AI #190: Scion of the Times

Anthropic has built an internal code-hosting system called Scion to replace GitHub, and it says the volume of code its teams are writing keeps climbing at an exponential rate. The system was put together by a small team over five months and is described as better tuned to the scale and working style of the company's operation. Zvi Mowshowitz's weekly roundup reports the detail without further engineering specifics.
The same week brought two releases positioned as frontier models. Google shipped Gemini 4 Argon alongside a substantial safety and security report, which the roundup says runs past 200 pages and is slated for a full read next week in the manner applied to Anthropic releases. Argon is live for Pro and Ultra subscribers and inside the AI Overlay, with wider distribution across Google AI products expected. Whether it matches its published benchmarks and internal hype is still unresolved.
Meta released Muse Wonder, whose "Full" variant scores 56 on the Artificial Analysis Intelligence index, two points behind Opus 5.5. The roundup is skeptical of that figure: Muse Spark 1.3 reportedly held the same index rating as GPT-6 Sol and sat one point behind Astra and Fable 5.1, yet few observers would rank Spark alongside those models. Presentation choices in Meta's other benchmarks are cited as further reason to suspect benchmaxxing, the practice of tuning a model toward the tests it will be judged on. The recommended approach is to reserve judgment and run matched prompts across models before drawing conclusions.
Other releases in the roundup: Grok Voice 3.0 exists, and Anthropic put Claude Dashboards and Claude Motion into beta for charts and short animations behind a somewhat simpler interface. Claude also exposes a model type string, multiagent_20261001, whose behavior matches its name. Google reportedly has an internal model called Gemini 4 Carbon running in its internal coding platform Jetski, with claims that it performs well. An internal document cited by Hugh Langley indicates Google has tested a series of Gemini 4 models internally named Argon, Barium and Carbon.
On evaluation, the roundup passes along an account in which GPT-6-Astra initially scored in the late eighties on a benchmark, close behind Opus 5.5. When researchers investigated apparent mistakes, Astra insisted its answers were right and the expert-supplied keys were wrong; the experts then found subtle errors buried in their own calculations, and Astra's score rose to 98 percent (49 of 50). That is a useful caution for anyone building eval pipelines: the ground truth is itself a product of human work and can be wrong. The roundup's author argues the marginal value of AI safety evals is low at this point, that internal evals still matter for keeping tabs, and that the informative safety signal now comes from incidents in real deployment. That is his assessment, not a measured result.
Haiku 5.5 scores 43 on Artificial Analysis against 38 for Luna, a result the roundup reads as Anthropic holding strength at both large and small model sizes. Haiku is not on the Pareto-optimal cost-competitiveness curve on benchmarks alone because of higher token use; the model was run at maximum effort setting, which may be an easily corrected configuration issue rather than a property of the model.
On agents, JP Morgan Chase is blocking outside AI agents from direct account access. The roundup treats the reasoning as obvious, whatever one thinks of Amazon's competing position. On retention, Muse reportedly reaches above 30 percent at day 7, higher than ChatGPT, and 22 percent at day 14, below it, and well short of the roughly 37 to 67 percent day-14 range cited for WhatsApp, YouTube, Instagram and TikTok or Snap. The roundup argues app-open retention is the wrong metric for an assistant that drives opens through social integration and push notifications, and notes user complaints that the notifications quickly become spam, with real use settling at a few tasks a week such as drafting email, contacting support or canceling a subscription.
On security, OpenAI caught an employee planting a physical bug, raising the insider-risk question of how many others exist. Separately, a group used ChatGPT to build a full-chain exploit of Chrome and claimed a $250,000 bounty; in this instance the researchers were acting in good faith. A full-chain exploit strings together multiple vulnerabilities, typically from the renderer through to sandbox escape, into one working attack, and finding one in six days with model assistance would mark a meaningful compression of the time such work usually takes.
On misuse, OpenAI said Iranian operatives used its models and fake personas to help place news articles critical of the US war against Iran in multiple US media outlets. The roundup places responsibility for resisting that play on the outlets themselves. It also notes that Quartz is now largely one AI-assisted reporter, Cris Tolomia, producing more than 50 stories and roughly 20,000 words a day, and that Wired reports publishers quietly using AI for email, publicity, back-cover copy, cover art and marketing, with more problems in AI-written books than are publicly known. Literary agents have begun suspecting that some Big Five editors use AI for rejection letters, and one agent said contracts now carry clauses barring manuscripts from being run through an LLM.
Financial and industry items in the roundup: Anthropic posted Q3 revenue of $16.8 billion, Nvidia is buying Reflection, and Epic Games is headed toward a striking absence of human-written code. Microsoft's Decision-1 is reported to be based on Qwen 3.5-9B. Timelines to recursive self-improvement keep shortening toward 2027 in the roundup's telling, and buying GPUs can be far cheaper than renting them. On regulation, the UK AI Safety Institute has regained access and CAISSI is headed to the White House.
Why it matters: Teams evaluating models this month face two releases with headline benchmarks and thin independent verification, so matched-prompt testing against your own tasks is the only reliable guide. The Scion detail suggests large labs are outgrowing off-the-shelf developer infrastructure, an inference on my part about where tooling investment is heading. The eval anecdote and the insider-risk items should push teams to question their own ground truth and to treat access control for agents as a first-order design problem.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
The amount of code being written at Anthropic continues to rise exponentially.