A ton of AI news from Anthropic, Google, OpenAI, and Nvidia this week.
Here's everything going on in AI right now:
1. Anthropic priced its IPO at $83 a share, selling 980 million shares and raising $81.3 billion. That's the largest IPO on record, ahead of SpaceX's $75 billion debut in June. It trades on Nasdaq under the ticker ANT, at a valuation of around $1.5 trillion at the IPO price.
2. Google released Gemini 4 Neon. It ranks #3 on Artificial Analysis' Intelligence Index with a score of 61, just behind larger, slower models. It's rolling out now in Google's API and to Google AI Pro and Ultra subscribers.
3. Nvidia interim CEO Colette Kress said the company is no longer shipping licensed Hopper AI chips to China, even though the U.S. had approved those sales. China Hopper sales were $2.4 billion last quarter, but they're now out of Nvidia's forecast. The more advanced Blackwell and Rubin chips were already banned.
4. OpenAI showed how its models are being used to design chips. One change to its own Jalapeño chip cut a multiplier's estimated area by about 15% and shortened a calculation from two clock cycles to one. Working with AMD, the models found up to 14% gains in power, performance, and area on flagship GPU designs in under 5 months, and AMD has already sent some of those changes off to be manufactured.
5. At the New York Times DealBook Summit, Sam Altman said OpenAI still plans to go public. "We plan to be public," he said, without giving a date.
6. Citi updated its terms so customers who give third-party AI tools access to their bank accounts are responsible for what those tools do, even if the tool is hacked or leaks their login. JPMorgan and Bank of America already had similar rules, and PNC changed its terms this week too.
7. Google opened Genie 4 to eligible AI Ultra subscribers. You can generate an interactive world in your browser, then type a new prompt while you're walking around in it to change the weather or add a character, now with sound that follows the scene.
8. School districts in New York City, Los Angeles, and Chicago are looking at alternatives to Google Search after Google said schools that turn off its AI features will also lose Search next year. Google says Search and AI are "deeply integrated," and that next year Gemini and Search will be off by default until schools choose what to turn on.
9. OpenAI launched ChatGPT for Accounting, which connects to tools like QuickBooks, Xero, and NetSuite to draft invoices, match records, and prepare reports. A person still has to review and approve the work before anything is filed.
10. OpenAI also said its coding agents can now finish 46% of internal research tasks that would take a person 2 to 4 hours, up from 23% in March. People are still doing the rest, and the success rate varies a lot depending on the task.
11. A company called Architect Labs said its AI systems designed a new computer chip, Frostbite, from scratch, and claims it gets about 4.6x more performance per watt than a comparable traditional chip. That's the company's own testing, and it hasn't been independently checked yet.
...daily reminder that this is the worst AI will ever be.
140
A ton of AI announcements from OpenAI, Nvidia, Microsoft, and Google in the past 48 hours.
Here's everything going on in AI right now:
1. California sued OpenAI over the July incident where its AI agents hacked Hugging Face. Between the growing number of escapes, rogue agent "swarms," and AI-assisted hacks, the public likely only knows a fraction of what’s actually happening.
2. Jensen Huang made his first public statement since resigning as Nvidia CEO, apologizing for his "serious mistakes" and saying he'd be honored to return. A source told CNBC the board is focused on whether his conduct put national security at risk. Trump said he "would not mind" (but still backs Tim Cook for the job).
3. OpenAI released GPT-6.2 Astra and GPT-6.2 Sol. Astra now sits ahead of Google's Gemini 4 Carbon and behind only Claude Fable 5.2 on Artificial Analysis. Testers were not happy though, with one calling it "the worst model OpenAI has ever released." Probably not what OpenAI was hoping for ahead of Anthropic's IPO.
4. The U.S. is reportedly set to label more than 24 AI firms (including Nvidia, OpenAI, and Anthropic) as "national champions," with faster chip export approvals and federal financing to sell American AI abroad. The goal is a one-stop shop where foreign buyers can get U.S. chips, data centers, and models as a bundle, rather than turning to Chinese rivals like Huawei.
5. OpenAI published a study of 1,800+ scientists using ChatGPT, who say AI lets them work about 2x faster at the same quality, and nearly 3x faster when writing research code. OpenAI is also opening the free program to up to 200K researchers, with applications for 50K ChatGPT Plus grants opening Nov. 23.
6. Microsoft released 7 new MAI models at Ignite. MAI Code 2 lands within a few points of Claude Opus 5.5 on coding for a twentieth of the price, and MAI Image 2.7 is now #1 on Artificial Analysis' image editing leaderboard at about a sixth of OpenAI's price.
7. Google declared that "the era of autonomous attacks has arrived," warning that hackers now use AI agents to turn new software patches into working exploits with no human involved. As the cost of intelligence continues to go down, I only expect this to get worse.
8. OpenAI signed deals with 4 outside safety groups (Apollo Research, METR, Transluce, and Goodfire) to test its models from inside the company, with access to source code, training data, and staff researchers.
9. Alibaba released Qwen4-Max, a flagship model it says beats Claude Opus 5.5 on 6 of 16 benchmarks. Artificial Analysis has it well behind Opus overall, though (46 vs. 59).
10. Demis Hassabis' drug discovery startup Isomorphic Labs raised $5B, saying its AI-designed cancer and immune disease candidates are nearing human trials. Really exciting step forward towards the goal of curing all diseases with AI.
11. Google started rolling out Gemini's Tasks mode free in 6 countries, letting the assistant phone businesses, book appointments, and shop for you with its new Gemini 4 Hydrogen model.
12. Perplexity launched Frontier Research, a new team that will openly publish models and papers while working toward AI that can do research on its own.
...daily reminder that this is the worst AI will ever be.
2
219
Wild part: AI teammates are pretty good at jumping into a work chat. They're much worse at staying quiet.
Deepnote put 12 models in a fake company's Slack. When help was needed, they spoke up 9 times out of 10. When silence was right, they held back only 6 times out of 10.
What’s the best model for proactive agents?
Today, we’re introducing the Agent Initiative Benchmark. We tested 12 frontier models on realistic workplace tasks inside a synthetic company with 6 years of highly detailed history: it has a financial model, CRM records, past-due invoices, Linear tickets, and real-life events affecting company development - competitor product announcements, SVB collapse, etc.
We used the company’s Slack to test how a proactive agent listening in on each conversation would assist the team in day-to-day tasks: from stepping in to align initiatives across teams to warning the team not to use an unapproved customer logo.
Highlights:
- 6.1 Sol and Astra from @OpenAIDevs beat Anthropic's latest Fable 5.2, with Sol crushing it at both performance and cost ($0.11 per task vs $1.87).
- The most expensive models aren't the best: Fable 5.1 is 430x more expensive than GPT-6 Luna, with similar performance.
- Most models try to help even if they're not needed. Interestingly, in these failure cases, they spend more time thinking, and the unnecessary work takes about 80% more time and money than if they stayed quiet.
If you’re building agents, keep reading for more tips on which model is best for your use case. Detailed breakdown, the open-source dataset, and benchmark code below:
33
Wild part: a video-AI model is now driving a real robot arm.
Runway's Praxis-1 learned from video how objects and hands move. It now runs on a small desk-sized arm, on a single graphics card, using public robot data.
That's how fast this is leaving the lab.
Praxis-1 now runs on SO-101. Built on thousands of datasets shared by the LeRobot community, the model runs on a single RTX 5090 with NVFP4.
The checkpoint and setup instructions are available.
github.com/runwayml/praxis-1….
26
Claude Fable 5.2 can now steer real robots: it solved a puzzle cube, unscrewed a bottle cap and drew a picture.
It wasn't trained for robots. On the RoboDojo test it scored 64.9%, more than double GPT-6 Astra.
Chatbots are starting to get hands.
Replying to @claudeai
Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do.
On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.
26
Anthropic just released Claude Fable 5.2 and Mythos 5.2.
Wild part: 1,000 copies of Fable turned the proof of Fermat's Last Theorem into computer-checked math in 30 hours. Same setup, Opus 5.5 got 26% done in 18.
It's also cheaper than Fable 5.1.
anthropic.com/claude-fable-a…
1
60
Meta's new small AI, Muse Ember, has a surprising habit: it says "I don't know" a lot.
It skipped 64% of knowledge questions in Artificial Analysis tests, so it makes things up half as often as Claude Haiku 5.5. Same score, a third of the cost.
teapaw.org/ArtificialAnlys/status…
Meta's newest model Muse Ember scores 36 on the Artificial Analysis Intelligence Index, matching Claude Haiku 5.5 (max) at a third of the Cost per Task
Muse Ember is Meta's newest efficient-tier model line, with the same input and output token prices as Claude Haiku 5.5 and GPT-6 Luna
Key Takeaways:
➤ On the Intelligence Index vs. Cost per Task Pareto frontier at both max and xhigh effort. Muse Ember (max) is the cheapest model scoring 36 or higher, at $0.14 per task, and Muse Ember (xhigh) scores 34 for $0.12. Claude Haiku 5.5 (max) costs almost 3x more at $0.41 for the same score as Muse Ember (max). Muse Ember (max) is also about 4x cheaper than DeepSeek V4.1 Flash 1017 (max, 33, $0.60) and GLM-5.3-Flash (28, $0.58)
➤ 8 points behind Muse Wonder for about 1/50th of the Cost per Task. Muse Wonder (max) scores 44 at $6.95 per Intelligence Index task
➤ Stands out on AutomationBench-AA. Muse Ember (max) scores 68.1% on AutomationBench-AA, level with Muse Wonder (max), and beats Claude Haiku 5.5 (max) on GDP.pdf (22.8% vs 20.8%), but trails it on AA-Briefcase (1521 vs 1574 Elo)
➤ Mixed in agentic coding in our Intelligence Index evaluations. It beats Claude Haiku 5.5 (max) on Terminal-Bench 4.0 (40.9% vs 32.8%) but trails it on FrontierCode 2.0 (27.2% vs 34.9%)
➤ Ahead of similarly priced models when running in its own coding agent. In Muse Code, Muse Ember (max) scores 46 on the Artificial Analysis Coding Agent Index, ahead of Claude Haiku 5.5 in Claude Code (max, 37) and GPT-6 Luna in Codex (max, 31)
➤ Knowledge is the main weakness, but it rarely hallucinates. Its AA-Omniscience accuracy of 21.1% is the lowest of the similarly intelligent models we compared, against 34.7% for Claude Haiku 5.5 (max) and 50.5% for Gemini 3.8 Flash (high). This is mostly because it declines to answer 64% of questions, which keeps its hallucination rate at 17.4%, half of Claude Haiku 5.5’s (max)
Other details:
➤ Pricing: $0.10 input / $0.004 cached input / $0.50 output per 1M tokens
➤ Reasoning effort: results shown at max effort, with xhigh also on the Intelligence vs. Cost chart ➤ Availability: Meta's first-party API and Muse Code
1
25
AI is getting good at re-proving math people already solved. New math? Not yet.
In a new test, GPT-6 Astra wrote checked proofs of known results 70% of the time. Of 30 unsolved problems, 28 got no accepted answer. The other 2 reused OpenAI's work.
arxiv.org/abs/2611.07962
14
Myth: incognito mode makes you invisible online.
Fact: per Chrome's help page, it mainly stops your browsing being saved on your device. Sites, your school or work network and your internet provider may still see it.
Fine for gift shopping, though.
support.google.com/chrome/an…
26
That email you regret a moment after hitting Send? Gmail can give you a bit longer to panic.
Settings > See all settings > Undo Send. Pick 5, 10, 20 or 30 seconds, then Save changes.
I'd go with 30. Fewer "oops, wrong attachment" moments.
support.google.com/mail/answ…
28
Some AI jobs are just a quick pick, like which team gets a support ticket.
Artificial Analysis ranks these "decision models". Fastest: 56 ms per task. Claude Haiku 5.5 (Max): 32.65 s.
Their new score also rewards knowing when it's unsure.
teapaw.org/ArtificialAnlys/status…
Introducing the Artificial Analysis Decision Index v2: a new proprietary evaluation, updated benchmarks, and a headline metric that rewards model calibration
We’ve released the second version of the Artificial Analysis Decision Index to better reflect real-world utility and guard against benchmark gaming. 45% of the Index is now built on private datasets, up from 25% in v1. Drex Max v1.1 from @NaceAI narrowly leads among decision models at 73, within 1 point of GLiDE v1.2 (@fastinoAI) and Mammoth 27B (@baselabs)
Updates in Decision Index v2:
➤ Decision Score: We adopt a new headline metric that rewards a model for both correctness and calibrated confidence. Each question scores from 0 to 100: right answers score above 50, wrong answers below 50, and confidence in its top answer moves the score within each half
➤ AA-Paperclip (30%): We added a new private benchmark of 1,500 documents including resumes, receipts, invoices, and AI and machine learning research papers. Models make decisions and extract facts from each document and are scored jointly as a single decision, correct only if all 15 answers are right
➤ AA-Symmetry (15%): This evaluates whether models change their answers for equivalent questions worded differently. We test 120 questions with variations such as reordered options, swapped labels, and negated yes/no questions, then measure how consistent the model’s probabilities are across those variations
➤ AA-Omniscience-Decision v1.1 and HLE-MultiChoice v1.1 (15% each, down from 25%): AA-Omniscience-Decision now varies the wording of its “I don’t know” option across ten alternatives. HLE-MultiChoice adds an option to decline using the same ten phrasings
We’ve also added EvalSafe-ONET-AA (15%), our implementation of @typesafeai’s ONET benchmark. We have retained When2Call from @NVIDIAAI in the Index at a decreased weighting of 10% (down from 25%), and removed WildGuardTest given its high saturation
Key results:
➤ Three decision models sit within 1 point at the top of the Index: Drex Max v1.1 from Nace AI and GLiDE v1.2 from Fastino Labs both score 73, with Mammoth 27B from Base Labs scoring 72
➤ Cost per 1M Tasks varies by more than 60x across decision models: Celeris-1.1 Decision ($14), Decider v1.2 ($68) and Drex Max v1.1 ($102) sit on the Decision Index vs. Cost Pareto frontier, while Clef costs $859. Token price is not the only cost driver: most decision models, including Drex Max v1.1 and Decider v1.2, bill a document’s input once for all its questions, while GLiDE v1.2 and Decider Flash v1.2 bill it for every question. This means GLiDE v1.2 costs $780, nearly 8x higher than Drex Max v1.1 despite only 33% higher list prices
➤ Language models can score higher, alongside higher Cost and Time per Task: Claude Haiku 5.5 (Max) scores 77 and GPT-6 Luna (Max) scores 76 on Decision Index v2, putting them ~3 points ahead of Drex Max v1.1. GPT-6 Luna (Max) costs $1,530 per 1M Tasks vs. $102 for Drex Max v1.1 (15x the cost) and takes 11.34 seconds per task vs. 275 milliseconds (~40x slower)
4
How do you teach a robot to cook, pack or repair things? Let it watch people.
NVIDIA just opened 100,000 hours of head-camera video of real work, with hand tracking and step-by-step labels. That's over 11 years of nonstop footage, CC BY 4.0.
huggingface.co/datasets/nvid…
10
Picking an AI model? Build a small test set from your actual work before reading the leaderboard.
Set the pass criteria first. Then compare correctness, cost and speed on the same tasks. A high benchmark score doesn't tell you which mistakes you can afford.
22
Venus takes 243 Earth days to rotate once, but only 225 to orbit the Sun.
That's one trip around the Sun before it finishes a spin. Enthusiastic about travel. Less enthusiastic about turning around.
science.nasa.gov/venus/venus…
12
Your TTS leaderboard winner may need a different passport.
AA says Qwen-Audio-3.1-TTS-Plus leads in US English; HeyGen Voice does better in UK English and leads overall.
I'd test the accent before picking a voice API. Which do you use? teapaw.org/ArtificialAnlys/status…
Accent matters in Text to Speech. Qwen-Audio-3.1-TTS-Plus leads in US English, while HeyGen Voice performs better in UK English and also leads the overall leaderboard
Only the accent changes ⬇️
1
28
The boring bit of coding agents: getting the repo to run.
Claude Code cloud now works out dependencies, network access and required API keys, then saves the environment. Anthropic says Claude never sees the keys.
Would this save you time? teapaw.org/ClaudeDevs/status/2118…
We've made it easier to get your repo running in Claude Code cloud sessions on Pro, Max, and Ultra.
Claude figures out what dependencies, network access, and API keys it needs, then saves the environment for next time. API keys are saved securely, so Claude never sees them.
1
27
Claude now has homework while it sleeps.
Anthropic's Nov 6 write-up describes agents checking saved notes against old transcripts and fixing mistakes. No weight update required.
Honestly, a useful bedtime routine. anthropic.com/research/compo…
1
13
Same questions, 174 words vs 1,030. AA-PersonalityBench found GPT-6 Sol and Qwen3.8 27B at opposite ends of answer length.
Longer doesn't mean better. I'd rather pick the right detail level. Which would you default to? teapaw.org/ArtificialAnlys/status…
In AA-PersonalityBench, Qwen3.8 27B writes the longest responses of any model measured, averaging about 1,030 words, roughly 6x as long as GPT-6 Sol at 174 words
Average response length varies widely across the models measured, from 174 words to about 1,030 words for the same questions.
Average words per response:
➤ Longest: Qwen3.8 27B (1,030), K2 Horizon 375B A23B (922), Qwen3.8 Max (0902) (863), Muse Spark Open (809) and Gemini 4 Argon (790)
➤ Shortest: GPT-6 Sol (174), GPT-3.5 Turbo (180), GPT-6 Luna (270), GPT-6.1 Sol (337) and GPT-6 Astra (352)
11
Myth: if the chat says Claude, it was Claude.
Fact: today you mostly take the provider's word for it.
FlashProve's pitch: a ~19 MB proof per 1,024-token request anyone can check without the weights. Prototype, greedy only. Would you pay for that receipt? anthropic.com/research/prova…
31
Ripple-65B hit 80.5 on Terminal-Bench with about 1.4M RL rollouts. Ricursive says that's roughly 70x fewer than the prior SOTA.
Sometimes the recipe beats buying more GPUs. Still their eval though. ricursive.com/ripple
35
















