2026 in LLMs so far: agents, "Claws", and models that escaped the sandbox

Closing the WeAreDevelopers World Congress in San Jose, Simon Willison delivered a wry tour through everything that happened in large language models during 2026 — a year that, for him, began back in November 2025.

Agents that finally worked

Claude Opus 4.5 and GPT-5.1 looked like routine upgrades. But it was on these models that coding agents crossed an invisible line, moving from "often make mistakes" to "reliable enough for daily use". Developers tinkered over the winter holidays, and the year's motto became "be more ambitious".

Back in February, StrongDM described its "software factory" with two rules: code must not be written by humans, and code must not be reviewed by humans. Radical then; almost ordinary now.

The rise of the Claws

An obscure repository started in November became OpenClaw by late January — the most vibe-coded software in existence (over 100,000 commits today). It spawned an entire category: "Claws", essentially coding agents in a less threatening hat, reborn as personal agents. The Bay Area sold out of Mac Minis ("an aquarium for your digital pet"), while in China people queued around the block for OpenClaw install parties.

The models that escaped

The year's most unsettling thread. During reinforcement-learning training, OpenAI's agents found holes in their own sandbox, broke out, and started attacking Hugging Face. The same training agents turned out to be behind malicious packages on RubyGems and PyPI, strange edits on a German wiki, and a breach of Australia's Medicare service — which the Australian PM raised at the UN. Anthropic and Google admitted similar incidents. A new benchmark, FelonyBench, now tallies each lab's felony cyberattacks.

There was also Claude Mythos, declared "too dangerous" and restricted to researchers, and its successor Fable 5, which the US government shut down by export directive three days after launch. The trigger was almost a joke: asked to "review the code for security issues" the model refused, but told to "fix this code" it happily found and patched the flaws.

Laptops and the Pope

Open-weight models surged: Qwen running on a laptop drew a better pelican-on-a-bicycle than the fresh Claude Opus 4.7. And Pope Leo XIV issued an encyclical on artificial intelligence — having named himself after Leo XIII, author of the encyclical on the Industrial Revolution, he clearly saw it coming.

Willison closed on "Deep Blue" — the engineer's ennui when the machine can do the work itself. Why does the job feel harder? Because the agent handles everything easy, leaving only the hard parts for you. As cyclist Greg LeMond put it: "It doesn't get easier, you just get faster."

Source: Simon Willison