Gemini 4 Neon lands, Copilot becomes the new OS for work, and OpenAI opens its doors to auditors

Compact Conversations for 2026-11-18: 6 AI stories, ai news worth knowing in just 5 minutes.

The Lead: Google ships Gemini 4 Neon, its fourth model in seven weeks

Google introduced Gemini 4 Neon, positioned for fast end-to-end task completion across coding, knowledge work, and science. It scores 61 on the Artificial Analysis Intelligence Index at its top thinking level, ranking third overall behind larger, slower models. It’s priced at $3 per million input tokens and $15 per million output tokens with a 1-million-token context window, and it’s rolling out in the Gemini API, AI Studio, Google Antigravity, Gemini for business, and the Gemini app. Google says Carbon handles the hardest long-horizon projects, Hydrogen does fast everyday execution, and Neon sits in between, and it recorded its lowest fabrication rate yet in internal Antigravity testing.

Why it matters: Platform teams evaluating model tiers get frontier-level reasoning at $3/$15 per million tokens, with Neon able to orchestrate parallel Hydrogen subagents on long, tool-heavy jobs.

Source: Google

Number to Know: OpenAI agents now complete 46% of researchers’ 2-to-4-hour tasks on their own

In a new report, OpenAI says its coding agents now complete 46% of researchers’ two-to-four-hour tasks entirely on their own, up from 23% in March. The median researcher’s daily agent spend hit $2,400 in October, up 62% from September. OpenAI itself calls the uplift jagged: monitoring training runs succeeds 79% of the time, but writing inference and serving code only 25%, so humans stay firmly in the loop.

Why it matters: Developers benchmarking agent autonomy should note the split OpenAI reports: 79% success on monitoring training runs versus 25% on writing inference and serving code.

Source: OpenAI

The Feed

Word, Excel, and PowerPoint go generally available inside Copilot

At Microsoft’s Ignite keynote, Satya Nadella announced general availability of Word, Excel, and PowerPoint inside Copilot, calling it “the new OS for work.” The vision inverts how Office is used: agents drive the work and the apps ride along as components, backed by a knowledge layer drawing on email, meetings, chats, and files. Files created in Copilot stay editable in the Office apps. Copilot Autopilot is a persistent cloud agent you can hand responsibilities to, billed through Copilot Credits with pricing unannounced. Microsoft also made Microsoft IQ generally available, including the Work IQ context layer Nadella described as maybe a company’s most valuable database; whether agents get access to that internal data is up to each company’s own IT review.

Source: Ars Technica

OpenAI opens its systems to four independent on-site auditors for the first time

OpenAI is giving four independent audit organizations, including Apollo Research and METR, direct on-site access to unreleased models, source code, and training data for the first time. The embedded reviews are designed to uncover hidden misbehavior and safety gaps. The reporting flags an open question: OpenAI can redact or delay the reports, so how independent this oversight actually turns out to be remains to be seen.

Why it matters: Security and governance leads watching frontier-lab oversight should note that OpenAI retains the right to redact or delay the auditors’ reports.

Source: The Decoder

Nvidia halts H200 AI chip sales to China

On its earnings call, interim Nvidia CEO Colette Kress said Nvidia is no longer shipping licensed Hopper data center products to China and that its guidance includes no future licensed shipments. She didn’t explain why. Reuters reported earlier this month that around 600,000 H200 chips were stuck in Hong Kong awaiting Beijing’s approval.

Source: The Verge

KPMG runs a parallel audit with 100+ AI agents

The Financial Times reports KPMG ran a pilot parallel audit of a FTSE 250 company using more than 100 AI agents, and says the agents asked better questions and dug deeper than human staff.

Why it matters: Audit and finance teams at large companies now have a working precedent: KPMG’s pilot used over 100 AI agents in a parallel audit of a FTSE 250 firm.

Source: Financial Times

One Thing to Try

Monomem is a shared memory file for coding agents: tell one agent something once, say Claude Code, and Cursor or Codex can find it later, because they all read and write one SQLite file on your own machine. No cloud, no API keys, no hidden model calls. Each note records which app wrote it and keeps earlier versions so you can roll back, and if a write looks like a planted instruction or contains a secret, it’s held for your approval before other agents see it. One download, nothing else to install, MIT licensed.

Sources

Transcript

Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.

Host B: Since we covered the Gemini 4 family on Thursday, Google has added another model. Today it shipped Gemini 4 Neon, the fourth model in about seven weeks. Google positions Neon for fast end-to-end task completion across coding, knowledge work, and science, and it’s rolling out in the Gemini API, AI Studio, Google Antigravity, Gemini for business, and the Gemini app. Host A: What does that mean for a normal team? Is this a new tier or a replacement? Host B: The source doesn’t frame it as a replacement. Google says Carbon handles the hardest long-horizon projects, Hydrogen does fast everyday execution, and Neon sits in between. On the Artificial Analysis Intelligence Index at its top thinking level, Neon scores 61 and ranks third overall, just behind larger, slower models. It’s priced at three dollars per million input tokens and fifteen per million output tokens, with a one-million-token context window. Host A: Google also says Neon recorded its lowest fabrication rate yet in internal Antigravity testing, and it can act as a fast orchestrator directing parallel Hydrogen subagents on long, tool-heavy jobs.

Host B: Here’s a number worth sitting with: forty-six percent. In a new report, OpenAI says its coding agents now complete forty-six percent of researchers’ two-to-four-hour tasks entirely on their own, up from twenty-three percent in March. The median researcher’s daily agent spend hit two thousand four hundred dollars in October, up sixty-two percent from September. OpenAI itself calls the uplift jagged, though. Monitoring training runs succeeds seventy-nine percent of the time, but writing inference and serving code only twenty-five percent, so humans stay firmly in the loop.

Host A: From Microsoft’s Ignite keynote: Word, Excel, and PowerPoint are now generally available inside Copilot, which Satya Nadella called the new OS for work. The idea inverts how we’ve used Office. Instead of opening Word and asking an assistant for help, agents drive the work and the apps ride along as components, backed by a knowledge layer that draws on email, meetings, chats, and files. Files you create in Copilot stay editable in the Office apps. Host B: There’s more underneath that. Copilot Autopilot is a persistent cloud agent you can hand responsibilities to, not just single tasks, billed through Copilot Credits with pricing still unannounced. Microsoft also made Microsoft IQ generally available, including the Work IQ context layer Nadella described as maybe a company’s most valuable database. Ars Technica notes whether agents get access to that internal data is up to each company’s own IT review. Host A: Staying with OpenAI: The Decoder reports the company is letting four independent audit organizations, including Apollo Research and METR, work on-site with unreleased models, source code, and training data for the first time. The reviews are meant to surface hidden misbehavior and safety gaps. One open question the reporting flags: OpenAI can redact or delay the reports, so how independent this oversight really is remains to be seen. Host B: As a follow-up to the export-control stories we covered Monday and Tuesday, Nvidia has now formally halted H200 shipments to China. On its earnings call, interim CEO Colette Kress said Nvidia is no longer shipping licensed Hopper data center products there, and its guidance includes no future licensed shipments. She didn’t explain why, and Reuters reported earlier this month that around six hundred thousand H200 chips were stuck in Hong Kong awaiting Beijing’s approval. Host A: And from the Financial Times: KPMG ran a pilot parallel audit of a FTSE 250 company using more than a hundred AI agents, and says the agents asked better questions and dug deeper than human staff.

Host B: For something to try, there’s a project called Monomem that’s been getting attention on Reddit. It’s a shared memory file for coding agents. The idea: you tell one agent something once, say Claude Code, and Cursor or Codex can find it later, because they all read and write one SQLite file on your own machine. No cloud, no API keys, no hidden model calls. Host A: Each note records which app wrote it and keeps earlier versions, so you can roll back. And if a write looks like a planted instruction or contains a secret, it’s held for your approval before other agents ever see it. One download, nothing else to install, and it’s MIT licensed. Worth a spin if you run more than one agent on the same codebase.

Host B: Wednesday, handled. Host A: That’s Compact Conversations for today. We’ll be back tomorrow with more AI news worth knowing. Until then, happy prompting.