Services or capabilities described in AWS documentation might vary by Region. To see the differences applicable to the AWS European Sovereign Cloud Region, see the AWS European Sovereign Cloud User Guide.

Preserved thinking

Overview

On Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5, and Claude Haiku 5.5, each thinking block is tied to the model and the conversation that produced it. On Claude Mythos 5.1, thinking blocks are tied to the model only; the conversation prefix isn't checked. The API checks that the system prompt, the tool list, and all messages before the block are unchanged when you replay that block in a later request. This protects the integrity of Claude's reasoning: reasoning produced under one set of instructions cannot be replayed under a different set.

If your harness sends conversation history back exactly as it received it, nothing changes for you. If you edit history between requests — injecting per-turn reminders, summarizing older turns, or rebuilding the system prompt — the API rejects the request by default.

What is checked on replay

Model compatibility

When you switch models partway through a conversation, the model check decides whether earlier thinking blocks reach the new model. Blocks that the new model can't read are dropped for that request; the API doesn't edit your messages, so the blocks are used again if you switch back to a model that can read them.

The block_binding request object (beta)

Add block_binding to the thinking object and include the beta value in anthropic_beta to control what happens when a prefix mismatch is detected:

{
    "anthropic_version": "bedrock-2023-05-31",
    "anthropic_beta": ["thinking-binding-controls-2026-08-01"],
    "max_tokens": 16000,
    "thinking": {
        "type": "adaptive",
        "block_binding": {
            "prefix_mismatch_behavior": "drop_block"
        }
    },
    "messages": [{ "role": "user", "content": "Your prompt here" }]
}

block_binding accepts one field, prefix_mismatch_behavior, which controls what the API does with a thinking block that fails the conversation prefix check. mismatch_behavior is accepted as a deprecated alias for this field. Setting both prefix_mismatch_behavior and mismatch_behavior in the same request returns a 400 invalid_request_error: cannot set both block_binding.prefix_mismatch_behavior and block_binding.mismatch_behavior; send only prefix_mismatch_behavior. Sent without the beta value, block_binding returns 400 thinking.adaptive.block_binding: Extra inputs are not permitted. Malformed values also return a 400 naming the field. On Claude Sonnet 5.5 and Claude Haiku 5.5, block_binding works only with thinking.type: "adaptive".

Controlling mismatch behavior

Neither value changes the model check, which always drops.

The input_transformations response array

When the beta value is sent, the response may include a top-level input_transformations array describing any blocks that were removed:

{
    "type": "message", "role": "assistant", "content": [ ... ],
    "stop_reason": "end_turn",
    "usage": {"input_tokens": 18234, "output_tokens": 911},
    "input_transformations": [
        {"type": "thinking_dropped", "path": "messages.3.content.0", "reason": "prefix_binding_mismatch"}
    ]
}

This is a top-level array (a sibling of usage), present only with the beta value; [] when no blocks were removed and absent otherwise. Each entry contains a path identifying the removed block and a reason: model_binding_mismatch (created by a different model), prefix_binding_mismatch (the conversation prefix changed), organization_binding_mismatch (created by another account or customer group), or end_user_binding_mismatch (created in a Claude app). In streaming responses, the array appears within the message object inside the message_start event. Removed blocks do not count toward input_tokens.

Organization-locked thinking

On Claude Sonnet 5.5 and Claude Haiku 5.5, each thinking block is also bound to the AWS account (or customer group) that produced it, in addition to the model and conversation-prefix checks. A block replayed from a different account, or one minted in a Claude app (claude.ai or Claude Code), is dropped before the model runs; the request still succeeds and the model does not see that earlier reasoning.

By default the drop is silent. With the thinking-binding-controls-2026-08-01 beta, each dropped block is listed in the top-level input_transformations array:

{
    "input_transformations": [
        {"type": "thinking_dropped", "path": "messages.1.content.0", "reason": "organization_binding_mismatch"}
    ]
}

Replaying thinking blocks within the same AWS account (or customer group) is unaffected. Without the thinking-binding-controls-2026-08-01 beta, the response does not include an input_transformations field. In the same-account case with the beta, the field is present as an empty array ([]).

This binding applies on Claude Sonnet 5.5 and Claude Haiku 5.5; earlier models are unaffected.

Error responses

When prefix_mismatch_behavior is "error", a prefix mismatch returns:

{
    "type": "error",
    "error": {
        "type": "invalid_request_error",
        "message": "messages.3.content.0: Invalid `signature` in `thinking` block. The block is bound to a different conversation. Remove the block, or set `thinking.block_binding.prefix_mismatch_behavior` to \"drop_block\". The `system` prompt differs from the one this block was created with."
    }
}

The last sentence names what changed (for example, the system prompt). If the request doesn't include the thinking-binding-controls-2026-08-01 beta value, the message instead ends with a reminder that prefix_mismatch_behavior requires that beta value.

This error is permanent for that request — an automatic retry loop will not clear it. When you catch it, either strip all thinking blocks from history and retry, or retry with prefix_mismatch_behavior: "drop_block" and the beta header.

Guidance for multi-turn and agentic applications