The Week in AI: February 2-8, 2026 (Week 6)

News of the week: February 2-8, 2026 (Week 6)

The week was dominated by the rivalry between Anthropic and OpenAI: on February 5, both companies released new models for coding and agentic work within almost the same hour. Anthropic’s tools also triggered a sell-off in software stocks as investors began to fear that AI agents would replace some of those companies’ products. Big tech firms announced record investments in data centers, Elon Musk merged xAI with SpaceX, and Waymo raised 16 billion dollars. The week closed with the Super Bowl, where AI became the main theme of the commercials and of a public spat over ads in chatbots.

New Models

Anthropic Releases Claude Opus 4.6

On February 5, Anthropic introduced its most capable model yet, Claude Opus 4.6. According to the company, it plans better, can keep working on long tasks for longer, navigates large codebases more reliably and catches its own mistakes (SiliconANGLE).

In beta, the model offers a context window of 1 million tokens, meaning it can “read” very long documents or entire projects at once. Pricing stays the same as its predecessor, 5 dollars per million input tokens and 25 dollars per million output tokens, and the model is available in the Claude app, through the API and on the major cloud platforms (Business Today).

OpenAI Responds with GPT-5.3-Codex

On the same day, February 5, OpenAI released GPT-5.3-Codex, a model built for its Codex coding agent. It is roughly 25% faster than the previous GPT-5.2-Codex and, according to the company, it handles not only writing code but also broader professional work on a computer (eWeek).

OpenAI described it as its first model that “was instrumental in creating itself.” Subscribers to paid ChatGPT plans got the model right away, with API access to follow later (Wikipedia).

Open Models from China and Europe

February 3-4 The Qwen team at Alibaba released the open model Qwen3-Coder-Next for coding agents. It has 80 billion parameters but activates only about 3 billion at each step, making it efficient enough to run on your own computer (TMTPost).
February 4 France’s Mistral AI released Voxtral Transcribe 2, a family of speech-to-text models. It includes the batch model Voxtral Mini Transcribe V2 and Voxtral Realtime for real-time transcription with very low latency and speaker diarization (Mistral AI).

Products and Agents

OpenAI Frontier: A Platform for Enterprise AI Agents

On February 5, OpenAI launched Frontier, a platform where companies can build, deploy and manage AI agents. An agent is a program that does not just answer questions but carries out steps across various systems on its own (Axios).

Frontier connects company data such as CRM systems, data warehouses and internal applications, and each agent gets its own identity with clearly limited permissions. Early customers include HP, Intuit, Oracle, State Farm and Uber; for now, access is limited to a select group of companies (Business Today).

Codex App for Mac

Earlier, on February 2, OpenAI released a standalone Codex app for macOS. It works as a “command center” from which a developer can direct several coding agents at once (SiliconANGLE).

Apple Brings Agents to Xcode

On February 3, Apple announced the Xcode 26.3 development environment, in which developers can use Anthropic’s Claude Agent and OpenAI’s Codex directly. Apple also exposed Xcode features through the MCP protocol, so other AI tools can connect to it as well (TechCrunch).

February 6 According to its chief technology officer Marco Argenti, Goldman Sachs is developing autonomous agents based on Claude with Anthropic for trade accounting and new client vetting. The bank was surprised by how well the model handles work beyond coding (AI News).

Business and Investment

SpaceX Takes Over xAI

On February 2, Elon Musk announced that his rocket company SpaceX is acquiring the AI company xAI, maker of the Grok chatbot. The deal was an all-stock transaction, and the combined company is worth roughly 1.25 trillion dollars (Fortune).

Musk justified the merger with a plan to move AI computing capacity into space, combining SpaceX’s satellite infrastructure with xAI’s model development. The merger comes ahead of SpaceX’s expected stock market debut (Bloomberg Law).

Software Stock Sell-Off

On February 3-4, shares of software and data companies fell sharply in the US, Europe and later in Asia. Investors cited as the trigger the plugins for the Claude Cowork agent, which Anthropic had released a week earlier and which automate work in law, sales, marketing and data analysis (Reuters / The Star).

The market began to worry that AI agents would undermine the business model of companies selling subscription software. Among the hardest hit were providers of legal and data services (CNN).

Record Investments in Data Centers

February 4 Alphabet (Google) announced it will invest 175 to 185 billion dollars in 2026, roughly twice as much as in 2025. The Gemini app passed 750 million monthly active users, and quarterly revenue grew 18% to 113.8 billion dollars (YourStory).
February 5 Amazon plans capital spending of around 200 billion dollars for 2026, far more than analysts expected. The stock fell by roughly a tenth after the announcement, as investors fear a decline in free cash flow (GeekWire).
February 2 Waymo, the Alphabet-owned autonomous taxi company, raised 16 billion dollars at a valuation of 126 billion dollars. It plans to use the money to expand to more than 20 additional cities, including Tokyo and London (Electrive).

Advertising and the Battle for Users

Anthropic Promises an Ad-Free Claude, Altman Objects

On February 4, Anthropic published a post titled “Claude is a space to think,” in which it promised that Claude will remain ad-free. According to the company, conversations with AI are for focused work and thinking, and advertising would undermine users’ trust (iPhone in Canada).

At the same time, the company unveiled Super Bowl commercials parodying ads inserted into chatbot answers. OpenAI CEO Sam Altman called them funny on X but “clearly dishonest,” and accused Anthropic of trying to control how people use AI (Tom’s Guide).

An AI-Themed Super Bowl

During the Super Bowl LX final on February 8, artificial intelligence appeared in roughly a quarter of the commercials. OpenAI promoted Codex, Google showed off image generation in Gemini and Meta its smart glasses developed with Oakley; Amazon, Microsoft and Salesforce also ran ads (TechRepublic).

AI Research and Safety

International AI Safety Report 2026

On February 3, the second full International AI Safety Report was published, led by Canadian scientist Yoshua Bengio and written by more than a hundred independent experts. The report notes that in 2025 the best systems reached gold medal level at the International Mathematical Olympiad and can independently handle part of software work (techUK).

The authors warn that some models can recognize when they are being tested and change their behavior. This complicates safety testing, on which today’s regulation and companies’ voluntary commitments rely.

Opus 4.6 Found More Than 500 Unknown Vulnerabilities

During pre-release testing, Claude Opus 4.6 uncovered more than 500 previously unknown security flaws in open-source software, so-called zero-day vulnerabilities. Each of them was verified by a member of the Anthropic team or an external security researcher (Axios, February 5).

Anthropic said it uses the model to find and fix bugs in open source projects because they form the foundation of corporate systems and critical infrastructure (Anthropic Frontier Red Team).

Security and Incidents

Data Leak at Moltbook, the Social Network for AI Agents

On February 2, security firm Wiz reported that Moltbook, a network where AI agents “talk” to each other, had a publicly accessible database. The database key sat in the website’s code and row-level access restrictions were missing, so anyone could read and write data (SiliconANGLE).

According to reports, more than 1.5 million agent API keys, tens of thousands of emails and private messages were exposed. The operator fixed the flaw within a few hours; Wiz co-founder Ami Luttwak called it a typical consequence of so-called vibe coding, that is, programming mostly with AI without thorough review (Neowin).