This Week in AI: March 2–8, 2026 (Week 10)

News of the week: March 2–8, 2026 (Week 10)

The week was dominated by the dispute between Anthropic and the US Department of Defense, which on March 5 formally designated the company a “supply chain risk”. Meanwhile, OpenAI released two new models, GPT-5.3 Instant and GPT-5.4, the latter able to operate a computer on its own. Google launched the cheapest model in the Gemini 3 family, and Nvidia and Broadcom confirmed that investment in AI chips keeps growing. The first figures also emerged on how AI is changing the job market and cybersecurity.

New Models

OpenAI releases GPT-5.4 with computer use

On March 5, OpenAI introduced GPT-5.4, available in ChatGPT, in the API and in the Codex coding tool. Alongside the base version, the company released GPT-5.4 Thinking (a model that “thinks” for longer before answering) and the more powerful GPT-5.4 Pro (TechCrunch).

According to the company, it is the first general-purpose model with built-in computer use: it looks at a screenshot and responds with mouse and keyboard commands. On OSWorld-Verified, a benchmark that measures work in an ordinary desktop environment, it scored 75.0% according to OpenAI, compared with 47.3% for the previous GPT-5.2 and 72.4% for humans (OpenAI).

GPT-5.3 Instant: less lecturing, fewer hallucinations

Two days earlier, on March 3, OpenAI rolled out GPT-5.3 Instant, its new default model for everyday conversations. It is meant to respond more naturally and skip unnecessary warnings and moralizing (TechCrunch).

According to the company, the rate of hallucinations, meaning made-up claims, in sensitive fields (medicine, law, finance) fell by 26.8% when using the web and by 19.7% without it. Plus and Pro subscribers and business customers got the model first (OpenAI).

Google: Gemini 3.1 Flash-Lite

On March 3, Google released a preview of Gemini 3.1 Flash-Lite, the fastest and cheapest member of the Gemini 3 family. It is aimed at developers through Google AI Studio and at businesses through Vertex AI (Google).

The price is $0.25 per million input tokens (a token is roughly part of a word) and $1.50 per million output tokens.
Developers can set how much the model should “think”, which saves money on high-volume tasks such as translation or content moderation (VentureBeat).

Open source: small Qwen3.5 models

On March 2, the Qwen team at China’s Alibaba released four small open Qwen3.5 models ranging from 0.8 to 9 billion parameters under the Apache 2.0 license. The smallest ones are designed for phones and other devices without a cloud connection (Artificial Analysis).

According to Artificial Analysis, the 9B model needs only about 6 GB of memory after compression (4-bit quantization) and leads models under 15 billion parameters on the MMMU-Pro benchmark.

Products and Agents

Codex Security hunts for bugs in code

On March 6, OpenAI launched a research preview of Codex Security, previously known as Aardvark. It is an AI agent that builds a threat model for a project, verifies discovered vulnerabilities in an isolated environment and proposes fixes (OpenAI).

During testing over the past 30 days, it scanned more than 1.2 million commits (saved code changes) and found 792 critical bugs. In open source projects, it has so far led to 14 assigned CVE identifiers, meaning officially registered vulnerabilities.

Claude sees record demand

Following the dispute with the Pentagon, the Claude app climbed to the top spot among free apps in the US App Store. On March 2, however, it reported elevated error rates due to the surge (CNBC).

On March 6, the company said it was seeing more than a million new sign-ups a day and that its daily active users had more than tripled since the start of the year (TechCrunch).
According to data cited by TechCrunch, ChatGPT uninstalls jumped 295% after OpenAI announced its deal with the Pentagon (March 2, TechCrunch).

More product news

March 4: Google made the Canvas feature in its search engine’s AI Mode available to all users in the US. Users can write documents or have a simple tool or dashboard built right in search (TechCrunch).
March 3: Apple unveiled the M5 Pro and M5 Max chips and new MacBook Pro models. Each GPU core has its own Neural Accelerator, so AI performance is up to four times higher than the previous generation, according to the company (Apple).
March 6: Meta will allow rival AI companies to offer chatbots on WhatsApp in Brazil, for a fee. The move follows action by Brazil’s antitrust authority (TechCrunch).

Business and Investment

Nvidia puts $4 billion into optics

On March 2, Nvidia announced it would invest a total of $4 billion in Coherent and Lumentum, two billion in each. Both make photonic components that move data inside data centers using light instead of electricity (CNBC).

For Lumentum, the deal also includes a commitment to buy laser components worth several billion dollars. Shares of both companies rose 12 to 15% on the news.

Broadcom: AI revenue doubles

Broadcom, a maker of custom AI chips, reported results for its first fiscal quarter on March 4. Revenue from AI semiconductors doubled year over year (CNBC).

CEO Hock Tan said revenue from AI chips alone could significantly exceed $100 billion in 2027 (CNBC).

Money and numbers

March 3: Bloomberg reported that Anthropic is nearing an annual revenue run rate of $20 billion. The figure extrapolates current monthly revenue to a full year (Bloomberg).
March 3: According to Crunchbase data, startups raised $189 billion in venture capital in February, most of it going to three companies: OpenAI, Anthropic and Waymo (TechCrunch).
March 5: Lio, a startup that automates corporate procurement, raised $30 million in a round led by Andreessen Horowitz (TechCrunch).

Regulation and Law

Pentagon formally designates Anthropic a risk

On March 5, the US Department of Defense confirmed that it had officially designated Anthropic a “supply chain risk”. It is the first American company to receive such a designation (TechCrunch).

Negotiations broke down over terms of use. Anthropic wanted a guarantee that its models would not be used for fully autonomous weapons or for mass surveillance of Americans. The department insisted it could use them for any lawful purpose (CNBC).

Anthropic argues that, under the law, the designation applies only to the use of Claude in department contracts, not to all customers of military contractors (Anthropic).
March 4: Some defense companies began phasing out Claude (CNBC). At the same time, a tech industry trade group expressed “concern” to Secretary Pete Hegseth (CNBC).
March 5: According to the Financial Times, the two sides returned to the negotiating table (CNBC).

OpenAI amends its Pentagon deal

Sam Altman admitted that the deal OpenAI signed with the Department of Defense at the end of February “looked opportunistic and sloppy”. The company added an explicit ban on using its models to surveil people in the US, including with data bought from brokers (March 3, CNBC, OpenAI).

On March 8, hardware lead Caitlin Kalinowski left the company over the deal. In her view, it was announced without clearly defined safeguards (CNBC).

More legal and political news

March 2: The US Supreme Court declined to hear an appeal by scientist Stephen Thaler. This leaves in place the rule that an image created autonomously by artificial intelligence cannot be protected by copyright, because the author must be human (CNBC).
March 2: Democrats in Congress are preparing a legislative response to the Anthropic dispute, including a ban on federal agencies retaliating against technology contractors (Axios).
March 3: A political action committee funded by the tech industry is spending millions against politician Alex Bores, author of New York’s AI transparency law, the RAISE Act (TechCrunch).
March 6: The White House was preparing to publish a list of state AI laws it considers overly burdensome and is pressing states to hold off on their own regulation until a federal framework is in place (Axios).

Research

Anthropic measures AI’s impact on the job market

On March 5, Anthropic economists published a study with a new metric called “observed exposure”. Rather than tracking what AI could theoretically do, it looks at which work tasks people actually automate with its help (Anthropic).

In computer and math occupations, language models could theoretically handle 94% of tasks, but actual use of Claude covers only 33% of them.
The authors found no systematic rise in unemployment in the most exposed occupations since late 2022.
They do suggest, however, that hiring of young workers in these fields has slowed. According to Fortune, the decline is 14%, though the result is only barely statistically significant (Fortune).

Security and Incidents

Claude finds 22 vulnerabilities in Firefox

On March 6, Anthropic and Mozilla published the results of their collaboration: over two weeks, the Claude Opus 4.6 model found 22 vulnerabilities in the Firefox browser, 14 of which Mozilla rated as high severity. Most were fixed in Firefox 148 (Anthropic, TechCrunch).

The company also examined whether the model could exploit the bugs. Out of hundreds of attempts costing about $4,000, it succeeded only twice, and only in a test environment without a sandbox, the isolation layer that protects a normal browser.

The Claude outage on March 2 was linked to a sharp spike in demand (TechCrunch).