This Week in AI: December 15-21, 2025 (Week 51)

News of the week: December 15-21, 2025 (Week 51)

The week before Christmas brought a dense flurry of news. Google rolled out its fast Gemini 3 Flash model as the default in the Gemini app and in Search, and OpenAI answered with a new image generator, the GPT-5.2-Codex model for programmers, and an app directory right inside ChatGPT. In finance, the headlines were dominated by OpenAI’s talks with Amazon and investors at a valuation of around 750 billion dollars, a record funding round for Databricks, and Micron’s results, which confirmed the hunger for memory chips. New York passed its own law on the safety of the most powerful models, and a power outage in San Francisco exposed the limits of robotaxis.

New Models

Google released Gemini 3 Flash and made it the default model

On December 17, Google introduced Gemini 3 Flash, a faster and cheaper sibling of November’s Gemini 3 Pro. According to Google, it offers Pro-level reasoning with the speed and price of the Flash line. The company says it outperforms the older Gemini 2.5 Pro while being roughly three times faster. (Google Blog)

On the SWE-bench Verified coding benchmark (fixing real bugs in open source projects), it scored 78%, according to Google.
It became the default model in the Gemini app, replacing version 2.5 Flash, and also in AI Mode in Google Search.
Developers pay $0.50 per million input and $3 per million output tokens (tokens are the chunks of text that models count in).

Nvidia bet on open Nemotron 3 models and bought SchedMD

On December 15, Nvidia announced the Nemotron 3 family of open models in three sizes: Nano, Super, and Ultra. The smallest, Nemotron 3 Nano, has 30 billion parameters, of which only about 3 billion are active at each step, and it can handle a context of up to one million tokens. It is available for download now, with the larger versions due in the first half of 2026. (TechCrunch)

At the same time, Nvidia acquired SchedMD, the company behind Slurm. Slurm is an open source job scheduler that distributes work across thousands of processors in supercomputers and is used by more than half of the world’s most powerful machines. The price was not disclosed, and Nvidia promised that Slurm will remain open source. (TechCrunch)

OpenAI: a new image generator and GPT-5.2-Codex

On December 16, OpenAI launched the GPT Image 1.5 model and a redesigned ChatGPT Images section. According to the company, images are generated up to four times faster, edits preserve details better, and the model can even handle small, dense text. In the developer interface (API), it is 20% cheaper than its predecessor. (Neowin)

Two days later, on December 18, came GPT-5.2-Codex, a version of the new GPT-5.2 model tuned for autonomous coding work and security tasks. Paying ChatGPT users got it in the Codex tools for the command line, development environments, and the cloud. (Gigazine)

Other models of the week

December 16 – Meta released SAM Audio, a model that extracts just the desired sound from a mix, such as a voice or a dog barking. The request can be typed as text, pointed out in a video, or marked as a time segment. (Gigazine)
December 17 – France’s Mistral introduced Mistral OCR 3 for converting scanned documents, tables, and handwriting into text. According to the company, it beats the previous version in 74% of comparisons. (Mistral AI)
December 18 – Google released FunctionGemma, a tiny open model with 270 million parameters. It runs directly on a phone and turns everyday sentences into specific app actions. (TestingCatalog)
December 19 – According to The Wall Street Journal, Meta is preparing an image and video model code-named Mango and a text model called Avocado for the first half of 2026. (TechCrunch)

Products and Agents

ChatGPT got an app directory

On December 17, OpenAI opened the door for developers to submit their own apps to ChatGPT for approval. Users will find them in a new app directory in the tools menu or at chatgpt.com/apps. The first approved apps are expected to roll out gradually from the start of the new year. (OpenAI)

Anthropic made Agent Skills an open standard

On December 18, Anthropic published the Agent Skills specification as an open standard. “Skills” are bundles of instructions, scripts, and resources that an AI agent loads when it needs to handle a specific task. Thanks to the opening, tools other than Claude can use them too. (AI Business)

The company also offered ready-made skills from partners, including Canva, Notion, Figma, and Atlassian.
Companies on the Team and Enterprise plans can manage skills centrally for all employees. (SiliconANGLE)
The move follows early December, when Anthropic handed the MCP protocol over to the Linux Foundation.

New ChatGPT rules for teenagers

On December 18, OpenAI added special principles for users aged 13 to 17 to its Model Spec, the document describing how its models should behave. For these users, safety takes priority even at the cost of limiting features, and the model should point them to help from trusted adults. The American Psychological Association contributed to the proposal. (OpenAI)

Business and Investment

OpenAI is in funding talks with Amazon and other investors

Reuters confirmed on December 17 that Amazon is in talks to invest roughly 10 billion dollars in OpenAI. According to The Information, OpenAI would use Amazon’s Trainium chips as part of the deal. A source described the talks as “very fluid,” and they may not end in an agreement. (RTE / Reuters)

Then on December 18, Bloomberg, citing The Information, reported that OpenAI is in preliminary talks to raise tens of billions of dollars at a valuation of around 750 billion dollars. That would be a jump of roughly 50% from the October valuation of 500 billion. (Bloomberg)

Databricks is worth 134 billion dollars

The data and AI platform Databricks announced on December 16 that it raised more than 4 billion dollars at a valuation of 134 billion. The round was led by Insight Partners, Fidelity, and J.P. Morgan Asset Management. The company reports an annual revenue run rate of 4.8 billion dollars, growing more than 55%. (The Star / Reuters)

Micron: demand for AI memory outpaces production

Memory maker Micron reported quarterly revenue of 13.6 billion dollars on December 17, up 57% from a year earlier. For the next quarter it expects revenue of around 18.7 billion, well above analyst estimates. The main driver is HBM, the fast memory stacked right next to AI chips. (SiliconANGLE)

More business news

December 17 – Amazon CEO Andy Jassy announced that AGI team head Rohit Prasad will leave at the end of the year. A new group under Peter DeSantis will bring together the Nova models, custom chips, and quantum computing. (GeekWire)
December 17 – Oracle shares fell about 5% after a report that investor Blue Owl will not back a roughly 10-billion-dollar data center project in Michigan for the Stargate project. Oracle says talks with another partner are on track. (Yahoo Finance / Reuters)
December 19 – The maker of the Cursor code editor acquired Graphite, a startup focused on code review. The price was not disclosed. (Fortune)
December 19 – Resolve AI, a startup founded by former Splunk executives, reached a valuation of 1 billion dollars in its first major round (Series A). (TechCrunch)

Regulation and Policy

New York passed the RAISE Act

Governor Kathy Hochul signed the RAISE Act on the safety of the most powerful AI models on December 19. This makes New York the second US state after California with comprehensive regulation of the sector. It came just one week after President Donald Trump signed an executive order intended to limit state AI regulation. (Axios)

Large developers must publish their safety protocols and report serious incidents within 72 hours.
Oversight will be handled by a new office within New York’s Department of Financial Services.

The Genesis Mission has 24 partners

The US Department of Energy announced memoranda with 24 organizations for the Genesis Mission, a national program for using AI in science, on December 18. The partners include OpenAI, Anthropic, Google, Microsoft, Nvidia, AMD, and AWS, among others. The goal is to connect their technologies with the computing power of 17 national laboratories. (insideHPC)

December 19 – According to Reuters, the Trump administration launched an interagency review of applications to export Nvidia H200 chips to China. Trump previously said the government would take a 25% cut of these sales. (investingLive)

Safety and Incidents

A power outage stopped Waymo robotaxis

On Saturday, December 20, a fire at a PG&E substation cut power to roughly one hundred thousand customers in San Francisco. Dozens of Waymo robotaxis stalled at intersections with dead traffic lights and snarled traffic. Waymo suspended service overnight and did not restore it until Sunday afternoon. (Gizmodo)

The incident showed that autonomous cars cannot yet handle large-scale infrastructure failures as smoothly as a human driver. The California regulator said it is looking into the events. (Gizmodo)