Python's release team spent the last couple of weeks proving that "ship on schedule" and "ship correct" aren't always the same promise. 3.15.0 was due October 1, a date locked in by PEP 790 more than a year ago. Instead, release manager Hugo van Kemenade pushed out a third release candidate on October 2 after testers found two bugs in the feature everyone actually asked for: lazy imports. Write lazy import a.b as c and, in one of the bugs, Python treated b as an attribute of a instead of importing the submodule at all. The other was sneakier: touching one lazily imported submodule silently pulled in a sibling submodule that was supposed to stay lazy. Neither is the kind of thing a quick unit test catches, which is exactly why they turned up this late. Final 3.15.0 is scheduled for today, three years after lazy imports first showed up as a proposal.
I'd rather the team delay than ship a known regression in a headline feature, and I like it a little less that it took a third release candidate to catch bugs this central to behavior that changes when your code actually runs. Lazy imports mean deferred side effects, and deferred side effects are exactly the kind of thing nobody wrote a test for in a production import graph. If your team cares about import-time behavior, budget real time for 3.15 testing before you flip the switch, and don't volunteer to be the project that finds the next edge case in prod.
Mistral spent the same week doing the opposite of careful. On October 6 it put Large 4, nicknamed "le Chonk" after a months-old internet joke the company apparently decided to run with, into API preview: a mixture-of-experts model with roughly 1.05 trillion total parameters and 49 billion active per token, trained from scratch on 3,800 Nvidia Blackwell GPUs in Mistral's own European data centers. Open weights are promised for October 27, after what the company says is three weeks of red-teaming. It's a real jump over Large 3, which scored 9 on Artificial Analysis's index. Large 4 scores 38. The catch is that 38 still lands behind DeepSeek's V4.1 Flash, a model roughly a third the size.
A trillion parameters bought Mistral a big improvement over its own last model and a result that a much smaller competitor already beats. That's less a knock on Mistral than a reminder that raw parameter count stopped being the thing that wins these comparisons a while back, and architecture and training recipe matter more than the number you can put in a press release. Still, Mistral remains one of the only labs outside China actually shipping open weights at this scale, and that's worth something on its own, before you even get to the benchmark table.
Meanwhile GitHub quietly reminded everyone building on Copilot that the model list under them isn't a stable foundation to build against. On October 2 it deprecated Gemini 3.5 Flash, Gemini 3.6 Flash, Kimi K2.7 Code, and Claude Opus 4.7 across every Copilot surface, chat, inline edits, agent mode, code completions, and pointed people toward Gemini 3.8 Flash, Kimi K3, and Claude Opus 5.5 instead. A second wave lands October 19, retiring GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, Gemini 3.7 Flash, and Grok 4.5.
None of this is unreasonable by itself. Models get superseded, providers retire old endpoints, and GitHub gave weeks of notice both times. But weeks of notice only helps if somebody's actually reading the changelog. If your team has a model ID hardcoded into a CI config, an extension manifest, or a Copilot policy setting somewhere, two deprecation waves inside three weeks is a pace that makes a standing process, someone who checks github.blog/changelog on a schedule, worth more than good intentions. The model you built a workflow around in August isn't guaranteed to exist in October, and at the rate these three stories moved this week, that's true of pretty much every vendor right now, not just GitHub.