Three threads run through today’s issue, and the first is scorecards. Google’s Gemini 4 Carbon leads the rebuilt Artificial Analysis index, NVIDIA and ElevenLabs report wins on tests they ran themselves, and Nathan Lambert asks how much any ranking deserves to be trusted.
The second is who runs the coding agents. Microsoft is renting out the software behind its own Copilot agents, GitHub’s Fusion picks the model for you, and Decagon lets agents tune its servers with a person signing off on every change.
The third is the bill. Anthropic now shows companies what their Claude Code spend buys, while Microsoft trimmed and Alphabet held their AI budgets and the chip landlords took the hit in the stock market.
Ahead on Paper: Scoreboards and Who Keeps Them
Google has a model at the top of the best-known AI ranking that most customers cannot use yet, and one of the ranking’s closest readers has doubts about scoring in general.
- Google’s Gemini 4 Carbon tops an AI ranking but goes to security teams first. Google says Gemini 4 Carbon scores 88.2% on a test of finding software flaws and ranks first on the new Artificial Analysis index, but it starts with a small group of cyber defenders and has no public date. Several of the figures come from Google’s own tests, and the post does not say how the guardrails change them.
- The AI scoreboard everyone cites just changed, and free models fell back. Nathan Lambert says the rebuilt Artificial Analysis index puts eight closed models above the best open-weight one, and that private tests ask for more trust than public benchmarks do. He still expects open labs to jump within two to three months and wants more indices with different biases.
Hear, Time, Compute: Speech Models and One Strange Chip
Two speech releases arrive with strong numbers that mostly come from their own makers, and a developer found a way to run a processor inside a language model.
- NVIDIA’s open speech model transcribes, translates and tells speakers apart. NVIDIA’s open Nemotron 3.5 Transcribe handles 40 languages and up to eight speakers, and NVIDIA also published the data pipeline behind it. Most results are NVIDIA’s own, and the speed figure of 100 hours of audio in under three minutes used a compressed version that has not shipped.
- ElevenLabs’ Cue pins every character of a transcript to the audio. ElevenLabs says its Cue model pins each character of a transcript to the audio, with 179 milliseconds of average timing error on whole recordings and about half a second of compute per hour. The benchmark is the company’s own, and no outside lab has published a comparison.
- A developer built a working computer inside an AI model’s weights. Divyansh Agrawal compiled a RISC-V processor into the weights of Alibaba’s Qwen 3.5 4B, so the model does exact arithmetic without a calculator tool and went from 48.6 to 100 percent on a long-integer test. It is slow, the checkpoint is 18.8 GB, and the checks are his own.
Who Drives the Agents: Renting, Routing and Self-Tuning
Three companies are taking the hard parts of running AI agents off your plate, and each leaves a cost or a judgment call in someone else’s hands.
- Microsoft rents out the software that runs its own Copilot agents. Microsoft Foundry now rents out the software that runs GitHub Copilot and Copilot Studio, with more than 20 models to choose from. The post gives no price for the hosted version and no comparison for what it calls industry-leading performance.
- GitHub Copilot now picks the AI model for you and bills for each one. GitHub’s Copilot Fusion sends each coding task to one or more models, and GitHub says its top level scored two points behind Claude Code at about half the API spend. Those figures are GitHub’s own reading of an Artificial Analysis index, and every model Fusion calls is billed at its standard rate.
- Decagon’s AI agents tune its model servers and find gains up to 2.5x. Decagon says an agent loop that tests changes on replayed production traffic lifted throughput per GPU by 1.05x to 2.50x across its endpoints, with a person approving every change that ships. Every figure is Decagon’s own, and its table does not name the endpoints.
Paying Up: Claude Code Bills and Cloud Budgets
AI spending is getting its first real audit, from the customers buying the tools and from the companies building the data centers.
- Anthropic starts showing companies what their Claude Code spend buys. Anthropic released reports that tie a team’s Claude Code output to its cost, as Jellyfish data shows spending on Anthropic-based tools up 10x this year to roughly 2 percent of engineering cost. TechCrunch links the launch to a planned IPO, and its account does not say how output is scored.
- Google and Microsoft stop raising AI spending plans, and cloud stocks slip. Microsoft cut its full-year capex forecast from $175 billion to $170 billion and Alphabet held its range, while CoreWeave, Nebius and IREN shares fell on delivery delays. Microsoft’s finance chief says spending should pick up in early 2027 because components cost more and some deliveries are late.
Today’s Quick Hits
- Vercel’s new CI only reruns the checks your change touches. Vercel opened a beta of Vercel CI, which reads a project’s Turborepo setup, reruns only the checks a change touches and pulls the rest from cache. Vercel says its own pipeline is now two to six times faster, a figure that is the company’s own.
- Meta quietly launches a pixel pet app, starting in the Philippines. Meta put Waddle World, a virtual pet app where users hatch, feed and talk to a pixel pet, into app stores in the Philippines. TechCrunch reports it is rated for adults and has no streaks, and Meta has not said where it goes next.
- GitHub’s Copilot Automations can now rewrite their own instructions. GitHub says Copilot Automations can now study how people react to their output and draft changes to their own prompts, tools or triggers. A person accepts or rejects each proposal, and the changelog offers no results showing the loop improves quality.
- OpenAI’s Dot assistants now take instructions by text message. OpenAI is previewing text-message access to Dot agents for US Pro subscribers, over iMessage on iPhone and RCS on Android, and a Dot can start the conversation itself. Engadget notes Plus subscribers are not included, and memory is shared with ChatGPT.
- Google opens one front door for apps inside Search and Gemini. Google opened a preview of its Connected Apps Platform, where outside developers build an app once and choose Search, the Gemini app or both. Google says more than 50 Connected Apps exist, and the post gives no dates, markets or developer counts.
- Mirendil opens a private beta of Lor, an AI built to run research. Mirendil put Lor, which it describes as software that does the work of a whole AI lab, into private beta. Its post on X claims the agents already write code and run experiments for days, and it links no benchmark, pricing or access terms.