Claude Preserved Thinking Errors: Which Earlier Context Must Stay Unchanged?

Identify which earlier context changed

Anthropic’s current help article describes a thinking block as an encrypted record of reasoning Claude may produce while answering. Multi-turn API calls send those blocks back with the system prompt, tools and earlier messages. Preserved thinking verifies that the context which produced a block still matches.

For affected models and accounts, changing the system prompt, tools or messages preceding that block causes an API error. The location of the edit matters: rewriting old history, compacting earlier turns or replacing tools mid-session cannot be treated as an edit that leaves existing blocks unaffected.

For example, suppose the first turn used tool set A to produce a thinking block. If the second turn directly replaces those earlier tool definitions with B while returning the original block, check this validation boundary. This is an illustrative scenario, not an API reproduction performed against a real account for this article.

Check the model and account-creation cutoff

Preserved thinking applies to all API accounts on Claude Fable 5.2 and Claude Mythos 5.2, with no older-account exemption. For Claude Fable 5.1, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5, the help article applies the rule to API accounts created on or after August 31, 2026 at 00:00 UTC. Earlier accounts are not subject to this preceding-context check on those models.

The scope covers Claude Platform organizations, Amazon Bedrock accounts, Google Cloud Vertex AI projects and Microsoft Foundry projects. Check the relevant account or project on the actual platform rather than substituting a personal subscription date. Language or location does not establish an exemption.

This applicability rule does not grant model access or change Mythos’s separate admission requirements. The help page only says “Updated this week”; it does not establish an exact publication or per-account enforcement time that this article can confirm. This is a description of current guidance, not a claim that a uniform rollout began today.

Preserve the original context before planning changes

Start by comparing the failing request locally with the original request that produced the relevant block. Check whether middleware, injected reminders or compaction logic changed the system prompt, tool definitions or preceding messages. Comparing only the final user question can miss the change. Keep real conversations, credentials and encrypted blocks out of public logs and tickets.

Anthropic points to three adjustment resources: preserved thinking documentation, Compaction (beta), and mid-conversation tool changes (beta). Compaction can summarize older turns while preserving recent turns word for word and can build the summary in the background. The tool-change mechanism adds or removes tools without editing earlier turns. Each requires checking current support conditions; neither is established as available for every platform, model and account.

When validating an integration change, cover unchanged continuation, earlier-history compaction, injected reminders, and adding or removing tools separately. Record actual errors, responses and usage. This is a proposed check sequence, not a runtime or performance acceptance test completed by this article.

Treat dropping thinking as a migration stopgap

For requests that need to modify earlier context, Anthropic describes an opt-in that drops affected thinking blocks so the request can proceed. The model does not see those blocks, and the response identifies which ones were dropped. A successful request therefore does not establish that the original reasoning context was preserved.

Anthropic warns that this can significantly increase output tokens and recommends it only as a stopgap while updating an integration. The reviewed help article provides no fixed increase, uniform charge, or copyable header and setting syntax. Check current technical documentation for the actual parameters and cloud-platform availability rather than putting a guessed switch into production.

Separate cross-account dropping from context errors

A separate rule applies to Claude Mythos 5.2, Claude Fable 5.2, Claude Sonnet 5.5 and Claude Haiku 5.5: thinking can be read only by the account that produced it or a linked account. Do not merge this model list or account-linking condition with the earlier-context rule governed by the creation cutoff.

If a request includes thinking from an unrelated account, the API drops it and continues without returning an error. The next response may be slower and use more tokens, similarly to continuation after compaction. The absence of an error does not establish that thinking context remained intact.

The help article explicitly lists automatic linking within the same Claude Platform parent organization, AWS organization or Google Cloud organization. This does not establish automatic linking across unrelated organizations or all Microsoft Foundry projects; verify the current platform-specific rules.

Use the response to separate compatibility from connectivity

For a custom API integration, check the platform and model, the relevant account or project creation time, whether preceding context changed, and whether the response reports dropped thinking. After restoring compatibility, compare output tokens, response time and task results. No fixed cost or speed change was measured for this article.

If the response identifies a context-validation problem, retrying the same request or changing proxy egress does not change the history already sent. Investigate networking separately when there is actual connection, DNS, TLS or proxy-authentication evidence. A proxy also does not change model eligibility or account-linking rules.

Anthropic says ordinary Claude Code, Claude Cowork and Claude.ai usage already handles thinking blocks, so users need not modify an integration for this change. Clients and agents that wrap the Messages API themselves should inspect their own history, compaction and tool-management logic.