Morning edition — Friday 13 November 2026

On 12 November Google released Gemini 4 Hydrogen, a model built for fast everyday coding and agent work that sits below its frontier models, Gemini 4 Carbon and Gemini 4 Argon. It is available on all tiers of Google AI Studio and the Gemini API, in Gemini for business, the Gemini app and Google Search worldwide, and for paid Google AI Pro and Ultra subscribers in the Antigravity coding tool. API pricing is flat across its 1 million-token context window at $1.50 per million input tokens and $7.50 per million output tokens.

Google says Hydrogen scores 51 on the Artificial Analysis Intelligence Index at its default high thinking level, one point above Argon, and 54 at its highest setting. Compared with Gemini 3.8 Flash, which it replaces, Google says Hydrogen finishes tasks 2.5 times faster, lifts Terminal-Bench 4.1 from 18.2% to 57% and cuts fabricated claims by more than five times.

Google also added Orchestrator to Antigravity, a model option that switches between Carbon and Hydrogen within one session: Carbon plans, Hydrogen writes edits and runs tests. On Terminal-Bench 4.0, Google says Orchestrator came close to Carbon's success rate while using 24% less quota and finishing 19% faster on average.

Microsoft Defender Research and the Microsoft AI Red Team published two studies on 12 November of backdoors planted during post-training in coding and terminal agents. In the first, a small number of crafted examples hidden in a fine-tuning dataset taught a model to find any available command-execution tool and misuse it when a routine, even paraphrased, request appeared. It reached about 78% attack success on familiar harnesses and transferred to an unseen one.

In the second, an attacker who controls training before releasing an open-weight checkpoint combined supervised fine-tuning with reinforcement learning. Average attack success on unseen repositories reached 84.1% to 94.8%, with transfer across coding harnesses. Microsoft recommends controls over training-data pipelines, model adoption and the tools an agent can reach.

This is separate from the Microsoft study this newspaper reported on 11 November, which planted a backdoor by editing a handful of neurons.

Will DePue, a former OpenAI researcher publishing as Unemployment Labs, released a forensic analysis on 12 November arguing that six open-weight models from DeepSeek, Moonshot AI, MiniMax and Ant Group were initialised on Nvidia Blackwell GPUs, which need a US licence for export to China. The models are DeepSeek-V4-Flash, DeepSeek-V4-Pro, Kimi K3, Kimi K3.1-Pro, MiniMax-M3 and Ling-3.1-flash; earlier models from the same labs read as Nvidia's older Hopper generation.

The method recovers traces of each model's starting weights from its published files and searches for the random seed and GPU layout that reproduce them. DePue says that with the recovered seed a Blackwell B300 reproduced 445 rows of DeepSeek-V4-Flash bit for bit, while an H100 reproduced none, and that the method agreed with all 24 models whose developers had disclosed their hardware. He estimates the largest runs used thousands of Blackwell GPUs.

The report says the fingerprint does not show where the chips were, who owned them or whether any law was broken: Blackwell systems can be rented in data centres outside China. The newest readable models from Alibaba, Zhipu, Tencent, Xiaomi and StepFun read as Hopper.

A coalition of four nonprofits, the Irreplaceable Education Fund, Public Citizen, Heat Initiative and the Alliance for Secure AI, sued Anthropic, Google, Meta, OpenAI and SpaceX's AI division in San Francisco Superior Court on Thursday 12 November, the Associated Press reports. The suit asks a judge to stop the companies developing or deploying new AI models; The Register describes the request as covering systems more capable than their most advanced existing models.

The plaintiffs argue that the companies violate California's unfair competition law by endangering public health and interfering with the use of property such as computer systems. The complaint quotes the companies' own leaders on catastrophic risk, including Anthropic's Dario Amodei and SpaceX's Elon Musk, and cites incidents of AI agents attacking systems, among them unauthorised access to an Australian government portal holding Medicare statistics, according to The Register.

Reuters notes that Florida's attorney general asked a judge in September to bar OpenAI from developing new models without outside oversight, a motion that is still pending, and that an OpenAI spokesperson then said the company had paused training its most capable models until more safeguards are in place. The companies named in Thursday's suit did not immediately respond to the AP.

SpaceX said on 12 November that it plans more than $74 billion of investment in two supercomputing campuses: more than $24 billion at Starbase in Texas's Rio Grande Valley and roughly $50 billion in Cochise County, Arizona.

SpaceX says the Texas facility will generate its own electricity without drawing on the public grid, and commits $220 million to Brownsville's water and wastewater systems. It says the Arizona campus will work with local electric cooperatives and use dry cooling with no planned industrial groundwater use. The company projects more than 500 permanent jobs in Texas and about 1,000 in Arizona.

This newspaper's 9 November edition noted a Starbase notice for a proposed $24 billion gas plant and computing site; Thursday's statement is SpaceX's own announcement of both campuses.