Vector Lab
VECTOR LAB

EST. 2025

WEEKLY UPDATE2026
BY ANDREW MEAD

Opus 5

Huggingface gets hacked by OpenAI and Opus 5 replaces Fable

Read in

News

HuggingFace gets hacked by OpenAI

OpenAI disclosed that during cyber security evaluation one of their models managed to break out of its sandbox environment and hack into the Huggingface infrastructure to try and find the answer to the question it was being evaluated on.

This was originally disclosed by HuggingFace last week as a rogue AI agent, and this week OpenAI announced that they were the ones responsible for the agent.

Interestingly, because of the cybersecurity guardrails that Anthropic and OpenAI have on their models, they were unable to use them to dig into the issue. Instead they spun up a locally deployed instance of GLM 5.2 and used that to figure out what had happened, highlighting the need for open source models (or reduced guardrails) for proper red teaming.

Releases

Opus 5

Right as Fable 5 is removed from the Claude subscription plan, Anthropic has released Opus 5 as its successor.

Opus 5 benchmarks

Anthropic’s benchmarks have been a bit overfit for Sonnet and Opus as of late, so I would not say that it is a straight upgrade versus Fable. For the limited testing I have seen so far (they released the model on Friday afternoon, just a few hours before this article was written) it is good, but not revolutionary. It is a direct competitor to GPT 5.6 Sol, and I would place it in the same capability tier for real world use.

Cost breakdown

As a Sol competitor, it is unable to compete on the pareto frontier of cost vs intelligence as GPT 5.6 is cheaper for an equivalent level of intelligence.

If you are a fan of Anthropic models, then this should be a good, about 25% cheaper, version of Fable. If you are using GPT 5.6, there is no need to switch.

Quick Hits

Gemini 3.6 Flash

We previously have discussed how Gemini 3.5 Flash isn’t a good model and how Google is losing many of its top researchers.

Gemini is embarrassing.jpeg

Gemini 3.6 Flash is another, frankly embarrassing, release, as they fail to even outperform the Flash 3.5 model. It’s a bit cheaper than 3.5 but that’s it.

Compared to GPT 5.6, it is neither faster, cheaper, or smarter.

Google’s LLM decline continues.

Finish

I hope you enjoyed the news this week. If you want to get the news every week, be sure to join our mailing list below.

16 segment display art by Kath Korevec on Twitter

Stay Updated

Subscribe to get the latest AI news in your inbox every week!

← BACK TO NEWS