What changed in Claude Code, Codex, Meta, and Grok? Four updates worth a look

A few recent AI updates raise familiar questions. What happens when code review never seems to finish? Does changing your default model also update older tasks? Can an assistant tell when something it remembers is out of date?

Here is one item each from Claude Code, Codex, Meta, and Grok. The announcements and articles date from October 12–15; we checked the sources again on October 23 before publication.

On October 15, Anthropic published how the Claude Code team does code review. One PR went through more than fifty review-and-fix rounds. The team subsequently capped rounds, narrowed later reviews to new changes, and stopped repeating findings.

That is a recognizable trap. Another review seems inexpensive, but each small suggestion can produce another edit and another round. The task keeps growing even when the original problem is already addressed.

Ask each finding to name a trigger, consequence, and location. After a fix, check that finding and the new changes. If several rounds do not converge, revisit the requirement and approach.

For example, a fix for an empty list crashing a page may prompt a suggestion to rewrite the whole list component. Decide whether that belongs in this task. Resolved bugs and newly introduced problems matter more than the number of comments.

The OpenAI changelog says GPT-5.5 retired from Codex for ChatGPT sign-ins on October 14. API-key-authenticated Codex sessions and the OpenAI API are unaffected. OpenAI recommends GPT-6.1 Sol where the plan and client support it.

Changing the model in a chat is straightforward. Older saved configurations are easier to overlook: workspace defaults, custom agents, scheduled tasks, and scripts may still name the retired model. A working current conversation does not tell you whether those settings have changed.

If you pinned GPT-5.5, search your configurations for it. Check whether each connection uses ChatGPT sign-in or an API key before deciding what needs updating. Ask an administrator to handle centrally managed settings.

After a change, run a familiar small task. Check that it starts, selects the intended model, and can use the files and tools it needs. Try a scheduled task manually rather than discovering an old setting when its output is due.

Meta announced Muse Wonder and Muse Wonder Full on October 12, emphasizing cross-conversation Active Memory and persistence through Active Focus. Wonder is offered through the Meta Model API and Muse Code; Full is offered to Muse Code subscribers. The Muse app rollout is separate.

We would start by checking what happens when a remembered fact expires. Can a newer requirement replace an older one? Does a fact belonging to one project stay out of another?

For a practice event, give an assistant a budget of 300, explicitly change it to 200 in the next conversation, then ask for a plan. Check which budget it uses and whether the old information can be found and corrected. That tells you more than asking whether it remembers you.

Developers should also inspect each access path separately. A model API, a coding product, and a personal assistant can offer different state management. A memory feature in one product does not establish that every API using the same model preserves your long-term history.

Grok Voice 3.0, announced October 14, is available through the Speech to Speech API. The company reports latency below one second and says agents using grok-voice-latest were switched automatically.

A smooth demo is appealing. In use, you also need to know whether the assistant notices a correction, handles a mistaken number, and drops an old instruction when you interrupt.

Try a sample request: ‘Remind me about the meeting at three on Friday—actually, make that four.’ Repeat it with background noise or interrupt the answer. Check the final time. If a calendar tool is involved, inspect the test calendar too: saying the right thing and writing the right record are separate checks.

Teams using a moving model alias such as latest should keep familiar test conversations. After an automatic switch, rerun corrections, interruptions, and tool-result checks. An old successful demo cannot indefinitely represent the current version.

If you use AI for code, watch for reviews that keep expanding the task. If you pinned a Codex model, check older task settings. If you work across conversations, correct a saved fact. For voice applications, test interruptions and changed instructions.

There is no need to rebuild your workflow for every announcement. Choose one small task you already do and note what became easier and what still needed fixing. That gives you something useful to compare next time.