The week in AI

Anthropic published three times this week. On 23 February it accused DeepSeek, Moonshot AI and MiniMax of distilling Claude through about 24,000 fraudulent accounts. The new scaling policy followed on 24 February, and on 25 February Anthropic acquired the computer-use startup Vercept. On 23 February OpenAI also stopped reporting SWE-bench Verified and called the benchmark contaminated and flawed.

On the product side, Alibaba's Qwen team released three more open Qwen3.5 models on 24 February. Perplexity launched an agent called Perplexity Computer on 25 February, and Google released Nano Banana 2 on 26 February.

The Responsible Scaling Policy (RSP) is the document in which Anthropic ties its safety measures to capability levels, called AI Safety Levels (ASL). Anthropic gives three reasons for the rewrite. Its pre-set capability thresholds proved ambiguous, especially in biology. The political climate turned against regulation. And it says the safeguards required at the higher ASL levels may be impossible for one company to put in place alone.

Version 3.0 therefore has two parts. One lists what Anthropic commits to do by itself. The other lists what it recommends for the industry as a whole. Anthropic also added a public Frontier Safety Roadmap of goals it calls "ambitious but achievable" and says it will grade itself against them in public.

The policy also adds Risk Reports, which are assessments of model risk that go to outside reviewers. Anthropic published the policy only, so no Risk Report from it was available yet.

Distillation means training one model on the outputs of another, so the student model picks up the teacher's behaviour without access to its weights. Anthropic says the three labs reached Claude through proxy services it calls "hydra clusters", which spread traffic across many accounts to avoid detection.

The counts are Anthropic's own. MiniMax accounts for over 13 million exchanges, Moonshot for over 3.4 million and DeepSeek for over 150,000. Each lab went after different skills, according to Anthropic. DeepSeek targeted reasoning and reward modelling, Moonshot targeted agentic reasoning, coding and vision, and MiniMax targeted agentic coding and tool use.

Anthropic says it watched MiniMax redirect its traffic to a newly released Claude model within 24 hours of the launch. The three labs' responses are not part of the record.

SWE-bench Verified is a set of real GitHub issues that a model must fix so that the repository's tests pass. OpenAI published the human-validated version in 2024, and it became the standard coding score in model launches.

OpenAI audited the hard tasks that models often failed, which made up 27.6% of the benchmark. It reports that at least 59.4% of those audited tasks had test cases that reject correct solutions. It also found signs that models had trained on the problems. Top scores had moved only from 74.9% to 80.9% over six months, and OpenAI reads that slowdown as the benchmark running out.

OpenAI recommends SWE-bench Pro in its place. Other labs had not yet said whether they would also stop reporting Verified.

Vercept built models that see a screen and operate software the way a person would. Anthropic is winding down Vercept's own product. Ross Girshick is the computer vision researcher behind the R-CNN line of object detectors.

Computer use is an area where Claude had improved fast before the deal. On OSWorld, a benchmark of desktop tasks, Anthropic's Sonnet models went from under 15% to 72.5%, by Anthropic's reported scores. The price of the acquisition is not in the record.

• Qwen3.5 gained three open-weight models on 24 February under Apache 2.0. They are 122B-A10B and 35B-A3B, which are mixture-of-experts models that use about 10 billion and 3 billion parameters for each token, and a dense 27B.
• Perplexity Computer launched on 25 February for Max subscribers. It takes a goal, plans the steps, browses, writes files and code, and sends each step to one of 19 models, by Perplexity's count.
• Nano Banana 2, also called Gemini 3.1 Flash Image, launched on 26 February across the Gemini app, Search and Ads. Google says it renders text accurately and pulls in web search to get subjects right, and it outputs images from 512 pixels to 4K, marked with SynthID and C2PA Content Credentials.
• Anthropic published the persona selection model on 23 February. It is a theory, with no experiment behind it, that pretraining teaches a model to simulate many characters and post-training picks one of them as the Assistant, so human-like traits are there from the start.

• Ross Girshick joined Anthropic on 25 February with his Vercept co-founders Kiana Ehsani and Luca Weihs when Anthropic acquired the company.