AI Coding News: August 1, 2026 — Codex Retires GPT-5.4, Copilot Adds Team Model Targeting
August opens with a cleanup notice rather than a new model launch. OpenAI has set a hard retirement date for GPT-5.4 and GPT-5.4 mini inside Codex, GitHub is rolling out a more granular way for enterprises to hand different Copilot models to different teams, and Cursor's new iPad app quietly became a serious code-review surface over the weekend. None of it is flashy, but all three changes are the kind that break a pinned config or a stale bookmark if nobody reads the notice. Here's what changed.
OpenAI Retires GPT-5.4 and GPT-5.4 Mini in Codex
OpenAI has confirmed that GPT-5.4 and GPT-5.4 mini will stop being available inside Codex for users signed in with a ChatGPT account on August 31, 2026. Anything still pointed at gpt-5.4 needs to move to gpt-5.6-terra, and gpt-5.4-mini callers should switch to gpt-5.6-luna — the two models OpenAI positions as direct successors following July's GPT-5.6 Sol/Terra/Luna rollout. The cutover is scoped to ChatGPT-authenticated Codex sessions; direct API access and Codex usage billed through an OpenAI API key are unaffected, so teams routing through the API rather than a ChatGPT seat have more runway.
There's already a preview of what the migration buys: OpenAI says it upgraded Auto-review in the ChatGPT app and Codex CLI from GPT-5.4 to GPT-5.6 Luna, making automated review roughly ten times cheaper to run at the same quality bar. For anyone with a saved prompt, CI script, or AGENTS.md convention that hardcodes GPT-5.4, the one-month window is the actionable part — a model swap that fails silently in a pinned config is a worse Monday than one that fails loudly today.
GitHub Copilot: Enterprise Teams Get Model Policy Targeting
GitHub shipped a new public preview on July 31 that lets enterprise admins move past a single enterprise-wide model switch. Enterprise teams model policy targeting lets an admin set a baseline set of Copilot models for the whole enterprise, then grant additional models to specific enterprise teams — giving a security team access to a model that a marketing-adjacent team doesn't need, for instance, without a blanket policy change.
Each model gets one of three states: Enabled for every member, Disabled for everyone, or Optional, which makes it assignable to individual teams. GitHub uses a least-restrictive evaluation, so a user who gets a model through membership in any enterprise team has it everywhere they work. Most enterprise customers get preview access starting August 3, and GitHub recommends creating enterprise teams and deciding which optional models to assign before flipping the feature on, since rollback is only guaranteed during the preview window. It's a direct follow-up to the July 29 default-model-enablement policy — that change decides what's on by default across an entire enterprise; this one decides who inside that enterprise can reach further.
Cursor for iPad Ships Full PR Review
Cursor's new iPad app, which went live on all paid plans on July 29, turns out to be more than a scaled-down client. The review screen covers an entire pull request — comments, CI checks, and approvals — so reviewers can reassign reviewers, read and resolve comment threads, and prompt the agent to act on feedback without dropping to a laptop. The layout takes advantage of the bigger screen with pinned sidebar chats for watching several agents at once and a split view that keeps a review open next to a live chat with full-size diffs.
The Apple Pencil integration is the detail worth calling out: reviewers can attach a screenshot, drop a comment at an exact point, or draw directly on the image, which maps markup-style feedback onto a code review in a way a phone screen can't really support. Combined with the inbox view Cursor added for both iPhone and iPad, it's a sign that mobile and tablet clients for these tools are moving from "check status" utilities toward places real review work happens.
The Takeaway
Nothing here is a new model or a new agent — it's three different teams tightening the screws on tools already in daily use. OpenAI is giving Codex users a firm deadline instead of a slow fade, GitHub is letting enterprises get more selective about who gets which model instead of forcing an all-or-nothing switch, and Cursor is betting that a tablet with a pencil is a legitimate place to review a pull request. If you run Codex through a ChatGPT seat, check for hardcoded GPT-5.4 references before August 31 — that's the one item on this list with a real deadline attached.
References & Further Reading
Comments
Share your thoughts and join the conversation