Voice, facts, and fixes: the week AI got more careful

A more reliable GPT-5.6, for everyone

OpenAI's headline move was pairing a meaningful quality upgrade with wider access. GPT-5.6 Sol improvements landed for Plus and Pro users, promising roughly 68% fewer factual errors on internal evals and more focused answers, plus a new effort slider to dial reasoning depth up or down. It's the kind of incremental but visible polish that makes daily use feel less like gambling on a model's mood. The same change is documented at the product level for anyone tracking the release notes.

The bigger structural shift: GPT-5.6 Luna becomes the default for Free and Go users, with unlimited text chats and a "Think" button for harder questions arriving next week — abuse guardrails still apply to file uploads, images, and other tools. That's a genuine expansion of capability to the bottom of the funnel, not just a discount on a paywall.

Voice gets its hands on files and Projects

Across the product surface, voice mode matured from a chat-only channel into something closer to a full workspace. ChatGPT Voice adds file uploads and Projects support, letting you upload images, docs, and PDFs mid-conversation and analyze them aloud, or drop into a Project and reference its recent chats, sources, and instructions. The same capability landed for ChatGPT Business and for Enterprise, Edu, and Healthcare workspaces, where "Live" also becomes the default Voice experience when admins enable it — no more Early Model Access toggle. Voice is now a first-class way to work, not a parlor trick.

Two other productivity notes: the ChatGPT for PowerPoint add-in reached general availability, letting you build and revise decks from a sidebar while keeping slides editable in PowerPoint itself — a meaningful step toward agentic work inside the Microsoft Office workflow. And Sign in with ChatGPT began rolling out across Airtable, GitLab, HubSpot, Notion, Supabase, and Vercel, which is OpenAI quietly building an identity layer for the connective tissue of the modern workplace.

Claude Code hardens its sandbox

While OpenAI pushed capability forward, Anthropic spent the week making Claude Code harder to break out of. The back-to-back releases read like a security audit in real time. Claude Code v2.1.221 fixed a Bash tool permission-check bypass where zsh could slip hidden commands through regex conditionals, added a Focus view to collapse tool activity, and introduced sandbox credential file masking on Linux and WSL. The very next day, v2.1.222 closed a worktree-isolation gap where destructive git commands could hit the main checkout, plus a PreToolUse auto-allow hook bypass in background tasks.

Then v2.1.223 fixed another Bash permission bypass and pulled 1M-context models back to a 200K auto-compaction window, while v2.1.224 added self-hosted environments via claude self-hosted-runner and cross-session messaging with SendMessage. The pattern — bypass, fix, new bypass, fix — is exactly what you'd expect from a tool that's becoming a trusted agent running arbitrary commands on developer machines. The v2.1.225 OAuth token loss fix and the reliability-focused v2.1.226 round out the week's churn. Security patches aren't glamorous, but for anyone running Claude Code unattended, they're the most important releases here.

Supporting movement across the rest of the stack

The lower-key releases still tell a coherent story of incremental polish. o1 and o3-mini gained Python-powered data analysis in ChatGPT, Codex CLI 0.147.0 added portable Agent Plugins and persistent conversation sections, and the Node SDK v7.4.0, Python SDK v2.53.0, and Go SDK v3.50.0 all picked up gpt-5.5 support and tool metadata fields — the plumbing for the model updates above. Enterprise & Edu admin controls for desktop app updates and education plugins for teachers and students round out a week where the big players expanded both reach and guardrails.