# GLM-5.2 with a 1M window, Codex Record and Replay

> Z.ai releases GLM-5.2, OpenAI teaches Codex to learn from demonstrated workflows, and the research list tilts toward looped coding models.

Published 13 Oct 2026. Coverage: 15 Jun 2026 to 21 Jun 2026. By Ali Asif.

## Items

1. **GLM-5.2** (Z.ai, Open weights): Z.ai's notes date GLM-5.2 to June 16, with a 1M-token context and better long-horizon coding. [Primary source](https://docs.z.ai/release-notes/new-released)
2. **Record and Replay in the Codex app** (OpenAI, Agents and coding): June 18 lets Codex on macOS turn a workflow you demonstrate into a reusable skill, with handoff between local and remote hosts. [Primary source](https://developers.openai.com/codex/extend/record-and-replay)
3. **Claude Platform release notes, June 15 to 18** (Anthropic, Agents and coding): Sonnet 4 and Opus 4 were retired June 15, and the SDKs added a code execution tool that keeps REPL state between calls. [Primary source](https://platform.claude.com/docs/en/release-notes/overview)
4. **JoyAI-VL-Interaction** (arXiv, Science): Real-time vision-language interaction, first on the weekly papers list. [Primary source](https://arxiv.org/abs/2606.14777)
5. **Learning from the Self-future: On-policy Self-distillation for dLLMs** (arXiv, Science): A distillation method for diffusion language models. [Primary source](https://arxiv.org/abs/2606.18195)
6. **LoopCoder-v2** (arXiv, Agents and coding): A coding model that loops once to scale test-time compute efficiently. [Primary source](https://arxiv.org/abs/2606.18023)

A quieter week after Fable 5, with a few things worth keeping.

## GLM-5.2

Z.ai's release notes date GLM-5.2 to June 16. It supports a 1M-token context and improves long-horizon and coding performance. Press coverage put the first access a few days earlier, and described the model as roughly 753B parameters with about 40B active under an MIT license. I could not check the model card in detail, so treat the size and benchmark figures going around as secondhand until Z.ai's own page says the same. The lineage is clear enough: GLM-5.1 was the eight-hour worker in April, and this one adds a longer memory.

## Record and Replay

OpenAI's June 18 Codex update added Record and Replay on macOS. You demonstrate a workflow, and Codex turns it into a reusable skill. The same release adds thread handoff between local and remote hosts. Two days earlier, OpenAI rolled more Codex features into the EEA, UK and Switzerland, including computer use, the Chrome extension and memories.

I think the interesting move here is the authoring model. Skills used to be written. Now they can be shown.

## Anthropic housekeeping

Anthropic's platform notes show Claude Sonnet 4 and Claude Opus 4 retired on June 15, and the SDKs picking up a code execution tool version on June 18 that keeps REPL state across calls. It is a small change that makes multi-step analysis in a sandbox much less awkward. The retirements had been announced on April 14, so anyone pinned to those model IDs had two months of notice. If you run production agents on pinned versions, that calendar is worth putting in a reminder.

## Papers

JoyAI-VL-Interaction, on real-time vision-language interaction, topped Hugging Face's weekly list. Two more are worth a look. One is a self-distillation method for diffusion language models, which are still a minority approach next to autoregressive models. The other is LoopCoder-v2, which loops once to cut the cost of test-time compute. Neither changes what I do on Monday, but diffusion language models are the research thread I would keep half an eye on, because if they get fast and good enough the serving math for long documents changes.

What I'd watch next: Gemini 3.5 Pro, which Google said would arrive in June. With nine days left in the month, nothing has appeared yet.

## My take

Record and Replay is the item I would test first. Most of my back-office work is repeated clicking through portals: supplier registration, quote downloads, vendor forms. If a skill can be made by doing the task once on screen, the cost of automating a ten-minute chore drops to almost nothing. I would keep a human on every step that sends something outside the company.

## Sources

- https://docs.z.ai/release-notes/new-released
- https://developers.openai.com/codex/extend/record-and-replay
- https://platform.claude.com/docs/en/release-notes/overview
- https://arxiv.org/abs/2606.14777
- https://arxiv.org/abs/2606.18195
- https://arxiv.org/abs/2606.18023

Canonical: https://heyaliasif.com/ai/2026-06-15-week-in-ai