# GPT-5-Codex, Grok 4 Fast, and Anthropic explains its bad month

> Week of 15 Sept 2025: GPT-5-Codex, Grok 4 Fast, Anthropic's postmortem, Magistral 1.2, Tongyi DeepResearch, Wan2.2-Animate.

Published 13 Oct 2026. Coverage: 15 Sept 2025 to 21 Sept 2025. By Ali Asif.

## Items

1. **Codex upgrades with GPT-5-Codex** (OpenAI, Agents and coding): The Codex CLI repo is where the GPT-5-Codex model, tuned for long agentic coding runs, is wired into the terminal. [Primary source](https://github.com/openai/codex)
2. **Grok 4 Fast** (xAI, Frontier): A cheaper Grok 4 variant with a 2M-token context window that merges reasoning and non-reasoning in one model. [Primary source](https://x.ai/news/grok-4-fast)
3. **Magistral Small 1.2** (Mistral AI, Open weights): The open reasoning model gains vision input and better tone in a September refresh. [Primary source](https://huggingface.co/mistralai/Magistral-Small-2509)
4. **Tongyi DeepResearch 30B-A3B** (Alibaba Tongyi Lab, Open weights): An open web-research agent model with 3B active parameters, trained for long multi-step searches. [Primary source](https://huggingface.co/Alibaba-NLP/Tongyi-DeepResearch-30B-A3B)
5. **Wan2.2-Animate 14B** (Alibaba Wan, Image, video, world): Animates a character from a reference video, or swaps a person in a clip, with open weights. [Primary source](https://huggingface.co/Wan-AI/Wan2.2-Animate-14B)
6. **A postmortem of three recent issues** (Anthropic, Policy and business): Anthropic traces weeks of degraded Claude answers to three infrastructure bugs and says how it will catch them. [Primary source](https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues)
7. **Anthropic Economic Index, September report** (Anthropic, Policy and business): Maps how Claude is used across countries and tasks, and where usage leans toward full automation. [Primary source](https://www.anthropic.com/research/anthropic-economic-index-september-2025-report)

A pile of coding and agent news this week, but the piece I keep returning to is a postmortem rather than a launch.

## Anthropic owns up

Anthropic published an engineering write-up on three infrastructure bugs that degraded Claude's responses between August and early September. One was a routing error that sent some requests to servers set up for a longer context window. Another corrupted output on some hardware. A third came from a compiler issue that affected how the model picked its top tokens. Users had been complaining that Claude felt worse, and the company says it didn't reduce quality on purpose.

I like this for a boring reason. When you run agents inside a company, the model is a supplier, and suppliers fail. What I want from a supplier is a clear account of what broke and what they changed. I also take the lesson for myself: I now keep a small set of fixed test prompts and rerun them, so a silent regression shows up in my own numbers first.

## The coding race

OpenAI released GPT-5-Codex, a version of GPT-5 tuned for agentic coding, delivered through the Codex CLI and the cloud product. The pitch is that it can work independently for long stretches on bigger tasks. I link the Codex repository because OpenAI's own announcement page blocked my link check.

xAI shipped Grok 4 Fast, a cheaper model with a 2M-token context window that folds reasoning and quick answers into a single set of weights. Whether it's good enough is a separate question, but a 2M window at a low price is tempting for tender-document work.

## Open models worth a look

Mistral refreshed its Magistral reasoning models with vision. Alibaba released Tongyi DeepResearch, an open web-research agent with only 3B active parameters, and Wan2.2-Animate for character animation and replacement in video.

Anthropic also published its Economic Index report on how Claude is used across countries. It's a useful read if you want evidence of where automation is actually landing, rather than another prediction. I'd pair it with the postmortem: one tells you what customers do with the model, the other tells you what the supplier does when the model slips.

## What I'd watch next

Alibaba's annual cloud event is coming up. Qwen has been shipping weekly, and I'd expect a big multimodal release.

## My take

The Anthropic postmortem is the one I'd forward to a client. Three separate bugs quietly made a model worse for weeks, and the company wrote it up in plain terms. If I'm building a business on these models I need that kind of honesty about failure modes, and I need my own monitoring so I'm not finding out from a user complaint.

## Sources

- https://github.com/openai/codex
- https://x.ai/news/grok-4-fast
- https://huggingface.co/mistralai/Magistral-Small-2509
- https://huggingface.co/Alibaba-NLP/Tongyi-DeepResearch-30B-A3B
- https://huggingface.co/Wan-AI/Wan2.2-Animate-14B
- https://www.anthropic.com/engineering/a-postmortem-of-three-recent-issues
- https://www.anthropic.com/research/anthropic-economic-index-september-2025-report

Canonical: https://heyaliasif.com/ai/2025-09-15-week-in-ai