· 07:00 · 7 min + 3 min Explained
GitHub's Copilot can now pick and mix AI models for you
GitHub's Copilot Fusion chooses the models for each coding task, as agents start four in ten pull requests. Also: Google's Gemini 4 Carbon goes to security teams first, big tech's AI money, Claude's spending reports, and Britain's deal for data centres.
Written by AI and checked against our editorial standards. Every claim links to its primary source. Not reviewed by a human editor.
Get it in your inbox at 7am every day ↓
Today in 60 seconds
- GitHub's Copilot Fusion picks and mixes AI models for each coding task. Developers and anyone paying for coding agents.
- Google's Gemini 4 Carbon tops a key ranking, but only cyber defenders get it. Anyone following the race between the big AI labs.
- Microsoft took in $10B from OpenAI, and Google takes AI ads abroad. Anyone wondering whether AI makes money.
- Claude Enterprise now shows what each team's AI spending buys. Managers and finance teams paying the AI bill.
- Britain offers faster grid access to data centres that invest locally. Anyone following where Europe's AI will run.
1. GitHub’s Copilot can now pick and mix AI models for you
The short versionGitHub's new Copilot Fusion chooses and combines AI models for each coding task. On an independent test, it came close to the top setup at about half the cost.
On Wednesday, GitHub said over 40% of pull requests began with an agent’s first draft in September, up from 17% in March. A pull request is a proposed change to a project’s code.
That leaves every team choosing which model to pay for. At its Universe conference, GitHub’s answer was to stop choosing.
1 of 2 Copilot Fusion
In Copilot, you can now pick “Copilot Fusion” instead of one model. You set the effort to Standard, High or Max, and Copilot decides which models do the work. Sometimes a cheap model drafts, and a stronger one steps in only if a check fails.
Artificial Analysis, an independent test group, put Fusion Max about two points behind Claude Code with Opus 5.5 at Max. But Fusion cost about half as much per task, VentureBeat reports. It works on every Copilot plan.
The catch: you still pay each model’s normal rate, and a benchmark is not your code.
2 of 2 Copilot comes with GitHub Enterprise
GitHub Enterprise now includes Copilot for every licensed user, first for customers who buy through a sales team. Instead of fixed AI credits per seat, the company pays for use from one shared budget. If it runs out, code completion still works.
It is like a phone plan that moves from a fixed bundle to paying per call. Meanwhile, the phone picks a good enough network for each call. Developers stop picking models, and the price of a finished task becomes what counts.
Source: github.blog · Official · github.blog · Official · github.blog · Official · venturebeat.com · Reported
2. Google’s Gemini 4 Carbon tops a key ranking, but only security teams get it
The short versionGoogle's new Gemini 4 Carbon ranks first on a widely used model ranking, but for now only cyber defenders can use it. Google also opened Gemini and Search to more outside apps.
1 of 2 Gemini 4 Carbon
New AI models usually go to everyone at once. Carbon, released on Wednesday, goes first to a group of cyber defenders. Developers, firms and consumers come later, and Google gives no date.
The reason is its skill. Google says it scored 88.2% on SEC-bench Pro, a test of finding security holes, and it can also exploit them. On Artificial Analysis’ index of general skill, it ranks first of 52 models.
The catch: the security scores come from Google’s own announcement, measured without safety filters. A model that finds holes this well could also help attackers.
2 of 2 Apps inside Gemini and Search
Google also opened a preview that lets developers put their apps inside Gemini and Search. Apps plug in through MCP, a common standard for connecting AI to outside tools. Over 50 apps already work in Gemini, and for now only US developers can join, TechCrunch reports.
The assistant is becoming the place where work happens. For an app, being found inside Gemini may soon matter as much as ranking high in Google search once did.
Source: blog.google · Official · artificialanalysis.ai · Official · blog.google · Official · techcrunch.com · Reported
3. Microsoft took in $10B from OpenAI, and Google takes AI ads abroad
The short versionNew results show how much money now flows through AI. Microsoft took in $10B from OpenAI in three months, and Google is bringing ads in AI search beyond the US.
Microsoft took in $10B from its deals with OpenAI in the quarter to September, TechCrunch reports from Microsoft’s filing. Google Cloud, which rents out computing power for AI, grew its revenue 103% from a year ago.
Search, Google’s old money machine, grew only 12%, slower than earlier this year. Meanwhile, Google is taking ads in AI Mode, its chat-style search, to other countries for the first time.
The catch: investors were not won over, and Alphabet’s shares fell after the results.
The money is real, and it is moving from the classic list of search ads into AI answers. If you use AI search outside the US, expect ads inside the answers soon.
Source: techcrunch.com · Reported · techcrunch.com · Reported
4. Anthropic shows companies what their Claude spending buys
The short versionClaude Enterprise now reports what each team gets done with Claude and what it costs. Teams can also share one monthly budget.
Spending on coding tools built on Anthropic’s models grew 10x this year, says Jellyfish, which tracks how engineers work, faster than most finance chiefs planned. It is still only about 2% of engineering costs, TechCrunch reports.
On Wednesday, Anthropic added new tools to Claude Enterprise. Smart reports read a sample of a team’s chats and show what the work cost and where it got stuck. With a shared monthly budget, heavy users run out less often while colleagues leave theirs unused.
The catch: this is the seller’s own report on what its product is worth.
Every finance team now asks the same question: what did we get for the AI bill? Whoever answers it best keeps the customer when the budget is cut.
Source: claude.com · Official · techcrunch.com · Reported
5. Britain offers faster grid access to data centres that invest locally
The short versionBritain's budget offers AI data centres a faster grid connection if they invest locally and protect water. The aim is to keep sensitive data in the country.
AI data centres need a lot of power. In Essex, grid delays pushed the power for one planned data centre into the 2030s, and Microsoft dropped its plans for it, Reuters reports.
In Wednesday’s budget, finance minister John Healey offered a deal called the British Data Centre Covenant. Operators that invest locally, create jobs and protect water supplies get priority on the grid, and may build their own power on site.
The catch: Britain estimates it needs to triple its AI-ready capacity by 2030, and the deal itself adds no new capacity.
"When our armed forces use AI to keep us safe, we want that data here," Healey said. Like keeping a country's gold in its own vault, more governments want their AI to run on their own soil.
Source: kfgo.com (Reuters) · Reported
Open source & local AI
NVIDIA’s free model transcribes 40 languages, and llama.cpp learns to talk
The short versionNVIDIA released an open model that turns hour-long recordings in 40 languages into text. And llama.cpp, the engine behind many local AI apps, can now listen and speak.
Meeting notes often go to a cloud service to be typed out. Two releases let you do it yourself.
1 of 2 Nemotron 3.5 Transcribe
NVIDIA’s new model is free, also for commercial use. It handles hour-long recordings and can tell speakers apart. NVIDIA says it ranks first among open models on a common English test.
The catch: for now it is built for big graphics cards, and lighter versions for PCs are still coming. It also only works on finished recordings.
2 of 2 llama.cpp 0.8.0
llama.cpp’s built-in server now does live voice chat and text to speech. It follows the same protocol as OpenAI’s realtime voice service, so tools written for that should be easy to point at your own computer.
Speech is one of the last things most people still send to the cloud. With free tools for both listening and talking, the audio can stay in your office, which matters under EU privacy rules.
Source: huggingface.co · Official · huggingface.co · Official · github.com/ggml-org · Official
Paper of the day
Cheap models can tell why sources disagree, almost like big ones
The short versionWhen an AI's sources disagree, it needs to know why. Researchers show that cheap "decision models" can usually tell, and pass only the hardest cases to a big model.
AI assistants often look things up first. When two sources disagree, the right reply depends on why: maybe one is out of date, or simply false.
Sorting that out usually takes a large model. The authors tested small “decision models” that just pick from a fixed list. Most matched or beat Gemini 2.5 Flash, at a fraction of the cost.
Their best setup is a cascade. The small model handles the easy cases and sends its least certain quarter to the stronger Gemini 3.8 Flash. For most models, that matched Gemini 3.8 Flash at 25 to 37% of its cost.
The catch: the paper is not peer reviewed, and the small models were worse at spotting a planted false source.
It is the same idea as Copilot Fusion in story 1. A family doctor sends only the hard cases to a specialist. Here, a cheap model does the easy part and calls the expensive one only when unsure.
Source: arxiv.org · Official
Explained: if you have a little more time
3 terms from today's news · 3 minWhat is model routing?
Model routing means that a system looks at each question first and decides which AI model should answer it. The system that makes that choice is called a router. AI models differ a lot in price. The strongest ones can cost many times more per word than small ones, but most questions do not need them. A router tries to spot which questions do. > **Think of it like this:** a family doctor sees every patient first. Most go home with simple advice, and only the hard cases are sent on to an expensive specialist. The specialist's time is spent where it makes a difference. There are a few common ways to do it. A simple router guesses from the question itself. It has often been trained on many earlier examples of when a cheap model's answer was good enough. A cascade works the other way round. The cheap model answers first, and a check decides whether a stronger model should take over. Some systems go further and let one model write and another review. It works. In a 2024 study, researchers at LMSYS sent most questions to a cheaper model. On one test, they kept 95% of GPT-4's quality and cut costs by over 85%. The catch is that a router can be wrong. If it sends a hard question to a weak model, the answer gets worse, and you may not notice. **Why it matters to you:** more and more AI tools now choose the model for you, often behind an "auto" setting or an effort level. That can make the same work much cheaper. But it also means you do not always know which model gave you the answer.Two more from today
What is an AI agent?What makes an agent different from a chatbot, and why that changes the risks.
An AI agent is an AI model that does not just answer, but acts. It decides its own next step, uses tools like a browser, a code editor or a payment page, and keeps going until the task is done.
That makes it useful for open-ended jobs. It also means mistakes can pile up, so agents need limits on what they may touch.
What is a benchmark?How AI models are tested and scored, and why a high score is not the whole story.
A benchmark is a fixed test for AI models: a set of tasks plus a way to score the answers. Every model takes the same test, so you can compare the scores.
Scores are useful but limited. A lab often runs the test on its own model, and test questions can leak into the data a model learned from.
How sure are we?Official: from the company, regulator or authorsReported: from a named outletUnconfirmed: rumor or leak