OpenAI tweaks GPT-5.5 Instant; Codex Goal Mode goes GA

OpenAI published release notes on May 29, 2026 that bundle changes across ChatGPT and Codex, including a GPT‑5.5 Instant style-and-quality update and broader Codex tooling improvements. The notes cover product changes, infrastructure work, and a slate of model retirements.

The ChatGPT changelog says GPT‑5.5 Instant received a targeted update to improve response style and quality, producing answers that are easier to read and better paced for practical tasks. OpenAI also removed canvas support from GPT‑5.5 models and moved writing and coding features into chat blocks.

OpenAI’s product post for GPT‑5.5 Instant lists concrete internal gains: a reported 52.5% drop in hallucinated claims on internal “high‑stakes” prompts and reductions in response length and line count versus GPT‑5.3 Instant. The company frames these as gains in factuality, concision, and utility.

The longer changelog in the Help Center clarifies timing and rollout: GPT‑5.5 Instant began rolling out earlier in May and now serves as ChatGPT’s default; legacy instant models remain temporarily available to paid users before sunset. The update is reflected both in the ChatGPT UI and the API (chat‑latest).

Practically, OpenAI says GPT‑5.5 Instant uses fewer words and fewer formatted lines on average — the company’s example reports roughly 30% fewer words and 29% fewer lines in some sample comparisons — and avoids overly long, bullet-heavy outputs that users have complained about. Those style changes aim to make answers quicker to scan while keeping needed detail.

On the Codex side, OpenAI’s May updates moved several experimental features to production. Help center notes dated May 21 and May 29 list Appshots, improved in‑app browser annotations, locked/remote computer use, and Goal Mode becoming generally available across the Codex app, IDE extension, and CLI.

Goal Mode is the most consequential change for developer workflows: it lets users define an outcome and success criteria and have Codex continue working toward that objective across sessions, potentially for hours or days. OpenAI and multiple developer writeups say Goal Mode now runs by default and is exposed across Codex surfaces.

Several community posts and tooling trackers describe Goal Mode as a form of persistent objective tracking with token budgets and autonomous continuation; users report best results when they give explicit, testable completion criteria such as “until this test suite is green” or “until grep returns zero matches.” That usage pattern helps Codex decide when a goal is truly done.

The release notes also called out desktop features that make long‑horizon agent work easier: Appshots on macOS provides a live window into an app to speed context capture, and Windows support for Computer Use lets eligible users remote‑control host machines from mobile Codex or ChatGPT clients. OpenAI notes regional limits on some Computer Use features at launch.

OpenAI used the same entries to continue a sweep of older models out of ChatGPT: the Help Center lists sunset dates for OpenAI o3 and GPT‑4.5 from ChatGPT in the weeks ahead, and earlier entries show prior retirements of GPT‑4o and other legacy models. Those deprecations apply to the ChatGPT UI and model picker; API changes are separate.

Developers and operators should also note infrastructure work mentioned in the May notes: OpenAI reports responsiveness, in‑app browser speed, and stability improvements intended to back longer agent runs and richer Codex interactions. Open-source trackers and the Codex CLI changelog show rapid CLI releases in mid‑May that brought Goal Mode into the cli and app packages, though some GitHub issues indicate edge cases remain where model‑exposed goal management tooling is incomplete.