Quick answer
On 16 November 2026 OpenAI signed agreements letting four independent safety organisations, Apollo Research, METR, Transluce and Goodfire, work inside the company with deep access to its internal models, training data and monitoring systems. It is OpenAI’s first phase of “embedded assessments”, and it landed while everyone was waiting for GPT-6.2 Astra (which arrived on 18 November). The same news cycle brought Tencent’s very cheap Hy4 Flash model, a detailed post on how Anthropic’s Claude Code team verifies AI-written code, and a Gemini Notebook feature that turns your notes into interactive websites.
Key takeaways
- OpenAI’s safety pilots are official. Outside experts get access to model checkpoints, source code and monitoring records, and some will publish findings.
- GPT-6.2 Astra previews were leaks, reportedly via unofficial API resellers. OpenAI released the model officially two days later.
- Tencent’s Hy4 Flash costs about $0.07 per million input tokens and, in Tencent’s own blind test, came close to Claude Haiku 5.5 on user satisfaction.
- Anthropic’s Claude Code team uses a “verify” skill, automatic test loops and a 39-item reliability checklist, and calls in a human after three failed fix attempts.
- Gemini Notebook can now build shareable mini websites from your sources when you type “make me a site”.
🛠️ Practical guides
Ready to put this into practice? These step-by-step guides answer two common follow-up questions:
- How to Make Claude Code Test Its Own Work With a Verify Skill — set up a simple verify skill so Claude Code runs and checks your app before it says a job is done.
- How to Turn Your Notes Into an Interactive Website With Gemini Notebook — turn a folder of notes, PDFs or sales figures into a clickable explainer, tracker or dashboard you can share.
In this guide
- Watch: the source video
- What are OpenAI’s new embedded safety assessments?
- Why was everyone waiting for GPT-6.2 Astra?
- What is Tencent’s Hy4 Flash?
- How does the Claude Code team check AI-written code?
- What can Gemini Notebook build now?
- What else happened in AI this week?
- What is confirmed and what is only claimed?
- What should you do about this news?
- Frequently asked questions
- The bottom line
- Related posts
This was a week where the biggest AI story was not a model launch but a promise about how models are checked before they reach you. Below, we separate what the companies actually announced from what was leaked, rumoured or simply speculated about, and pull out the practical bits you can use straight away.
This round-up is based on the 17 November news video from WorldofAI, with each major claim checked against official announcements where we could find them. Because we are publishing a few days later, we have noted where things have moved on since.
Watch: the source video
What this video covers: WorldofAI’s AI news round-up for 17 November: leaked GPT-6.2 Astra previews and a hidden benchmark, reports of OpenAI release delays and its new safety partnerships, Gemini 4 Neon and Oxygen spotted in logs, Tencent’s Hy4 Flash, Microsoft’s expected MAI Code 2 Flash, Hugging Face models on OpenRouter, a Codex capacity-error fix, Claude Code updates and verification, OpenAI’s science and healthcare news, Gemini Notebook sites, bigger Antigravity context windows, Sam Altman’s reported AGI remark, OpenCode Browse and an AI treaty march.
Video by WorldofAI (@intheworldofai). The video was sponsored by Mammouth, an AI platform that bundles several models from $10 a month. We have no link with the sponsor.
What are OpenAI’s new embedded safety assessments?
Embedded assessments let outside safety experts work inside OpenAI, with access to some of the same internal resources its own researchers use. OpenAI announced the first phase on 16 November 2026 with four organisations:
| Organisation | What it will look at |
|---|---|
| Apollo Research | Whether models could secretly work against their developers’ intentions, and whether OpenAI’s safeguards would catch it |
| METR | Misaligned behaviour in OpenAI’s largest reinforcement learning training run to date, including failures existing tests may have missed |
| Transluce | How well OpenAI’s internal tests catch deception, manipulation and “reward hacking” (a six-week pilot starting in early December) |
| Goodfire | Using interpretability tools to monitor training runs and debug unwanted behaviour |
Access can include model checkpoints, source code, training and evaluation data and monitoring systems. Transluce says it is not accepting payment from OpenAI and will report publicly what access it asked for and what was refused. OpenAI links the work to commitments it made under the White House Accord on Super Intelligence.
Why it matters to you: if you build your business on AI tools, independent checks are one of the few ways to know whether a provider’s safety claims hold up. For background on one of the partners, see AI Interpretability: Why Goodfire’s Research Is Good News for Safety.
💡 Tip: Outside testers don’t replace your own checks. Before relying on any model for customer-facing work, run a quick test of your own. Our guide How to Test an AI Model for Made-Up Answers Before You Rely on It shows you how.
Why was everyone waiting for GPT-6.2 Astra?
Because about 75 days had passed since GPT-6 Astra, longer than OpenAI’s usual gap between flagship releases. The creator showed demos said to come from early GPT-6.2 Astra access, including a detailed dungeon-crawler game and a 3D harbour scene built entirely in code.
🔎 Unconfirmed / rumour: The previews were reportedly obtained through Chinese API relay services, which resell access to models. A “hidden” benchmark result (89% versus 67% for GPT-6 Astra on a biology test) was said to have been found in the code of an OpenAI web page. The creator said he had not verified it. Reports that OpenAI delayed Astra over safety worries also came from unnamed sources and deleted posts. Treat all of this as unconfirmed.
What has happened since: OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 18 November. For a proper look at the released model, read our GPT-6.2 Astra Review: Cheaper and Efficient, but Is It Worth Switching?.
⚠️ Warning: Never use unofficial API “relay” services to get early model access. You have no idea who sees your prompts or data. Our guide How to Spot a Fake AI Model Leak Before You Download Anything covers the warning signs.
What is Tencent’s Hy4 Flash?
Hy4 Flash is a low-cost AI model from Tencent, released on 16 November 2026, aimed at everyday office tasks rather than heavy coding. Here is what Tencent announced:
| Detail | What Tencent says |
|---|---|
| Price (list) | About $0.07 per million input tokens and $0.281 per million output tokens |
| Launch offer | 50% off via Tencent Cloud until 30 November; free in its WorkBuddy app until then |
| Size | 224 billion parameters in total, only 12 billion active at a time |
| Context window | Up to 1 million tokens, with image and other multimodal input |
| Blind test | 85 internal users, 259 tasks: satisfaction close to Claude Haiku 5.5, at a much lower cost per task |
| Open weights | Tencent plans to release them |
In the creator’s quick tests, Hy4 Flash did a nice job of a simple animated web scene but made mistakes on a detailed SVG drawing compared with Qwen 4 Max. Tencent itself recommends its bigger Hy4 model for complex software work.
Our take: very cheap models are worth testing for repetitive jobs like summarising documents or tidying spreadsheets. Use our guide How to Test Whether a Cheaper AI Model Is Good Enough for Your Work, and remember to check where a provider stores and processes your data before sending anything sensitive.
How does the Claude Code team check AI-written code?
Anthropic’s Claude Code team makes Claude prove its code works before it is committed, rather than relying on more and more AI reviewers. The team set this out in a blog post published on 16 November. The main ideas:
- Run the app, don’t just read the code. Claude tests changes end to end, the way a person would, in a browser, simulator or local server.
- A “verify” skill stored in the project tells Claude exactly how to start and smoke-test the app before committing.
- Three strikes, then a human. Claude tries to fix failing checks itself; if a check still fails after three rounds, it stops and asks a person.
- A 39-item reliability checklist, each item based on a past bug or incident, triggered only when a change touches the relevant area.
- Different kinds of tests: end-to-end tests of real user journeys, “property” tests that throw thousands of random inputs at a rule, and comparisons between old and new versions.
The creator said Claude now writes and reviews nearly 90% of the team’s code, which is why verification matters so much. You don’t need a big team to borrow the idea. Our guide How to Make Claude Code Test Its Own Work With a Verify Skill shows how to set up a simple version.
Also in Claude Code this week, the creator reported that the desktop app can now test apps on a foldable iPhone model in Apple’s iOS simulator, in public beta.
Try this today: next time an AI tool says “done”, ask it to show you the app running and to list exactly what it tested. If it can’t, it hasn’t verified anything.
What can Gemini Notebook build now?
Gemini Notebook can now turn your sources into interactive websites, such as explainers, dashboards and project trackers. According to Google’s 16 November announcement, you type “make me a site” in the notebook chat with a description, then open the result in a new tab or share it. It is rolling out on the web to users aged 18 and over, including Google Workspace customers, with mobile to follow.
For small businesses this is a quick way to make a client-facing explainer, a sales dashboard from a few documents or a simple project tracker. Our step-by-step guide How to Turn Your Notes Into an Interactive Website With Gemini Notebook shows you how, including how to check it for mistakes.
What else happened in AI this week?
- Gemini 4 Neon and Gemini 4 Oxygen were spotted in Google’s back-end logs. Neon has since launched; read our Gemini 4 Neon Review: What the Tests Show on Speed, Cost and Quality. Oxygen, possibly a new Flash model, is still unconfirmed.
- Codex capacity errors fell sharply. A chart shared in the video showed the share of daily Codex users hitting a capacity error dropping from a peak of 6.89% to about 0.085% by 12 November, after OpenAI improved how it handles retries and overloads.
- Microsoft’s MAI Code 2 Flash appeared in a leaked model card ahead of Microsoft Ignite.
- Hugging Face and OpenRouter were reported to have linked up, so Hugging Face models can be routed through OpenRouter with automatic failover.
- OpenAI shared research in which more than 1,800 scientists reported nearly doubling their productivity with ChatGPT. These are self-reported figures.
- OpenAI expanded into healthcare admin, with computer-use agents and more than 20 healthcare plugins used by organisations including Abridge, Ambience, Cigna and CVS Health.
- Antigravity context windows for Gemini 4 Carbon, Hydrogen and Argon were reported to have doubled from 256K to 512K tokens.
- OpenCode Browse 0.1.0, a browser tool for AI agents rewritten in Rust, claims 25 times lower memory use.
- Nvidia: CNBC reported the board is not considering bringing Jensen Huang back as CEO. Background: Jensen Huang Resignation: What It Could Mean for Your AI Bill.
- An AI treaty march brought over 1,000 people onto the streets in six countries, described in the video as the largest AI safety protest so far.
🔎 Unconfirmed / rumour: Sam Altman was reported to have said in an interview, “We didn’t have AGI. Now we do.” The creator stressed that the context was unclear. An Anthropic researcher was also shown predicting “broadly superhuman AI” in about 1.3 years. These are opinions and remarks, not evidence; treat them with caution.
What is confirmed and what is only claimed?
| Claim | Status | Source |
|---|---|---|
| OpenAI embedded assessments with Apollo, METR, Transluce and Goodfire | Confirmed | OpenAI and partner announcements |
| Tencent Hy4 Flash price and specs | Confirmed | Tencent |
| Claude Code team verification approach | Confirmed | Anthropic blog |
| Gemini Notebook interactive sites | Confirmed | |
| Early GPT-6.2 Astra previews via relay services | Unconfirmed leak | Creator and viewers |
| Hidden 89% Astra benchmark | Unverified | Social media report |
| Astra delayed for safety reasons | Unconfirmed | Unnamed reports |
| Gemini 4 Oxygen | Unconfirmed | Back-end logs |
| Altman “now we do” AGI remark | Unclear context | Reported interview quote |
What should you do about this news?
- Add a “prove it” step to AI work. Whether it is code, a report or a spreadsheet, ask the AI to show its checks, not just its answer.
- Test a cheap model on one repetitive job. Hy4 Flash and similar models could cut costs for routine tasks, if the quality holds up for you.
- Try Gemini Notebook sites for one client explainer or internal dashboard, and check every fact before sharing.
- Ignore leaked-model shortcuts. Wait for official releases rather than using unofficial resellers.
Your next step: pick one AI task you rely on every week and write down how you would know if the AI got it wrong. That one sentence is the start of your own verification checklist.
Frequently asked questions
What are OpenAI’s embedded assessments?
A pilot in which four independent safety organisations, Apollo Research, METR, Transluce and Goodfire, get deep access inside OpenAI to assess model alignment and safeguards. OpenAI announced it on 16 November 2026.
Has GPT-6.2 Astra been released?
Yes. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 18 November 2026, two days after this news round-up.
How much does Tencent Hy4 Flash cost?
The list price is about $0.07 per million input tokens and $0.281 per million output tokens, with 50% off via Tencent Cloud until 30 November 2026.
What is a verify skill in Claude Code?
A set of instructions saved in your project that tells Claude how to start your app and smoke-test it, so it checks its work before committing changes.
Is Gemini Notebook’s website feature free?
Google says it is rolling out on the web to Gemini Notebook users aged 18 and over, including Workspace customers. Check your own account, as rollouts can take a couple of weeks.
Was the video sponsored?
Yes, by Mammouth, an AI platform that bundles several models from $10 a month.
The bottom line
The most useful idea this week is simple: AI work needs checking, and the companies are starting to say so out loud. OpenAI is letting outsiders inspect its models, and Anthropic is making Claude prove its code runs. You can apply the same principle on a small scale today, by asking every AI tool you use to show its working before you trust the result.
Related posts
- GPT-6.2 Astra Review: Cheaper and Efficient, but Is It Worth Switching?
- AI Interpretability: Why Goodfire’s Research Is Good News for Safety
- Gemini 4 Neon Review: What the Tests Show on Speed, Cost and Quality
