Kimi K3.1 Pro Beats GPT-6 Astra, Ranks No. 4 in AI Index
ALSO: Anthropic’s junior hires drop to 9%, ChatGPT PIN lock
Welcome to Guru99 AI Report!
Top Story: Hi there, a Chinese model just broke into AI’s global top four, and anyone can now download it for free. Also in this issue: why Anthropic is hiring far fewer juniors, ChatGPT’s new PIN lock, and GM’s 300,000 virtual batteries.
🚀 China’s Kimi K3.1 Pro Cracks AI’s Global Top Four
Brief Buzz:
Chinese AI lab Moonshot AI has launched Kimi K3.1 Pro, an upgrade to its K3 model from July. It’s already live in the Kimi apps and API, with free-to-download weights scheduled for November 4. Moonshot says it’s on par with, or very close to, the best closed models.
- No. 4 on Artificial Analysis’ Intelligence Index — above GPT-6 Astra and behind only Claude Opus 5.5, Sonnet 5.5, and Meta’s Muse Wonder Full.
- Split verdict vs. Opus 5.5 in Moonshot’s own tests: ahead on long coding projects and expert exams, behind on terminal coding and 3D design.
- $2/$10 per million input/output tokens — a fifth of Astra’s rate — though it uses more tokens than similarly priced models.
- 64% on the Artificial Analysis cyber defense index, the highest publicly available score, with zero refusals. Only OpenAI’s GPT-6 Sol (vetted access) scores higher at 69%.
- Offense deliberately capped: Moonshot removed hacking-related training data from the open release, cutting its internal exploit score from 90.5% to 81%.
💡 Why Should You Care?
Top-end AI is getting cheaper and will soon be downloadable. Any company could run K3.1 Pro on its own servers — a win for budgets and security audits. The downside: once the weights are public, no one can ever recall them.
💼 Anthropic’s Junior Share of Tech Hires Falls to 9%
Brief Buzz:
Anthropic turned the lens on its own workplace. A new study from its Economics team examines how AI agents are changing hiring and the nature of work inside the company. The result: far fewer junior technical hires and a big jump in output per person. All figures come from Anthropic’s internal data.
- Junior roles have fallen from 27% to 9% of Anthropic’s technical hires since early 2025.
- 6-to-1: by August, AI agents directed by employees were logging 6 hours of work for every 1 hour of human work.
- 5x the code commits per engineer compared with six months earlier — and agent code was removed or reverted no more often than human code.
- ~80% idle: in interactive sessions, agents spend most of their time waiting for a human to respond. The human is the bottleneck.
- Internal moves up 50% among tenured staff, as Anthropic builds an early-warning system to flag roles AI might disrupt.
💡 Why Should You Care?
This is one company grading itself, so it doesn’t represent the whole economy. Still, it suggests entry-level coding is thinning out, and the skills that will land you a job are judgment, autonomy, and creative use of AI.
🔒 ChatGPT Can Now Hide Your Secrets Behind a PIN
Brief Buzz:
OpenAI has given ChatGPT a locked folder. Locked Chats lets you move sensitive conversations into a PIN-protected folder, out of reach of anyone who borrows your phone or glances over your shoulder. It began rolling out to most plans on October 20.
- Chats unlock with a 4-digit PIN; supported devices can also use Face ID, Touch ID, or device unlock.
- New messages in locked chats won’t update your regular memory; everything previously remembered stays as it was.
- Unlike Temporary Chat, where conversations are deleted, locked chats are saved so you can return to them anytime.
- It’s a privacy curtain, not a vault — no end-to-end encryption, and retention and model-training settings are unchanged.
- Forget your PIN? You can reset it, but that permanently deletes all your locked chats, with no recovery.
💡 Why Should You Care?
People confide health worries and relationship problems in chatbots. Yet 75% of partnered Americans told Pew in 2019 that they’d shared their phone passcode — so a phone passcode alone isn’t enough.
🧪 OpenAI Buys Orchestra Bio to Untangle Drug Development
Brief Buzz:
OpenAI has acquired Orchestra Bio, a four-year-old startup whose AI agents help biopharma teams manage the heavy logistics of drug development. The team and technology will join Rosalind, OpenAI’s life sciences suite. Deal terms weren’t disclosed.
- A new US medicine takes roughly 10 to 15 years from target discovery to approval, according to OpenAI.
- Ripple effects are the problem — a failed manufacturing batch can delay a clinical trial. Orchestra’s agents map what’s affected and compare possible fixes.
- Humans still decide: agents handle study and timeline tracking, and scientists use that analysis to choose the next step.
- The goal is to connect GPT-Rosalind research insights to downstream studies and decisions. Moderna and Novo Nordisk are already Rosalind partners.
- Big labs are buying in: the deal follows Anthropic’s reported ~$400M acquisition of biotech startup Coefficient Bio in April.
💡 Why Should You Care?
A discovery only helps patients once years of research, partners, and trials fall into place. OpenAI is betting agents can speed up that handoff — though it hasn’t said by how much.
🐋 DeepSeek Open-Sources the Engine Behind Its Low-Cost AI
Brief Buzz:
DeepSeek has open-sourced DeepInfer, the software it uses to serve its own models in production. It’s designed to run DeepSeek-V4.1-Flash on your own hardware, with the stated aim of lowering compute costs and energy use. It’s not the full stack.
- Two chip families: Nvidia’s Blackwell (SM100) GPUs and Huawei’s Ascend 950. Older Nvidia chips such as Hopper aren’t included.
- Apache 2.0 licensed, shipping with DeepSeek’s custom kernels and speculative decoding (drafting tokens ahead to speed up responses).
- Missing pieces: a KV cache system (the model’s working memory), prefill/decode splitting, and load balancing — you’ll need to build those yourself.
- DeepSeek’s published speed tests all ran on Nvidia GPUs, so Ascend performance remains an open question.
- No Blackwell or Ascend chips? According to The New Stack’s analysis, you can study the code and port its tricks into your own engine.
💡 Why Should You Care?
Much of AI’s cost comes down to running models efficiently. This release hands self-hosters DeepSeek’s own playbook — and lists Huawei’s chips alongside Nvidia’s.
🔋 GM Tested 300,000 Battery Designs Before Building One
Brief Buzz:
GM has previewed its fourth-generation EV platform: a new battery chemistry paired with a central computer built for AI. The interesting part is how much of it was designed and tested in software first. No models or launch dates have been announced.
- 300,000+ battery designs ran through computer models before a single physical prototype was built, according to GM.
- LMR (lithium manganese-rich) batteries target 33% higher energy density than low-cost LFP cells at a similar cost.
- Early results point to about 20% more range per kWh, according to GM.
- A central computer delivers 40x the processing power for AI features while using 30% less wiring.
- Cars will test their own electrical connections on the assembly line, removing the final electrical check.
💡 Why Should You Care?
Testing is moving into software: GM is screening batteries virtually and building self-checks into the cars. For QA professionals, simulation and digital-twin testing skills are becoming the new career currency.
Hey! I’m Krishna Rungta
Founder of Guru99.com, Editor-in-chief & Technology Expert
Was this email forwarded to you? Sign up for free here.
