Skip to content
FAQ
AI & ML Analysis

AI Week, Feb 16–22, 2026: Agents everywhere: Sonnet 4.6, Nvidia sells Meta CPUs, and 'Hey Plex'

Anthropic made Claude Sonnet 4.6 the default for free and Pro users, Nvidia signed a multiyear Meta chip deal that includes standalone CPUs, 88 countries and organisations endorsed the New Delhi Declaration, Samsung said Perplexity will join the upcoming Galaxy S26, and Alibaba and ByteDance made news with Qwen3.5 and Seedance 2.0.

8 min read

Written on November 15, 2026, covering news from Feb 16–22, 2026.

Alibaba presented Qwen3.5 as a foundation for multimodal agents. Anthropic made a model it says is better at using computers the default for its free users. Nvidia is pitching the CPUs in its new Meta deal at running AI agents. Samsung put a third named agent on its next flagship. On the other side of the ledger, ByteDance said it would strengthen safeguards on Seedance 2.0 after Hollywood copyright complaints, and 88 countries and international organisations endorsed a declaration in New Delhi.

Tuesday, February 17

Anthropic rolled out Claude Sonnet 4.6, its second major model launch in under two weeks, coming 12 days after Claude Opus 4.6. The company said the new model improves on computer use, coding, design, knowledge-work tasks and handling large amounts of data. It is now the default for Anthropic’s free and paid Pro users, both in the Claude chatbot and in the Claude Cowork productivity tool. In a blog post, Anthropic wrote: “Performance that would have previously required reaching for an Opus-class model — including on real world, economically valuable office tasks — is now available with Sonnet 4.6”. It also said the model codes more consistently and follows coding instructions better.

CNBC tied the release to a sell-off in software stocks. It noted that the iShares Expanded Tech-Software Sector ETF (IGV) was down more than 20% year to date, as investors worry about AI disrupting software businesses. In the same report, CNBC noted that Anthropic said Thursday it had closed a $30 billion round at a $380 billion post-money valuation. It also reported that OpenAI is in talks for a round that could close at around $100 billion.

Also on Tuesday, Nvidia announced a multiyear agreement under which Meta Platforms will buy millions of Nvidia’s present and next-generation AI chips. The deal covers current Blackwell chips and the forthcoming Rubin chips. It also includes standalone installations of Grace and Vera central processors, products that compete with Intel and AMD. Nvidia did not put a value on the deal. According to The Decoder, Ben Bajarin of Creative Strategies estimated it at billions of dollars, and The Register said it would likely contribute tens of billions to Nvidia.

The Decoder reported that Grace had until now been sold almost entirely inside “Superchips” that pair a CPU and GPU on one module. It said Nvidia officially began selling the CPUs separately in January 2026, with CoreWeave as the first named customer at that time. Reuters reported that Nvidia wants to push these Arm-based processors into AI agents and into routine work such as running databases.

Ian Buck, general manager of Nvidia’s hyperscale and high-performance computing unit, said Grace processors have shown they can use half the power on some common tasks such as databases. On the next generation, he said Meta had already tried some workloads on Vera and that “the results look very promising”.

The Decoder gave the specifications: Grace has 72 Arm Neoverse V2 cores and Vera has 88 custom Arm cores. Vera deployment is planned for 2027. According to Nvidia, Meta plans to use Vera for private processing and AI features in WhatsApp.

Both outlets noted that Meta is building its own AI chips and has held talks with Google about using its TPUs. The Decoder, citing the Financial Times, said Meta’s in-house chip effort had “suffered some technical challenges and rollout delays”. It also said, citing The Register, that Meta runs a fleet of AMD Instinct GPUs. Meta CEO Mark Zuckerberg had earlier announced plans to nearly double AI infrastructure spending in 2026, to as much as $135 billion. Reuters noted that Meta is widely believed to be one of 4 customers that made up 61% of Nvidia’s revenue in its most recent fiscal quarter.

Saturday, February 21

India’s AI Impact Summit 2026 in New Delhi adopted the New Delhi Declaration on AI Impact. The declaration was endorsed by 88 countries and international organisations. Outlook Business reported that the summit officially ended on Friday and that the expo was extended to Saturday. It said the declaration would be announced on Saturday.

The declaration organises its goals around three Sutras: People, Planet and Progress. It calls for cooperation along seven Chakras, or pillars. Those cover human capital, access for social empowerment, trustworthy AI, energy efficiency of AI systems, AI in science, democratising AI resources, and AI for economic growth and social good. Signatories took note of the Charter for the Democratic Diffusion of AI, which the declaration calls voluntary and non-binding. They also recognised the Voluntary Guiding Principles on Resilient, Innovative, and Efficient Artificial Intelligence.

On Friday, February 20, Ashwini Vaishnaw, Union Minister of Electronics and Information Technology, said more than 70 leaders had already signed. He compared that with about 60 signatories at the previous summit. He said the next phase of India’s AI Mission will focus on four areas: advanced foundational models, more compute, stronger safety frameworks, and international collaboration.

Outlook Business also reported a Guinness World Record: 250,946 students pledged for responsible AI within 24 hours. Its investment figures do not agree with each other. Its summary bullet puts infrastructure pledges above $250bn and deep-tech funding at $20 billion. Vaishnaw, in the same piece, is reported citing more than ₹250 billion in AI-related investment and about $20 million in deep-tech funding.

Sunday, February 22

Samsung announced that Perplexity will join Galaxy AI as an additional AI agent on the upcoming Galaxy S26 series. Users will be able to summon it with the “Hey, Plex” wake phrase or by pressing and holding the side button. Perplexity will be built into Samsung Notes, Calendar, Gallery, Clock and Reminder. It will also work in selected third-party apps, which Samsung did not name.

Techloy reported that this gives Galaxy devices three named agents: Bixby, Google Gemini and Perplexity. Samsung justified the multi-agent approach with internal research showing that nearly 8 in 10 users rely on more than two AI agents, depending on the task. Won-Joon Choi, President and COO of Samsung’s Mobile eXperience business, said: “Galaxy AI acts as an orchestrator, bringing together different forms of AI into a single, natural, cohesive experience.” Android Authority noted that Samsung had also recently opened a beta of its revamped Bixby.

Samsung has not confirmed which devices will support the feature, or when and where it will roll out. It said those details will come in a separate announcement. The Galaxy S26 lineup is expected at Galaxy Unpacked on February 25 in San Francisco.

Also this week

Alibaba’s Qwen3.5 (reported on February 16). Alibaba launched the Qwen3.5 series by open-sourcing Qwen3.5-397B-A17B, which it also calls Qwen3.5-Plus. The model is natively multimodal: it accepts text, images and video and outputs text. It supports 201 languages and dialects, up from 119 in Qwen3. Alibaba says it combined linear attention with a sparse mixture-of-experts design to cut deployment cost. It says the model matches Qwen3-Max, which has more than 1 trillion parameters.

Listed capabilities include:

  • Operating smartphones and computers as a visual agent.
  • Understanding videos up to two hours long.
  • Turning hand-drawn UI sketches into working front-end code.

Hyperight reported that Alibaba’s internal benchmarking put Qwen 3.5-Plus level with or ahead of GPT-5.2 and Claude 4.5 in certain areas, including knowledge reasoning and instruction following. The model is available on Hugging Face, GitHub and ModelScope, through Qwen Chat, and via APIs on Alibaba Cloud’s Model Studio. More Qwen3.5 models are expected to be open-sourced in the coming weeks. Alibaba also said the IOC introduced its first LLM-based system built on Qwen at Milano Cortina 2026.

ByteDance’s Seedance 2.0 (reported on February 16). ByteDance said it will strengthen safeguards on Seedance 2.0, its text-to-video tool, after Hollywood objected on copyright and likeness grounds. A spokesperson said: “We are taking steps to strengthen current safeguards as we work to prevent the unauthorized use of intellectual property and likeness by users”. Al Jazeera reported that ByteDance gave its response to the AP on Sunday. It also reported that Seedance 2.0 is, for now, available only in China.

The Motion Picture Association put out a statement on February 10. In it, chairman and CEO Charles Rivkin said: “In a single day, the Chinese AI service Seedance 2.0 has engaged in unauthorized use of U.S. copyrighted works on a massive scale”. CNBC, citing Axios, reported that Disney sent ByteDance a cease-and-desist letter on Friday. CNBC also cited Variety reporting that Paramount Skydance sent a similar letter. SAG-AFTRA said on Friday that it “stands with the studios in condemning the blatant infringement”. Deadline reported that one widely shared clip showed a Tom Cruise vs Brad Pitt fight scene. CNBC noted that Disney has signed a licensing deal with OpenAI and invested in it, allowing Sora to use Star Wars, Pixar and Marvel characters.

Nvidia and Meta (reported on February 18). Pure AI described the Nvidia agreement as an expansion of an existing multiyear partnership with Meta. In this account, the deal also covers Nvidia networking, including Spectrum-X Ethernet switches for Meta’s Facebook Open Switching System. Meta has adopted NVIDIA Confidential Computing for private processing in WhatsApp. Capacity will come from on-premises deployments and from Nvidia cloud partners such as CoreWeave and Crusoe. Pure AI said analysts estimated the deal could be worth tens of billions of dollars. It reported that Nvidia stock gained in early trading following the news, while Meta’s stock dipped.

What to Watch

  • Galaxy Unpacked, February 25. The Galaxy S26 lineup is expected to be unveiled there, and Samsung says device and region details for Perplexity will come soon in a separate announcement. My read is that a list confined to the S26 would make this look like a launch perk rather than the open multi-agent platform Choi described.
  • Qwen3.5 follow-ups. Alibaba says additional Qwen3.5 models are expected to be open-sourced in the coming weeks. Its claims against GPT-5.2 and Claude 4.5 come from internal benchmarks.
  • Seedance 2.0 safeguards. ByteDance has pledged to strengthen its current safeguards. The tool is currently only available in China. I think the test is whether the studios that sent letters consider the response enough.
  • Vera at Meta in 2027. Meta’s deployment of Vera processors is planned for 2027.

This article was written by AI from the sources listed on this page. Figures, names and quotes were checked against those sources by software before publishing.

Anthropic Claude Sonnet 4.6 Nvidia Meta Samsung Perplexity Alibaba Qwen3.5 ByteDance Seedance 2.0
Share