Skip to content

Weekly digest

12 – 18 October 2026

9 posts

OpenAI had most of this week’s developer news. In Codex and ChatGPT Work, GPT-6.1 Sol now gets 500,000 tokens of context by default, up from 272k. OpenAI says it will raise that again but hasn’t said when. GPT-6 Astra Pro now lets several agents work on one task at once in ChatGPT, though only for Pro users in the Chat tab.

On the Google side, a Threads post says Gemini 4 Argon is reaching Pro subscribers in Antigravity with its own weekly quota, which Google hasn’t confirmed. On Arena it sits fifth, behind Opus 5.5 and GPT 6.1 Sol.

Codex and ChatGPT Work give GPT-6.1 Sol 500k context

Half a million tokens is what GPT-6.1 Sol now holds by default in Codex and ChatGPT Work, after OpenAI raised its default context length to 500,000. Think of context as the model’s desk: the bigger it is, the more repo files and chat history stay in view before older material has to go.

OpenAI also plans to push the default higher, but it hasn’t said when or to what number, so 500k may not stay the ceiling for long. The source names only Codex and ChatGPT Work. Until this change, the default sat at 272k.

Testing Catalog ThreadsDiscuss

GPT-6 Astra Pro splits one task among parallel agents

Several agents can work on different parts of the same task at once in ChatGPT, and you can follow each one’s progress as it goes. When an agent finishes, you open its completed piece and review the work. OpenAI announced this as part of an upgrade to GPT-6 Astra Pro.

The same update brings faster responses and better use of plugins and tools. OpenAI didn’t put a number on the speedup. The feature is available to Pro users in the Chat tab only.

Testing Catalog, which flagged the update, wonders whether Team accounts will get it too, and OpenAI’s post doesn’t say. Team users, meanwhile, keep running their agents one at a time.

OpenAI TwitterTesting Catalog ThreadsDiscuss

GPT-Live-1.1 makes ChatGPT Voice say “checking...” less

That recurring “checking...” in voice mode is getting quieter. OpenAI’s new GPT-Live-1.1 model powers Voice in ChatGPT, Codex and dots, and it’s rolling out today to all users, with API access “coming soon.”

The update also brings better accents and new voices in French, German, Hindi, Mexican Spanish, Korean and Japanese. Per Testing Catalog, the constant “Checking..” should be gone, which had been annoying during vibe coding sessions in Codex.

In our view, “less checking” is a soft promise: OpenAI gave no figure for how much shorter the pauses got, and no date for the API.

OpenAI TwitterTesting Catalog ThreadsDiscuss

Gemini 4 Argon reportedly lands in Antigravity for Pro users

According to a post on Threads, Google is rolling out Gemini 4 Argon to Google AI Pro subscribers in Antigravity, with a dedicated weekly quota just for Argon. Google’s official materials describe Argon as rolling out but don’t confirm the Pro-tier access or the separate quota, and the post says nothing about how big that quota is. Until now, Argon access we’d covered ran through the paid tier of the Gemini API.

Testing Catalog ThreadsDiscuss

Gemini 4 Argon lands fifth on Arena, behind Opus 5.5

Fifth place on Arena’s overall ranking is where Gemini 4 Argon landed, according to a post on Threads, behind Opus 5.5, GPT 6.1 Sol, Fable 5.1 and GPT 6 Astra. Three of those four are the very models Google picked for its own comparison table of 19 benchmark scores. The leaderboard is live, so the order may have shifted since the post went up.

Testing Catalog ThreadsDiscuss

Meta’s Muse Wonder Full reportedly stays off the API

A Threads post, so far unconfirmed by Meta’s documentation or a second outlet, says Meta is rolling out Muse Wonder and Muse Wonder Full on the Meta Model API platform, but Wonder Full, “our most capable model, running at max reasoning,” goes only to Muse Code subscribers. Muse Wonder gets the API slot and is pitched as “state-of-the-art at real-world coding”: it plans long jobs, calls tools in parallel, hands parts to sub-agents and recovers from errors. The post cites no benchmarks or prices, and the latest independently covered release, per MarkTechPost, is Muse Spark 1.3.

Testing Catalog ThreadsDiscuss

Linus Torvalds uses Claude Code on a hefty kernel change

Per Phoronix, Linus Torvalds is working on what the site calls a hefty Linux kernel improvement with Claude Code. According to Penguin Alley, Torvalds also committed a kernel bug fix after a debugging session he described as “from hell,” crediting AI assistance with “much of the grunt-work.”

Hacker NewsDiscuss

Anthropic offers Blackstone a double-digit yield on chips

Per the Financial Times, Blackstone is being offered a double-digit yield in Anthropic’s latest chip financing. The FT piece is paywalled, and its headline gives neither the exact rate nor the size of the deal. For comparison, Reuters earlier reported that Broadcom is lending Anthropic up to $42 billion for its buildout.

Hacker NewsDiscuss

Anthropic’s June credit rating: the top tier of junk

Per the Financial Times, Anthropic was assigned the highest junk credit rating in June. Junk, or speculative grade, covers everything below investment grade, so the highest junk rating sits just under that line. The FT piece is paywalled, and its headline doesn’t name the rating agency or the exact grade.

Hacker NewsDiscuss