends in March, which means paying extended-access rates from 8 January 2027; to avoid that, run Sonnet 4.5 on the Opus 4.1 dates.

A minimal lifecycle check to run nightly per Region:

            import boto3
from botocore.exceptions import ClientError

OLD = ["anthropic.claude-opus-4-1-20250805-v1:0",
       "anthropic.claude-sonnet-4-5-20250929-v1:0"]

for region in ["ap-south-1", "us-east-1", "eu-west-1"]:
    bedrock = boto3.client("bedrock", region_name=region)
    for model_id in OLD:
        try:
            details = bedrock.get_foundation_model(modelIdentifier=model_id)["modelDetails"]
            print(region, model_id, details.get("modelLifecycle", {}).get("status"))
        except ClientError as err:
            print(region, model_id, err.response["Error"]["Code"])
          

Route model choice through a gateway so step 8 is a config change with instant rollback. Our LiteLLM, Portkey, Kong and TrueFoundry comparison covers the options, and our guide to catching silent agent failures in CI/CD shows how to gate step 5.

How eCorpIT can help

eCorpIT is a Gurugram-based technology organisation, founded in 2021, assessed at CMMI Level 5 and MSME certified, with senior-led engineering teams working across AWS, Microsoft and Google platforms. Our LLM migration and cost optimisation service inventories Claude calls across your AWS accounts and Regions, ports requests to Sonnet 5.5 or Opus 5.5, and gates the cut-over with evals built from your own traffic. Talk to us at /contact-us/.

Last updated: 7 November 2026.

References

1. Model lifecycle (Legacy), Amazon Bedrock User Guide — AWS

1. Claude Sonnet 4.5 model card, Amazon Bedrock User Guide — AWS

1. Model deprecations — Claude Platform Docs, Anthropic

1. Model retirement schedule for Microsoft Foundry, last updated 23 September 2026 — Microsoft Learn

1. Partner model deprecations, last updated 6 November 2026 — Google Cloud

1. Pricing — Claude Platform Docs, Anthropic

1. Claude Sonnet 5.5 migration guide — Claude Platform Docs, Anthropic

1. Amazon Bedrock pricing — AWS

1. Claude Opus 4.1 model card, Amazon Bedrock User Guide — AWS

1. Amazon Bedrock foundation model price list data, published 12 October 2026 — AWS

1. Claude Sonnet 5.5 model card, Amazon Bedrock User Guide — AWS

1. Claude on Amazon Bedrock (Opus 4.6 and earlier) — Claude Platform Docs, Anthropic

1. Claude in Amazon Bedrock (Opus 4.7 and later) — Claude Platform Docs, Anthropic

1. Claude Sonnet 5 model card, Amazon Bedrock User Guide — AWS

1. Claude Opus 5.5 model card, Amazon Bedrock User Guide — AWS

1. Claude Opus 4.8 model card, Amazon Bedrock User Guide — AWS

1. Claude Haiku 5.5 model card, Amazon Bedrock User Guide — AWS

1. Supported Regions and models for inference profiles, Amazon Bedrock User Guide — AWS

1. Claude Sonnet 4.6 model card, Amazon Bedrock User Guide — AWS

1. Model lifecycle, Amazon Bedrock User Guide — AWS