This Week in AI: February 9-15, 2026 (Week 7)

News of the week: February 9-15, 2026 (Week 7)

The week was dominated by money, speed and copyright battles. Anthropic raised 30 billion dollars at a 380 billion valuation, while OpenAI launched a lightning-fast coding model running on Cerebras chips and started testing ads in ChatGPT. China delivered a wave of new models from Zhipu, MiniMax and ByteDance, and the Seedance 2.0 video generator immediately triggered a sharp protest from Hollywood. On the stock markets, fear of AI’s impact spilled over from software companies into real estate and logistics, and the UN approved its first scientific panel on artificial intelligence.

New Models

OpenAI: GPT-5.3-Codex-Spark runs on a chip the size of an entire wafer

On Thursday, February 12, OpenAI introduced GPT-5.3-Codex-Spark, a smaller and faster version of its coding model GPT-5.3-Codex. It is designed for real-time work, where a developer keeps refining code together with the model and does not want to wait for a response (TechCrunch).

The hardware is the most interesting part. The model runs on the Wafer Scale Engine 3 chip from Cerebras, which takes up an entire silicon wafer and carries 4 trillion transistors. It is the first result of a multi-year agreement between the two companies worth more than 10 billion dollars.

For now, Spark is available as a research preview for ChatGPT Pro subscribers in the Codex app. For OpenAI, it is also a signal that it does not want to depend solely on chips from Nvidia (OpenAI).

Google: a significantly stronger Gemini 3 Deep Think

On February 12, Google released a major update to its Gemini 3 Deep Think mode. It is a “thinking” mode that takes more time to answer and targets scientific, research and engineering tasks (Google).

According to Google, it scores 48.4% on Humanity’s Last Exam without tools and 84.6% on the abstract reasoning test ARC-AGI-2.
In Codeforces programming competitions it has an Elo rating of 3455, and on problems from the 2025 International Mathematical Olympiad it reaches gold medal level.
The mode is available to Google AI Ultra subscribers in the Gemini app, and researchers and companies can apply for access through the Gemini API.

A day earlier, on February 11, Google DeepMind described Aletheia, a research agent built on Deep Think. It worked through 700 open Erdos problems and solved four of them on its own (Google DeepMind).

A Chinese wave before the New Year: GLM-5, MiniMax M2.5 and Seed 2.0

Ahead of the Chinese Lunar New Year, local companies released several strong models at once. Zhipu AI (brand Z.ai), listed in Hong Kong since January, launched GLM-5 on February 11, a model focused on coding and agentic work (Caixin, The Star / Reuters).

GLM-5 reportedly has 744 billion parameters, of which only 40 billion are active at each step (a “mixture of experts” architecture, meaning a combination of specialized subnetworks). The weights are freely downloadable under the MIT license (Hugging Face).
MiniMax released M2.5 and the faster M2.5-Lightning on February 12. On the SWE-Bench Verified coding benchmark it reports 80.2%, on par with Claude Opus 4.6, at roughly a tenth of the cost per task (MiniMax).
ByteDance unveiled the Doubao Seed 2.0 family in Pro, Lite and Mini variants on February 14. It positions the Pro version against GPT-5.2 and Gemini 3 Pro (TechNode).

Seedance 2.0 and the Hollywood storm

On February 12, ByteDance officially launched the Seedance 2.0 video generator, which can combine text, images, audio and video as input (ByteDance Seed). Clips featuring characters from famous films spread quickly across social media.

The reaction was immediate. On February 12, Motion Picture Association chief Charles Rivkin called on ByteDance to “immediately cease its infringing activity” and spoke of unauthorized use of American works “on a massive scale” (MPA).

The next day, Disney sent a cease and desist letter. It states that the service comes preloaded with a “pirated library” of characters from Star Wars and Marvel (Axios). The actors’ union SAG-AFTRA also objected to the use of performers’ voices and likenesses (SAG-AFTRA).

Products and Agents

ChatGPT starts showing ads

Since February 9, OpenAI has been testing ads in ChatGPT in the US. They are shown to logged-in users of the free version and of the low-cost Go plan at 8 dollars a month, while Plus, Pro and business subscribers will not see them (Euronews).

The ads are labeled as sponsored and kept separate from the answers themselves. They are selected based on conversation topics and earlier chats, but advertisers only receive aggregated data on impressions and clicks. Meanwhile, Anthropic used the Super Bowl to promote the slogan “Ads are coming to AI. But not to Claude.”

The creator of OpenClaw heads to OpenAI

On Sunday, February 15, Sam Altman announced that Peter Steinberger, the author of the popular open source agent OpenClaw (formerly Clawdbot and Moltbot), is joining OpenAI. The agent runs on the user’s computer and handles tasks on their behalf, such as emails or working with files (Bloomberg).

At OpenAI, Steinberger is set to lead development of the next generation of personal agents. The OpenClaw project will move to a foundation and remain open, with continued support from OpenAI (Peter Steinberger’s blog).

February 13: OpenAI permanently retired the GPT-4o, GPT-4.1, GPT-4.1 mini and o4-mini models from ChatGPT. According to the company, only about 0.1% of users chose GPT-4o each day, yet some of the model’s fans still protested (Engadget, OpenAI).
February 12: Anthropic made it possible for companies to buy Claude Enterprise directly on the web without talking to sales. Customers pay per seat plus actual usage at API rates (Claude).
February 10: According to court filings, the first OpenAI device designed by Jony Ive will not arrive before the end of February 2027. The company will also not use the name “io” for its hardware (MacRumors).
February 12: Waymo began fully autonomous operations with the sixth generation of the Waymo Driver. It is cheaper, has fewer than half the cameras of the previous generation and is built to handle winter weather (Waymo).
February 12: Spotify co-CEO Gustav Soderstrom said the company’s best developers have not written a single line of code since December. Tools built on Claude Code do the programming for them (TechCrunch).

Business and Investment

Anthropic: 30 billion dollars at a 380 billion valuation

On February 12, Anthropic announced a 30 billion dollar Series G funding round. It values the company at 380 billion dollars, and the round was led by Singapore’s sovereign wealth fund GIC and the investor Coatue (Anthropic).

Annual run-rate revenue reached 14 billion dollars and, according to the company, has grown more than tenfold for three years in a row.
The Claude Code coding tool alone has an annual revenue run rate of more than 2.5 billion dollars.
More than 500 customers spend over a million dollars a year on Claude, and eight of the ten largest US companies on the Fortune 10 list are customers.

AI fears shake the markets

The sell-off that started with software companies spread during the week to sectors investors consider vulnerable to automation. Insurers, brokers, real estate services and logistics were next in line (The Star / Reuters).

Shares of the real estate firm CBRE fell almost 9% on February 12 after its CEO acknowledged that AI could reduce demand for office space. Some analysts, however, called the moves exaggerated and driven more by market sentiment (Fortune).

February 9-11: Co-founders Tony Wu and Jimmy Ba left Elon Musk’s xAI within 48 hours. After the merger with SpaceX, only six of the twelve founders remain at the company (The Decoder).
February 10: Cisco unveiled the Silicon One G300 networking chip with 102.4 Tb/s throughput for AI data centers, along with new fully liquid-cooled switches (Cisco).
February 11: Anthropic pledged to pay all the costs of upgrading the power grid for its data centers and to cover any electricity price increases for households caused by its consumption (Anthropic).

Regulation, Politics and Law

The UN has a scientific panel on AI, the US voted against

On February 12, the UN General Assembly approved, by 117 votes to 2, the 40 members of the Independent International Scientific Panel on AI. The United States and Paraguay voted against, calling the panel an overreach of the UN’s mandate (AP).

The members were selected from more than 2,600 candidates and will serve three years. The panel will publish an annual report on the risks, opportunities and impacts of AI, with the first expected for the July Global Dialogue on AI Governance (UN).

OpenAI accuses DeepSeek of “distilling” American models

On February 12, OpenAI sent a memo to the House of Representatives Select Committee on China claiming that China’s DeepSeek trains its models on the outputs of American models. This technique is called distillation: a smaller model learns to imitate the answers of a stronger one (Bloomberg).

According to the memo, DeepSeek employees concealed their identity through VPNs and intermediary services and queried the models using automated scripts. Committee chairman John Moolenaar described it as a “steal, copy and destroy” strategy (Yahoo Finance).

Claude reportedly helped in the operation against Maduro, says WSJ

According to the Wall Street Journal, the US military used Claude during the January operation in which it captured former Venezuelan president Nicolas Maduro. It accessed the model through Anthropic’s partnership with the data company Palantir (The Star / Reuters).

The report came out on February 13; Reuters could not independently verify it, and it is unclear what role Claude played. Anthropic did not comment on specific operations, although its policies prohibit using Claude to support violence or surveillance.

February 12: Anthropic donated 20 million dollars to Public First Action, a new bipartisan organization focused on AI education and policy. It cites a survey according to which 69% of Americans believe AI regulation is insufficient (Anthropic).

Research and Education

February 11: The Aletheia agent from Google DeepMind scored roughly 90% on IMO-ProofBench Advanced (olympiad-level proofs). According to the team, one of the scientific papers was produced entirely without human intervention (Google DeepMind).
February 13: Anthropic partnered with the nonprofit CodePath, the largest US provider of college computer science education. Claude Code will be brought into courses for more than 20,000 students, over 40% of whom come from families earning less than 50,000 dollars a year (Anthropic).