OpenResponses 2.9 release notes

Submitted October 1, 2026, and released October 2 as build 54. The App Store What’s New text is release_notes.txt, and the promotional text is promotional_text.txt. Every change was found while filming the app’s demo videos on 2.8; the details and their tests are in CHANGELOG.md under 2.9.

Tool steps

What 2.8 showed	Cause	2.9
Every step stayed on Queued	Rows closed only on response.output_item.completed, which the Responses API never sends; it ends items with response.output_item.done (openai-python’s stream event types, read October 1, 2026)	output_item.done closes hosted tool rows; in_progress, searching, interpreting and generating events show Running
A failed browser step showed Completed	Only outputs starting Error: counted, and browser failures read Error processing …	An output that starts with “Error” fails the step (ChatViewModel.functionOutputFailed)
MCP steps went from Queued to Completed, and a tool error was lost	response.mcp_call.in_progress names no tool; mcp_tool_execution_error has content, no message, so the item didn’t decode	in_progress shows Running; MCPToolError reads content, or a bare string
code_interpreter_call	Hosted tool items have no name	Rows read Code Interpreter, Web Search, File Search, Image Generation, and a Code Interpreter step shows its code

Text

• On the streaming path taken with Computer Use and with models older than GPT-5.6 or specialized ones, a new text item now starts a new paragraph (paragraphBreak, lastTextItemIds).
• The failure note that stands in for an answer clears when the answer starts. A tool loop completes a response every round, and each completion forgot the notes, so a round of only tool calls left the note above the answer.
• Reasoning summaries render inline Markdown (InlineMarkdown).

Charts and files

• appendArtifact keeps one entry per file id, preferring a copy that loaded and a real file name over the bare id.
• Code Interpreter keeps a displayed plot as a file named by its own id. When the code also saved the figure, the reply cited one chart twice; a plot that matches a saved file from the same container (same shape, and under 3% of pixels differing strongly at 32 by 32) shows once, as the saved file.
• ChatMessage.listedArtifacts leaves a picture out of Generated Files while images holds that very picture, so a browser screenshot that replaces images brings a chart back to the list.
• Saved text and error artifacts keep their content when a conversation is reopened.

Verification

Full unit suite (the set CI runs) on simulator “OpenResponses tests” 73D9ECD6: 343 executed, 0 failures, 3 skipped (the live docs-page test without its variable, and two search-by-meaning tests that need Apple’s on-device model, which a new simulator doesn’t have yet). Two reviewer passes over the diff; their findings are fixed and covered by tests.