OpenAI released GPT-5.4 (as GPT-5.4 Thinking) and GPT-5.4 Pro across ChatGPT, the API and Codex. The API version supports context up to 1 million tokens — by far OpenAI's largest — and adds tool search, which resolves tool definitions on demand rather than front-loading all of them.

On OSWorld-Verified, the desktop-navigation benchmark, GPT-5.4 hits 75.0% versus 47.3% for GPT-5.2 and above the 72.4% human baseline. GDPval rises to 83.0% and SWE-Bench Pro to 57.7%. OpenAI reports it is 33% less likely to make individual claim errors and 18% less likely to produce an erroneous response than GPT-5.2.

Two days earlier OpenAI had shipped GPT-5.3 Instant, which cuts hallucinations by 26.8% with web search and tones down over-cautious refusal behavior.

Judge Jesus Bernal of the Central District of California denied xAI's motion for a preliminary injunction in X.AI LLC v. Bonta, leaving California's AB 2013 in force.

The law, effective January 1, requires developers of generative AI systems publicly available in California to post a summary of the datasets used to train them. xAI argued the regime took its trade secrets under the Fifth Amendment and compelled speech under the First. The court found xAI had standing but had not shown a likelihood of success, calling its trade-secret pleading "generalized, abstract" and noting xAI had used the word "dataset" with ease throughout its own complaint.

Speaking at the Morgan Stanley Technology, Media & Telecom conference, Jensen Huang said Nvidia's recent investments in OpenAI and Anthropic will likely be its final stakes in both, reasoning that once the two labs go public — expected later this year — the opportunity to buy in privately closes.

Nvidia finalized roughly $30 billion in OpenAI's $110 billion round, well short of the $100 billion figure floated in September, and put about $10 billion into Anthropic. TechCrunch notes the IPO-timing explanation is hard to square with how late-stage investing usually works, and the circular structure of a chip vendor funding its own largest customers remains the standing critique.

Google introduced Gemini 3.1 Flash-Lite, the fastest and most cost-efficient model in the Gemini 3 series, rolling out in preview through the Gemini API, Google AI Studio and Vertex AI.

Pricing is $0.25 per million input tokens and $1.50 per million output — roughly one-eighth the cost of Gemini 3.1 Pro — with configurable thinking levels so one model can be dialed from high-volume classification up to deeper reasoning. Google reports 2.5x faster time-to-first-answer-token and 45% higher output speed than Gemini 2.5 Flash.

Alibaba's Qwen team released the Qwen3.5 Small Model series — 0.8B, 2B, 4B and 9B dense models, Apache 2.0 licensed, 262K context and natively multimodal, with the 9B pitched as competitive with much larger models.

A day later the team's most visible technical leader left: Junyang Lin announced he was stepping down from Qwen, drawing "end of an era" reactions from colleagues and Hugging Face staff. Qwen post-training head Yu Bowen was reported to have resigned as well, and Tongyi Lab held an emergency all-hands. Alibaba Group CEO Eddie Wu approved Lin's resignation on March 5, putting Alibaba Cloud CTO Zhou Jingren in charge of Tongyi Lab and installing ex-Google DeepMind researcher Zhou Hao over post-training. Alibaba shares fell about 4%.