Codex expands beyond coding while Claude and OpenAI harden security and cost controls

Codex becomes a platform for knowledge work

OpenAI made its biggest play yet to turn Codex from a developer tool into a general-purpose work platform. Six new role-specific plugins bundle 62 apps and 110 skills for data analytics, creative production, sales, product design, public equity investing, and investment banking — available in preview for Business and Enterprise customers. More notable is the launch of interactive Sites, which lets Codex create and share live websites and apps via URL directly from a conversation. Annotations now work on documents, spreadsheets, and slides, allowing in-place refinement without leaving the chat. The CLI kept pace with enterprise features: monthly credit limits, cloud-managed config bundles, and remote-control pairing for support workflows. A separate CLI update added Desktop handoff on macOS and Windows, plus file-path exposure for local image attachments so follow-up edits are more reliable. Activity insights and shareable profile cards arrived in the Codex app, along with improved Computer Use startup readiness.

Refusal billing and advisor cost controls on Claude

Anthropic quietly shipped a pair of API changes that matter for production budgets. The Claude API no longer charges for requests that return stop_reason: "refusal" without any generated output — a meaningful cost saver for workloads with aggressive guardrails. Separately, the advisor tool now supports a max_tokens parameter to cap the advisor model's output per call, reducing both latency and token spend on tasks that only need short responses. Both changes are live on the API.

Hardening sessions, permissions, and platforms

Security work dominated the week. Claude Code v2.1.160 adds prompts before writing to shell startup files (.zshenv, .zlogin, .bash_login) and ~/.config/git/, and blocks acceptEdits from writing build-tool config files that grant code execution. v2.1.161 fixed a managed-settings auth bypass where forceLoginOrgUUID policies were blocking third-party provider sessions alongside the org pin. v2.1.162 fixed WebFetch permission rules not applying to preapproved domains and Windows permission-rule handling with backslashes, while v2.1.166 hardened cross-session messaging so relayed messages no longer carry user authority — plus added a fallbackModel setting and retry logic for API failures. Across all these versions, Anthropic fixed dozens of smaller bugs: background session data loss, startup hangs on read-only config directories, and MCP timeout behavior.

OpenAI matched the security momentum. ChatGPT Lockdown Mode is now available to all logged-in users as an opt-in setting, restricting live web browsing, deep research, agent mode, and file downloads to reduce prompt-injection exfiltration risk. For Enterprise and Edu workspaces, admins gained granular App permissions — "Always ask," "Any changes," or "Important actions" — settable per connected app. Plugin sharing in Codex is now available by default for eligible Enterprise workspaces, with an MDM off-switch for admins.

Model updates and platform changes

Anthropic deprecated Claude Opus 4.1 with retirement on the API scheduled for August 5, 2026, recommending migration to Opus 4.8. Enterprise Claude plans gained custom admin roles for billing and privacy access without full ownership. OpenAI introduced GPT-Rosalind, a model update combining GPT-5.5's agentic coding with stronger drug-discovery reasoning, alongside the new LifeSciBench benchmark. ChatGPT memory now updates automatically to reduce stale context, and Plus/Pro users get twice the memory capacity. The SDKs all gained moderation endpoints: Python v2.41.0, Node v6.42.0, and Go v3.39.0 all added responses.moderation and chat_completions.moderation support. Smaller bug-fix releases shipped for the ChatGPT mobile apps and Claude desktop clients.