Claude Sonnet 4.5 on Amazon Bedrock: a drop-in swap for the agent loop

(Editor's note: the brief came in as "Sonnet 5.5" but the model actually shipping on Bedrock is Sonnet 4.5 — model ID anthropic.claude-sonnet-4-5-20250929-v1:0. We're reviewing what you can actually call today.)

What it is

Claude Sonnet 4.5 is Anthropic's current Sonnet-tier model, available on Amazon Bedrock as a managed endpoint. For developers already on Bedrock, it's a model-ID swap — same InvokeModel and Converse APIs, same IAM, same VPC setup. The Bedrock model card lists which endpoints and APIs are supported for Claude Sonnet 4.5.

The release ships alongside two API-level features worth knowing about: context editing and a memory tool on the Claude Developer Platform, which — paired with Sonnet 4.5 — are aimed at letting agents handle longer-running tasks without hitting context limits.

The next-work-session test

Concrete scenario: you have a Bedrock-backed coding agent running Sonnet 4 behind the Converse API, with an eval harness scoring it on an internal SWE-bench-style set. Next session, you change the model ID to global.anthropic.claude-sonnet-4-5-20250929-v1:0, rerun the eval, and compare. No new SDK, no new auth path, no new deployment.

What actually changes: tool-use trajectories on multi-step tasks, and — if you adopt them — how you manage context across long agent loops. The context editing feature is reported to cut token use by 84% in a 100-turn web search evaluation, and the memory tool allows persistence across sessions for long-running agents. Treat the 84% as Anthropic's claim on their own benchmark, not a result you should expect on your workload until you've measured it.

Pricing

Verified via Anthropic's and partner docs. Base cost for Sonnet 4.5 is $3 per million input tokens and $15 per million output tokens. On Bedrock specifically, routing matters: Sonnet 4.5 is available on Amazon Bedrock, and if you use the global.anthropic.claude-sonnet-4-5-20250929-v1:0 inference profile you avoid the premium, whereas us.anthropic.claude-sonnet-4-5-20250929-v1:0 costs 10% more.

Prompt caching and batch pricing exist as separate line items — check the Bedrock model card before you forecast spend.

What we'd actually use it for

Running the model behind an existing Agent SDK loop and A/B-ing it against Sonnet 4 on a frozen eval set. Tool-use heavy tasks — multi-file refactors, flaky-test triage, retrieval + synthesis pipelines — are where the delta is most likely to show up. The rebrand of the Claude Code SDK to the Claude Agent SDK signals it's positioned as the way to build general-purpose agents, especially paired with Sonnet 4.5.

What we would not do: assume the SWE-bench headline numbers carry over to your repo. Rebuild your eval, run both models, look at tool-call success rates and token spend per completed task.

Limits

• The context and memory features are Claude Developer Platform capabilities. On Bedrock you'll want to confirm which are exposed through the Bedrock API surface vs. only the Anthropic-direct API — the Bedrock model card is the source of truth here.
• The 10% premium on us.* inference profiles is a real tax on default copy-paste code; teams who don't audit model IDs will overpay.
• Vendor benchmark deltas (SWE-bench, context-editing token savings) are vendor-run. Reproduce on your own eval before you ship a cost-savings claim to your PM.
• Nothing here fixes the hard parts of agent engineering: tool schema design, retry policy, failure-mode logging. The model is a dependency, not a strategy.

Try it if

• You're already on Bedrock and your eval harness can swap a model ID in minutes.
• You run long tool-use loops where context bloat is a real cost line.
• You want to A/B against Sonnet 4 before your next sprint planning.
• You need enterprise AWS controls (IAM, VPC, CloudTrail) and can't move to a direct Anthropic API.

Skip it if

• You're not already running evals — swapping models blind is how regressions ship.
• Your workload is single-turn Q&A on short prompts. The agent-loop improvements won't matter and Haiku-tier models will be cheaper.
• You need the newest features the day they land — Bedrock typically trails the direct Anthropic API on feature parity.
• Your budget is dominated by output tokens and you haven't tried prompt caching or batch mode yet. Fix those first; the model swap is a smaller lever.