"Extended thinking" is Claude's mode for working through hard problems step by step in a dedicated reasoning phase before producing the final answer. It improves results on complex analysis, math, and multi-step planning, at the cost of extra output tokens and latency. Both extended thinking and its successor, adaptive thinking, are supported for Claude on Amazon Bedrock. The catch for anyone writing request bodies today: the classic budget_tokens configuration only works on the model generations that still support manual thinking.
Two mechanisms, split by model generation
| Model | Manual extended thinking (budget_tokens) | Adaptive thinking |
|---|---|---|
| Claude Haiku 4.5 | Supported | No |
| Claude Opus 4.6 / Sonnet 4.6 | Supported | Supported |
| Claude Opus 4.8 / 4.7, Sonnet 5 | Returns an error (400) | Supported (on by default on Sonnet 5) |
| Claude Fable 5 | Returns an error (400) | Always on; cannot be disabled |
With manual extended thinking, you enable thinking and set an explicit token budget. With adaptive thinking, the model decides how much to reason, and you steer depth with an effort setting (for example, effort defaults to high on Opus 4.8) rather than a fixed budget. If you send thinking: {"type": "enabled", "budget_tokens": N} to Opus 4.7 or later, or Sonnet 5, the request fails with a validation error — replace it with adaptive thinking plus an effort level.
Configuring a thinking budget
On the current "Claude in Amazon Bedrock" surface, the request body has the same shape as Anthropic's first-party Messages API:
from anthropic import AnthropicBedrockMantle
client = AnthropicBedrockMantle(aws_region="us-east-1")
response = client.messages.create(
model="anthropic.claude-haiku-4-5-20251001-v1:0",
max_tokens=8000,
thinking={"type": "enabled", "budget_tokens": 4000},
messages=[{"role": "user",
"content": "Reconcile these two quarterly reports..."}],
)On the legacy InvokeModel surface, the same fields go in the JSON body alongside the mandatory "anthropic_version": "bedrock-2023-05-31", max_tokens, and messages. AWS recommends the Converse API for a unified parameter set across models, but either legacy operation carries the thinking configuration in the request body. Note that Claude Sonnet 5 is not available on the legacy surface at all.
Budgeting and billing rules
Thinking budget tokens are a subset of max_tokens, are billed as output tokens, and count toward your rate limits. So in the example above, up to 4,000 of the 8,000 maximum tokens may be spent reasoning, leaving the remainder for the visible answer — size max_tokens with both phases in mind. At Claude Haiku 4.5's $5 per million output tokens, a fully used 4,000-token budget adds about two cents per request; on Opus-class models it is proportionally more, so reserve large budgets for requests that need them.
Latency needs the same respect. Thinking phases can make responses substantially longer to complete: Bedrock's inference timeout for recent Claude models is 60 minutes, but AWS SDK clients default to a 1-minute read timeout — AWS recommends raising it (for example, botocore read_timeout of 3600) for long generations. Streaming the response is the better user experience for anything interactive.
Operational fine print
Multi-turn conversations with tools: when you post tool results back, return the entire unmodified thinking block, including its signature, that accompanied the tool request. Thinking blocks are verified cryptographically, and modified blocks are rejected. Whether previous thinking blocks are preserved across turns varies by model — recent Opus and Sonnet generations keep them (billed as input tokens); Haiku models strip them automatically.
Guardrails: Amazon Bedrock Guardrails does not evaluate reasoning content blocks. Your content policies apply to the final response text, not the thinking phase — factor that into any safety review.
budget_tokens sized to the hardest requests in the route. On Opus 4.7+ and Sonnet 5, don't fight adaptive thinking — pick an effort level and let the model allocate.Where to go next
For the concept itself, see the extended thinking explainer; for cost planning, budgeting thinking tokens. The Bedrock feature matrix shows where thinking sits among the other capabilities.