Health intelligence, agent safety, and enterprise controls

Health intelligence reaches free users

The week's most impactful release was GPT-5.5 Instant's improved health intelligence, now available to all free ChatGPT users. OpenAI reports that production factuality issues in health responses have fallen by 71% over the past two months, and physician-rated responses now surpass both older models and human-written responses across accuracy, communication, and completeness criteria. With over 230 million weekly health and wellness queries on ChatGPT, this means a substantial cohort of users now receives materially better answers for interpreting lab results, understanding insurance, or preparing for appointments — no subscription required. The release signals a broader shift: health is becoming a primary evaluation frontier for general-purpose assistants, not just specialized medical AI.

Agent safety and hardening

Claude Code continued its aggressive safety push with v2.1.183's destructive git command blocking. Commands like git reset --hard, git checkout -- ., git clean -fd, and git stash drop are now blocked unless the user explicitly asked to discard local work, and terraform destroy/pulumi destroy/cdk destroy are similarly gated. This is a direct response to one of the most common failure modes in AI coding agents: a well-intentioned but catastrophic command that obliterates hours of work. The same release fixed fullscreen TUI corruption on Windows Terminal under heavy subagent load — a quality-of-life fix that affects anyone running complex multi-agent workflows on Windows.

The safety tightening extended to Claude Code v2.1.178, which introduced wildcard parameter matching for tool permission rules — you can now write Agent(model:opus) to block Opus subagents, for example — and fixed an OAuth credential race that could crash the process via stale file descriptors. Meanwhile, Codex CLI 0.142.0 added configurable token budgets that track usage across agent threads and abort turns when exhausted, plus an indexed web-search mode that restricts direct page access to server-approved URLs, giving teams control over what the agent can read.

Enterprise admin tooling matures

Both labs shipped meaningful enterprise controls this week. OpenAI launched credit usage analytics and granular spend controls for ChatGPT Enterprise, letting admins view usage trends across users, products, and models in the Global Admin Console, and set default, group-specific, or individual spend limits. On the SDK side, both the Python and Go SDKs added admin spend alerts API support. Anthropic launched enterprise-managed auth for MCP connectors, allowing admins to authorize and authenticate connectors across the whole organization through their identity provider — no individual setup required.

Two more channels for Slack integration arrived: ChatGPT Business and Enterprise & Edu workspaces can now use Slack connector actions — joining channels, creating reminders, uploading files — with OAuth scope controls managed by workspace admins. Record & Replay also launched as a Codex feature for macOS Business users, letting teams demonstrate a workflow once and turn it into a reusable skill (initially outside the EU, UK, and Switzerland).

Platform updates and the ecosystem

Anthropic's SDK support for code execution tool code_execution_20260120 rolled out across Python, TypeScript, Go, Java, Ruby, PHP, and C# — no beta header required — available on Claude Fable 5 through Sonnet 4.5+. Claude Code Artifacts entered beta for Team and Enterprise plans, turning raw output into shareable visual pages. Codex 0.141.0 encrypted remote executors via Noise relay channels and pinned bundled SQLite to fix a WAL-reset corruption bug. And the MCP tunnels management API migration moved from the Admin API to the Claude API with a new beta header and workspace-level IAM scope — a breaking change with a migration window for existing users.