GPT-5.5 retires, Claude Code retunes permissions, and the SDKs rebuild for realtime

A model generation turns over

The clearest story of the week is a retirement. OpenAI is retiring GPT-5.5 from ChatGPT, ChatGPT Work, and Codex — API access is unaffected, but anything built on the chat surfaces has to move. For Codex specifically, GPT-5.5 retired for users signed in with ChatGPT at noon PDT on October 14, while API-key-authenticated sessions kept working. The practical fallout is configuration drift: workspace defaults, saved model settings, managed configurations, custom agents, scheduled tasks, and scripts that still name gpt-5.5 need to be pointed at GPT-6.1 Sol. If you administer a workspace, this is the week's one genuinely urgent chore.

Claude Code retunes defaults and mod plumbing

Claude Code shipped a run of updates, and the headline is a default change: Claude Code's default permission mode on third-party providers now applies to claude -p and Agent SDK sessions with no explicit permission mode when telemetry is off or on a fresh install, and fork subagents are on by default across the CLI, Agent SDK, desktop app, and VS Code. The same update stopped the Read tool from loading image files of any size into memory — sources over 128 MB are now refused — and added a row in /config to toggle idle compaction off.

The rest of the week was about giving mod authors and terminal users more control. Claude Code added vim mode's % motion and multi-line input for mods, plus an environment variable that bounds unattended retry when a usage limit resets too far in the future; it also fixed a WorktreeCreate hook's worktree being removed on quit without asking, and sign-in refresh timeouts that caused unexpected sign-outs. Earlier, Claude Code brought /advisor to AWS and OpenTelemetry GenAI span attributes to trace spans, alongside mod authoring features like virtualText prompt decorations and a prompt.change hook, and more than 35 fixes across subagent tracking, worktree handling in reftable repositories, and resume/compaction message loss.

On the platform side, the Compliance API now returns the instructions members saved for Claude, and fast mode is generally available for Opus 5.5 — the beta header is no longer required, a default rate limit applies to Start, Build, and Scale tiers, and fast mode reaches Claude in Microsoft Foundry.

ChatGPT grows collaboration surfaces

ChatGPT's week was mostly new surfaces to work in. ChatGPT Slides lets you build a deck from a conversation, files, or an outline, edit it directly or by asking, and collaborate in real time with shared comments — then present from ChatGPT or export to PowerPoint, PDF, or Google Slides. GPT-6 Pro picked up parallel agents that split a task across different parts of the work with visible progress, plus faster responses and better plugin and tool use. Voice mode got GPT-Live-1.1 and a broader set of Live voices covering French, German, Hindi, Japanese, Korean, and Mexican Spanish, and Sites lost its fixed cap on how many you can create.

For administrators, ChatGPT Enterprise retired admin-managed sync connectors in favor of real-time Google Drive, SharePoint, and Teams apps, with new content controls under Safety & Security Controls. The Admin Console consolidated External access controls into one page for organization defaults and per-app permissions, and both Enterprise and Business admins can now manage up to 25 plugins at once — the same bulk plugin management arrived for Business workspaces. Enterprise also began a Sites admin preview with data residency and tunnels, and OpenAI opened Rosalind Academic Plan applications for US university life-sciences labs.

The SDKs rebuild around realtime

Across the Python, Node, and Go SDKs, the same feature wave landed in near-lockstep, and realtime is where the engineering went. All three added opt-in bounded connection recovery for Realtime — WebSocket session recovery in Go, with Python fixing beta Realtime byte reads on websockets 13 — alongside typed Live sideband events and an ultrafast service tier for Agents, plus Realtime Translation WebSocket sessions in Python.

The same releases carried a long list of shared API additions: prompt-based voice creation and cache settings, regional processing residency options, safety identifiers on image requests, browser and computer toolsets, and beta session spending limits. Go's parallel run brought Decisions delegation, triggers, and Live events, typed Live sideband events and beta connections, prompt-based voice creation and residency options, and updatable project API key expiry, while organization Decisions usage reporting and a degraded external storage status rounded out the shared surface. The API itself gained native Decisions integration in GPT-Live, letting apps act while the user is still speaking.

Elsewhere, Codex CLI added a searchable /config panel and immediate steering, so a follow-up sent mid-response redirects the turn and interrupts stop active Code Mode cells.