Anthropic Python SDK 1.12.0 adds claude-haiku-5-5 and typed computer and browser tool calls

Anthropic's Python SDK v1.12.0 adds support for a claude-haiku-5-5 model identifier and typed tool calls for the computer and browser toolsets, according to the release notes dated 2026-10-07. The release is a changelog entry on GitHub, not a model announcement. It contains no pricing, context window, benchmark scores, parameter count, or training cutoff for the model. Pricing not yet disclosed in the source.

What changed

Model and tool support

• claude-haiku-5-5: the api feature commit adds the model ID alongside typed computer and browser toolset tool calls. The commit message does not describe the model's capabilities.
• Model capabilities: three additions to the capabilities schema. disabled is now a valid value in the thinking types, and web search and code execution support are now represented.
• /v1/models: lifecycle stage fields and a filter are added. Model objects also gain a new line field, which has its own documentation update. The SDK notes do not define the values line can take.
• Managed Agents: the web_fetch tool config gains url_sources. Documentation for the multiagent and thread descriptions was also updated.

Admin and platform API

• RBAC roles gain a display_name field, and the existing name field is deprecated.
• The list workspaces call gains an include_default parameter.
• Docs now state the agent tools limit as 256.

Bug fixes

• Spend limits: listing spend limits now sends the beta header by default.
• Empty strings: the client now sends empty strings in query params and multipart forms. Previously they were evidently not sent.
• Memory tool: a local edit made after a memory upload whose response was lost is now kept (PR #990). This closes a data-loss path in the memory tool's sync logic.

Internal changes

The chores include tracking whether the client timeout was set explicitly and refactoring deprecated models. Several previously skipped tests were re-enabled, including batch results raw and streaming response tests (PR #978). A REVIEW.md with review instructions was added, and docs code blocks were reformatted with ruff. These changes do not affect public behavior.

The full diff is available as v1.11.0...v1.12.0.

What this means

The SDK shows that a Haiku 5.5 identifier exists in Anthropic's API surface, but nothing more. Without a published model card or pricing page, treat the model's availability, tier, cost, and capabilities as unconfirmed. Developers should not hardcode claude-haiku-5-5 into production until Anthropic documents it.

The more durable changes are in the metadata layer. Lifecycle stage fields and a filter on /v1/models let applications discover which models are current, in preview, or deprecated, instead of maintaining hardcoded lists. The capability flags for web search, code execution, and disabled thinking let clients check a model's features before building a request. Both reduce breakage when model lineups change.

Typed computer and browser toolset tool calls matter for agent builders. Typed responses cut down on manual parsing of tool payloads and make type checking possible in agent loops.

The name to display_name deprecation on RBAC roles is the one item that needs migration. Teams that read role names from the admin API should move to display_name. The empty-string fix is a quiet behavior change that could alter request payloads for code that relied on the old behavior.