Gemini 4 Carbon Tops the AI Index, DeepSeek Raises ¥105B, Agibot Pitches 'Worken'

AI Morning Brief for October 30, 2026: Gemini 4 Carbon Tops the Index, DeepSeek Raises ¥105.5B, and Agibot Bets on Real Work Hours

On October 30, 2026, money and technology in the AI world both fired simultaneously. Early this morning, Google released Gemini 4 Carbon, which topped the global Intelligence Index with a score of 59. DeepSeek was reported to have completed a second funding round of ¥105.5 billion at a valuation of ¥615.5 billion. 01.AI (Zero One Things) received several hundred million yuan from Hangzhou state capital. OpenAI is one step away from closing a $30 billion round. Agibot used its Innovation Day to change the evaluation standard for robots from "units shipped" to "hours worked on the job." This page deserves three minutes of your time.

This is not a daily log; it covers the five things you need to know today. Each item gives "what it is, why it matters, and what it means for you." A comparison table at the end lets you scan everything at once.

📦 Bonus resource: a directly runnable Python script make_ai_brief.py at the end for scraping news sources and auto-generating a morning brief. Copy and paste to run.

Google Gemini 4 Carbon: Scores 59, Tops Intelligence Index
DeepSeek Second Round Raises ¥105.5B, Valuation ¥615.5B
01.AI Receives Hundreds of Millions from Hangzhou State Capital, HQ in Linping
OpenAI Nears $30B Round at $1.25T Valuation
Agibot Innovation Day: 10,000 Units Shipped Only Proves Manufacturing Ability
Today's AI Highlights Comparison Table
Closing Thoughts: Where Is Today's Money Flowing?

1. Google Gemini 4 Carbon: Scores 59, Tops Intelligence Index

At around 2 a.m. Beijing time on October 30, Google released its new flagship model Gemini 4 Carbon, just 15 days after the previous generation Argon was fully rolled out. Parameter count, architecture, and technical report have not been disclosed, but the performance speaks: in the just-released Intelligence Index 5.0 from Artificial Analysis, Carbon scored 59, taking first place globally ahead of Claude Opus 5.5 (58) and GPT-6 Astra (57). The previous generation Argon scored 50.

Where does it excel? The most striking is long-context capability. In the official results, Carbon scored 96.2% on the GraphWalks long-context task, while Astra got 71.8% and Opus only 66.8%. That is a wide gap. Carbon won 15 of 17 official evaluations. On the security front it also pulled a "triple win": vulnerability detection 88.2%, exploit validation 54.8%, and vulnerability patching 71% - first place in all three.

The caveat: these security tests ran with guardrails off. In production with guardrails, scores may differ. And Carbon's launch token pricing is half off (input ~$2 per million, output ~$10 per million), clearly aimed at grabbing market share.

Why it matters: The Gemini 4 series iterated from Argon to Carbon in half a month, showing that leading closed-source vendors have compressed release cycles to "weekly" speed. For developers, the clearest signal is that the bar for long-context plus reasoning has been raised again. What used to be impossible - reading an entire codebase in one pass - is now a default capability.

Gemini 4 Carbon References

2. DeepSeek Second Round Raises ¥105.5B, Valuation ¥615.5B

If Gemini is "technology firing," then DeepSeek is "capital firing." On October 30, The Information reported, citing two people with direct knowledge, that DeepSeek's second funding round reached ¥105.5 billion with about 30 investors, valuing the company at ¥615.5 billion (the report did not specify pre-money or post-money).

Where does the money come from? Tencent and CATL remain the two largest investors: Tencent contributed about ¥10 billion, and the CATL system (including Zeng Yuqun) roughly ¥5-10 billion. China Mobile and CITIC Securities each put in about ¥8 billion. ChangXin Memory, SMIC, vivo, Meituan, Xiaohongshu, Midea, Geely, China Telecom, CICC, and others each invested ¥1-3 billion. Founder Liang Wenfeng personally put in another ¥2 billion this round. Overall, state-affiliated investors hold about 27%, giving this round a clear "domestic compute + semiconductor supply chain" collective flavor.

The caveat: The ¥105.5B figure was revised upward from an initial target of about ¥50B, to ¥100B, and finally to ¥105.5B. The "doubling" of the round is about loading up ammunition before an IPO. DeepSeek has already chosen CITIC Securities to assist with listing preparations, targeting the STAR Market, with the earliest possible IPO in 2027. Industry consensus expects completion between 2027 and 2028.

Why it matters: This is the ceiling-level event for a single funding round in China's large-model space. An unlisted AI company pushed its valuation to ¥615.5B in one round, showing that capital markets' imagination for a "China version of OpenAI" has fully opened. For practitioners, it means domestic large models have a thicker ammunition store for the "compute arms race" - and that is precisely the confidence behind open-source models' ability to stay cheap and iterate quickly.

DeepSeek Funding References

3. 01.AI Receives Hundreds of Millions from Hangzhou State Capital, HQ in Linping

If DeepSeek is "national team unity," then 01.AI is "local state capital poaching talent." On October 30, 01.AI, founded by Kai-Fu Lee, announced a partnership with Hangzhou city and Linping district. Combined investment from city- and district-level state capital totals several hundred million yuan, with headquarters officially moving to Hangzhou's Linping District, which will serve as a base for an industrial AI transformation hub.

Background: Founded in 2023, 01.AI is one of the "Six Little Tigers" of Chinese large models and has released three self-developed models. At the end of 2024, it began a strategic pivot from "frontier large-model research" to "industrial AI deployment," focusing on delivering practical AI solutions for enterprises. That pivot is exactly why it can secure government order-backed funding.

Why it matters: This is not pure financial investment; it's "scenarios for capital." Hangzhou opens its local digital economy and physical industry scenarios to 01.AI. State capital supplies money; the company supplies models and deployment capability. For 01.AI, this is a key step toward a Pre-IPO push, targeting an IPO in 2027 and orders of ¥1.5 billion in 2026. For local government, it's part of vying to become the "first city for AI innovation." When large-model companies start exchanging scenarios for money, the industry has moved from "competing on parameters" to "competing on deployment and getting paid."

01.AI References

4. OpenAI Nears $30B Round at $1.25T Valuation

While China is grabbing people and money, OpenAI across the ocean is not idle. Around October 22, OpenAI was close to closing a round of up to $30 billion at a pre-money valuation of $1.25 trillion, co-led by UAE investors and investment firm Coatue. This funding is positioned as bridging capital before a potential OpenAI IPO.

Some details: Earlier reports said the round was priced at a pre-money of about $1.4 trillion with no lead investor, offering a fixed valuation to potential backers. That later narrowed to a pre-money of $1.25 trillion led by a UAE consortium. For reference, in October 2024 OpenAI's valuation was only $157 billion - nearly 8x growth in two years. Meanwhile, Nvidia and SoftBank each completed their final $1 billion investment in the previous round, both trying to lock in their positions before this large round closes.

Why it matters: What does $30 billion mean? It's slightly larger than DeepSeek's ¥105.5B. It is one of the largest single funding rounds in AI history. By filling its war chest before an IPO, OpenAI is signaling it will spare no expense to expand compute and commercial footprint during the "land-grab" phase. For the whole industry, this funding chain (UAE sovereign funds → Nvidia → SoftBank) essentially links up the "arms dealers" of AI.

OpenAI Funding References

5. Agibot Innovation Day: 10,000 Units Shipped Only Proves Manufacturing Ability

The previous four events were about spending money, but Agibot's Innovation Day in Shanghai on October 29 pulled the conversation from money back to "can robots actually do work?"

Memorable quote: Co-founder Peng Zhihui (also known as ZhiHuiJun) said: "10,000 units shipped only proves manufacturing ability; 10,000 units that can actually work on the job is the victory of physical AI." This statement directly poured cold water on the "shipment worship" - producing a robot is not the achievement; getting it to finish work well on a production line is what counts.

Data support: As of 2026, Agibot shipped 5,000+ units in 2025, about 40% of the global share; in H1 2026 it shipped 9,700 units with a 43%+ share; the 20,000th unit came off the line in September, and H1 revenue was nearly ¥2 billion. On 3C production lines, its task success rate reaches 99.99% with a cycle time of 22 seconds per part. At Chimelong in Hengqin, nearly 500 robots are deployed, serving over 10,000 visitors daily. These numbers show that "working on the job" is no longer just a slogan.

New metric Worken: Agibot proposed a "measurement" for physical AI - Worken (Effective Intelligent Work Hours), defined as "active robot count × autonomous work duration × output efficiency ratio." Behind this is a conceptual shift: in the digital AI era, thinking is measured in tokens; in the physical AI era, work should be measured in "effective intelligent work hours." A robot that needs human takeover every 10 minutes has limited autonomous work time and thus low value.

Why it matters: When robot manufacturers start defining value by "effective work hours" rather than "units shipped," the industry has entered a phase of competing on "who can reliably work on the job and keep producing." On the other side, OpenAI's GPT-6 Astra often takes more than 18 seconds for a single inference, and a near-trillion-parameter model running in the cloud can take over ten minutes for a simple task - a clear case of "smart brain but no body yet." Agibot uses training with 4 robots versus 100 real robots, boosting training success from 50% to 98%, paving the way for the "brain + body" connection.

Agibot Innovation Day References

6. Today's AI Highlights Comparison Table

If you compress today's main threads into a "symptom → cause → solution" table, you can see clearly: money and technology are both accelerating, and money is getting smarter.

Symptom (what you see)	Cause (logic behind it)	Solution (what to do)
Gemini 4 Carbon tops the index in half a month, half-price tokens	Leading closed-source vendors compress iteration to weekly cycles, price war begins	Watch for the "affordability" of long-context + reasoning; make reading an entire repo in one pass a default ability
DeepSeek single round ¥105.5B, valuation ¥615.5B	Domestic compute and semiconductor supply chain "group up," loading up before IPO	Stop hesitating on domestic large-model compute; open-source low prices hide high investment
01.AI trades scenarios for Hangzhou state capital	Large models shift from "competing on parameters" to "competing on deployment and getting paid"; local governments want industrial AI scenarios	If building enterprise AI solutions, prefer locations with scenarios and state backing
OpenAI raises $30B at $1.25T valuation	Pre-IPO bridge round; Middle East sovereign funds and arms dealers collectively stake out	Watch AI infrastructure and compute chains; the leading pattern has 8xed in two years
Agibot proposes Worken (effective work hours) instead of shipment count	Industry shifts from "can build it" to "can work long"; physical AI needs a new value metric	Evaluate robot vendors by "effective hours + on-the-job success rate," not just unit count
Daxiao ACE-Brain-0.6, PoQ RoboAPI released one after another	Embodied models move from "demo" to "publicly accessible APIs"; spatial anchoring becomes a hot spot	Embodied/robotics developers can try these spatial-anchoring APIs to cut costs and speed up

Daxiao and PoQ Robot References

7. Closing Thoughts: Where Is Today's Money Flowing?

These five events may look like "technology competition + capital frenzy," but strung together they form a clear line - money and technology are accelerating together, and money is getting smarter.

On price: Gemini 4 Carbon launched with half-price tokens, showing that even leading closed-source vendors are now fighting over "cost per task." DeepSeek's single round of ¥105.5B and OpenAI's $30B show that "who has compute and who has scenarios" is deciding the next round of the landscape.

On standards: Agibot replaced shipment count with Worken (effective intelligent work hours), and Daxiao surpassed Astra with spatial anchoring. The industry is moving from "comparing parameters" to "comparing who can actually do the work."

When to pay attention: If you build agents, embodied intelligence, or enterprise AI transformation, this day is worth noting. Whether on "cost affordability" (Carbon, SpaceBunny), "compute ammunition" (DeepSeek, OpenAI), or "deployment metrics" (Agibot's Worken), the rationale for investing in AI deployment is getting stronger.

When not to panic: If you're just watching, much of this is still in the "research preview" or "release-as-demo" stage. Don't be swept away by "¥615.5B valuation" or "$30B funding." These numbers describe capital betting on the future, not that general-purpose replacement has already happened.

Companion code: make_ai_brief.py (directly runnable)

The script below is a companion resource for this article. It scrapes news sources (using publicly available RSS/JSON as a demo source) and auto-generates a morning brief in Markdown. Copy and paste to run.

# -*- coding: utf-8 -*-
"""
make_ai_brief.py — demo script to auto-generate an AI morning brief in Markdown

Function: Fetch titles and summaries from several public news sources, assemble a briefing using the "symptom → cause → solution" structure, and output a .md file.
Parameters: none. On run, generates ai_brief_YYYYMMDD.md in the current directory.
Returns: nothing; writes a file directly.
"""

import json
import re
import urllib.request
from datetime import date


# Simulated news sources (replace with actual RSS / news API URLs in production)
SOURCES = [
    {
        "title": "Gemini 4 Carbon tops Intelligence Index",
        "summary": "Google released Gemini 4 Carbon this morning; Intelligence Index 5.0 scores it 59, beating Opus 5.5 (58) and Astra (57).",
        "symptom": "Flagship model iterates in half a month, half-price token launch",
    },
    {
        "title": "DeepSeek raises ¥105.5B in second round",
        "summary": "DeepSeek is reported to have raised ¥105.5B in its second round at a ¥615.5B valuation, with ~30 investors.",
        "symptom": "Single round exceeds ¥100B; domestic compute groups up",
    },
    {
        "title": "Agibot proposes Worken to measure physical AI",
        "summary": "Agibot's Innovation Day proposed Worken (effective intelligent work hours) = active robot count × autonomous work duration × output efficiency ratio.",
        "symptom": "Industry shifts from comparing shipment to effective work hours",
    },
]


def fetch(url: str) -> str:
    """Fetch content from the given URL and return a string.

    Args:
        url (str): target URL.
    Returns:
        str: fetched text; empty string on failure.
    """
    try:
        # Pretend to be a browser to avoid some sources rejecting the request
        req = urllib.request.Request(url, headers={"User-Agent": "Mozilla/5.0"})
        with urllib.request.urlopen(req, timeout=10) as resp:
            return resp.read().decode("utf-8", errors="ignore")
    except Exception:
        return ""  # empty string on failure, keeps script running


def clean(text: str) -> str:
    """Strip HTML tags and extra whitespace, return clean title.

    Args:
        text (str): raw HTML text.
    Returns:
        str: cleaned plain text.
    """
    tag_re = re.compile(r"<[^>]+>")          # matches HTML tags
    text = tag_re.sub("", text)              # remove tags
    text = re.sub(r"\s+", " ", text).strip()  # compress whitespace
    return text


def build_brief(items: list) -> str:
    """Assemble the brief's Markdown from a list of items.

    Args:
        items (list[dict]): each dict has title/summary/symptom.
    Returns:
        str: assembled brief in Markdown.
    """
    today = date.today().strftime("%Y-%m-%d")
    lines = [f"# AI Brief {today}", ""]
    for i, item in enumerate(items, start=1):
        lines.append(f"## {i}. {item['title']}")
        lines.append("")
        lines.append(f"- **Symptom**: {item['symptom']}")
        lines.append(f"- **Quick look**: {item['summary']}")
        lines.append("")
    lines.append("## Comparison Table")
    lines.append("")
    lines.append("| Symptom | Cause | Solution |")
    lines.append("|---|---|---|")
    for item in items:
        lines.append(f"| {item['symptom']} | See details above | Focus on deployment metrics |")
    lines.append("")
    return "\n".join(lines)


def main() -> None:
    """Main entry: fetch data and write the brief file.

    Args: none.
    Returns: none.
    """
    # Demo: use built-in material; in production, change to fetch(url) per source
    items = [{"title": s["title"], "summary": s["summary"], "symptom": s["symptom"]} for s in SOURCES]
    md = build_brief(items)
    filename = f"ai_brief_{date.today().strftime('%Y%m%d')}.md"
    with open(filename, "w", encoding="utf-8") as f:
        f.write(md)
    print(f"Generated: {filename}")


if __name__ == "__main__":
    main()

After running, it generates a file like ai_brief_20261030.md containing a brief skeleton organized as "symptom → cause → solution." Replace the SOURCES entries (or convert them to real RSS fetching) and you get a personalized briefing.

When to use this brief: Use it as an "index + skeleton" when you want to quickly scan what's happening in AI each day. But if you need deep analysis, it can only circle the key news; the real judgment is yours.

This column, "LLM Large Model Technology," covers everything from Transformer attention mechanisms to LangChain, RAG, and MoE deployment, with a runnable snippet per article, updated weekly, focused on practical models and agent practices. Subscribe from the column page to read in order.

Comments

Top 1 from juejin.cn, machine-translated. The original thread is authoritative.

掘一二三桶金

Great stuff, thanks for sharing