OpenAI rolled GPT-6 to all 1.2 billion ChatGPT users this week, pairing it with Intelligent UI — answers that include charts, buttons, forms and calculators built inline. Mistral shipped Large 4: a 1.05 trillion parameter model trained entirely in European data centers, with open weights coming at end of October. Manus raised $500 million at a $4 billion valuation, its first funding since China forced Meta to unwind a $2 billion acquisition of the company in April. Today we have:
Featured Materials 🎟️
News of the week 🌍
Useful tools ⚒️
Weekly Guides 📕
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
Bonus Materials 🎁
We’ll Be at TechCrunch Disrupt and AI Week — Come Find Us 👋
Two big events are happening next week. TechCrunch Disrupt runs October 13–15 at Moscone West in San Francisco — 10,000+ founders, investors, and operators in one place. AI Week New York runs the same week on the East Coast.
We’ll be at TechCrunch Disrupt. If you’re an AI startup, a fellow creator, or just someone who wants to talk shop — we’d love to meet in person.
We’ll make sure we connect.
Featured Materials 🎟️
OpenAI Rolls GPT-6 to All ChatGPT Users — With Answers That Actually Do Things 🎛️
On October 7, OpenAI deployed GPT-6 to all paid ChatGPT tiers with a new capability called Intelligent UI. The model now composes answers from text, visuals and interactive elements, choosing the mix based on the question. Ask about mortgage options and you get a working calculator. Ask about market data and you get a live chart you can adjust. Ask how to file a form and you get the form itself, pre-filled with your context.
Paid tiers (Plus, Pro, Business, Enterprise) run on GPT-6 Sol. Free and Go tiers, which got access on October 8, run on GPT-6 Luna. Both are the same models that launched at DevDay on September 29. The difference from that launch: Intelligent UI now renders natively in Chat, not just in Codex or API responses.
What specifically ships: interactive explanations where changing an input updates the result; native charts generated from data in the conversation; embedded forms and calculators; and tappable decision flows. None of it requires a plugin. The model decides what format serves the answer.
OpenAI said it trained GPT-6 to treat format as part of the response, not an afterthought. That is the meaningful shift. A text answer that tells you to calculate something is not the same as an answer that does the calculation and shows you the input levers. For 1.2 billion weekly users, most of whom are not developers, this is what “AI that does things” looks like.
For most of ChatGPT’s users, Intelligent UI is the most visible change to what AI looks like since ChatGPT launched. The response itself is now the interface.
Source: OpenAI
Mistral Launches Large 4: Europe’s First Trillion-Parameter Open Model 🌍
On October 6, Mistral AI launched Mistral Large 4 in public preview — a 1.05 trillion parameter mixture-of-experts model trained from scratch in Mistral’s own European data centers, on 3,800 Nvidia Grace Blackwell GPUs. Open weights are scheduled for end of October. The company has not yet published the license.
The architecture activates 49 billion parameters per token, keeping inference cost close to a mid-size model while the full parameter count sits at a trillion. Context is 1 million tokens. The model handles text and images natively. Training data covers more than 160 languages, including every official EU language.
Mistral is positioning Large 4 as Europe’s answer to Chinese open-weight models like DeepSeek and Kimi, which have dominated the frontier open-source tier for most of 2026. On Vals AI’s benchmark index, Large 4 leads among open-weight models from the US and Europe. Against closed frontier models like GPT-6 Astra and Claude Opus 5.5, it trails — but at $1.36/$4.18 per million input/output tokens, it prices at roughly the same level as GLM-5.3 from China.
The model ID in Mistral Studio is mistral-large-4-0. It is available now via API; open weights follow after Mistral completes additional security tests. The company calls it “ML4” internally and “le Chonk” everywhere else.
A trillion-parameter open-weight model trained exclusively in European data centers is the clearest statement yet that European AI infrastructure is not a future ambition. It is running production jobs today.
Source: Mistral AI
Manus Raises $500M at $4B — First Round Since China Forced the Meta Divorce 🤖
On October 8, Butterfly Effect — the Singapore-headquartered parent company of Manus — announced it had closed a funding round of more than $500 million. Boyu Capital and IDG Capital co-led. Tencent, HSG (formerly Sequoia China), and ZhenFund participated. No valuation was disclosed in the announcement, but September reports had put the target at $4 billion, double the company’s last-known figure.
The context makes the number more meaningful than the number itself. In December 2025, Meta agreed to acquire Manus for more than $2 billion. In January 2026, Chinese regulators opened an investigation. In April, Beijing’s National Development and Reform Commission prohibited the deal and required both companies to unwind it. Manus resumed independent operations in August, deleted user data required by the separation terms, and by October had closed the largest AI agent funding round in Chinese startup history.
Manus builds general-purpose AI agents for research, coding, and web automation. It reported annualized recurring revenue of $100 million at the time Meta acquired it. A report from June 2026 put that figure above $500 million. The company is now reportedly considering a Hong Kong listing.
The round is the clearest evidence yet that Chinese AI capital did not follow talent to Singapore — it followed Manus.
Beijing blocked a $2 billion exit, the company raised $500 million independently, and its ARR quintupled during the acquisition freeze. The geopolitical disruption made Manus more valuable, not less.
Source: TechCrunch
Keep your mailbox updated with practical knowledge & key news from the AI industry!
News of the week 🌍
Anthropic Ships Claude Haiku 5.5 ⚡ — Released October 7, completing the Claude 5.5 family. $0.10/$0.50 per million tokens — roughly 75% cheaper than Haiku 4.5. Context expands from 200K to 1M tokens. Adds effort controls and computer use (OSWorld 2.1: 72.4%). Anthropic also halved Sonnet 5.5 cache-read costs in the same release.
Claude Now Works Inside Google Docs, Sheets and Slides 🔗 — Public beta October 6, all paid plans. One Marketplace add-on opens a Claude sidebar in whichever file you have open: in-place edits in Docs, formulas and pivot tables in Sheets, new slides in Slides. Separate connectors let you create and edit Google files from inside Claude directly.
SAP Acquires TechWolf to Map Every Job in Every Enterprise 🌍 — October 6. Ghent, Belgium-based TechWolf maps skills and tasks across enterprise roles using evidence from actual work systems rather than job descriptions. It joins SAP SuccessFactors and keeps serving non-SAP customers from Ghent.
Isomorphic Labs in Early Talks to Raise at Least $40B 🧬 — Bloomberg reported October 8 that the Google DeepMind drug-discovery spinout is raising at a valuation of at least $40B, with some sources citing $50B. Round not yet closed. Best known for AlphaFold 3.
Biren Technology Raises $515M 🌍 — Shanghai-based Biren placed 130M new H-shares on October 8, its third major raise in 2026 after its Hong Kong IPO and a $892M placement in July. Builds AI GPU chips as a Nvidia alternative in the Chinese market.
Perplexity Open-Sources pplx-embed-v2 ⚡ — MIT license, October 7. Two ColBERT models (0.6B edge + 9B retrieval) sharing one embedding space: index at 9B quality, query at 0.6B cost. The 9B scores 92.4% on MADQA. Handles text, images, and rendered PDFs.
Useful tools ⚒️
⭐ FastRouter.ai — Unified LLM gateway routing requests across 200+ models via one OpenAI-compatible API. Set rules by cost, latency, or quality; FastRouter picks the model. Includes failover, BYOK, RBAC, prompt management, guardrails, observability, and MCP gateway support.
Rill Browser — Browser built around Claude Code and Codex. Opens your project, passes context to your existing agent, and shows all tasks across sessions in one place. Organizes tabs by theme and turns browsing history into a searchable timeline. Uses your existing Claude or Codex subscription.
Opengeni — Deploy AI agents to production in hours: sandboxes, credential handling, persistent memory, human approval flows, 100+ integrations, and full cost visibility. Apache-2.0, self-hostable or cloud. For teams who want to ship agents without building agent infrastructure first.
Spira Maxima — Give it a script or recording and get back a finished social video: presenter, motion graphics, B-roll, captions, music. The model composes each element rather than applying a fixed template. $0.15/second at launch (50% intro discount).
Fuse AI — Assemble a custom GTM stack from pre-built AI modules: lead enrichment, email sequences, CRM sync, intent signals, and reporting. For founders who want a full GTM system without buying and integrating five separate tools.
Share this post with friends, especially those interested in AI!
Weekly Guides 📕
How I Stopped Overpaying for AI Models — Our guide this week: a 15-minute monthly workflow for auditing your model spend, routing cheaper models to the right tasks, and using OpenRouter to compare costs across providers. Published October 8, directly tied to the Haiku 5.5 and Mistral Large 4 launches. If you are still running Haiku 4.5 or paying frontier prices for classification tasks, this is where to start.
Claude Haiku 5.5 Migration Guide — Official Anthropic docs: swap the model ID, handle breaking changes (computer use tool name changed, server-side fallback removed), adjust for 30% more tokens on the same text, and configure the new effort parameter. Per-platform table covering Anthropic API, Bedrock, Google Cloud, and Microsoft Foundry.
How to Use the Mistral Large 4 API — Get an API key from Mistral Studio, make your first call with curl/Python/TypeScript using model ID mistral-large-4, configure reasoning effort, send images, use function calling and structured JSON output. Pricing: $0.68/$2.09 per million tokens.
AI Meme of the Week 🤡
AI Tweet of the Week 🐦
Bonus Materials 🎁
A $20 AI Agent De-Anonymized 1.52 Million Georgia Ballots — Princeton researcher Max Springer pointed a commercial AI agent at a paper describing a known flaw in Georgia’s ballot scanners. It produced a working exploit in hours, reconstructing the cast order of 1.52 million May 2026 primary ballots from public records alone. Georgia issued emergency rulemaking. The secret ballot has existed since 1888. It took $20 and an afternoon to break it.
DOOM Now Runs Entirely Inside a SQL Database — Developer mariuz ported DOOM to Firebird SQL compiled to WebAssembly. Every frame — BSP traversal, physics, rendering — executes as a SQL query. JavaScript handles only keyboard input and canvas. Playable in your browser. The same developer followed up with Quake, Quake 2, and Quake III Arena in the same database.
AI Is Decoding Sperm Whale Language — Project CETI released new findings from years of recording sperm whale codas with ML: vowel-like features in clicks, regional dialects, and what may be a phonetic alphabet. Researchers from Harvard, MIT, Imperial College and UC Berkeley. Whether any of this is language in the way humans mean the word is a question the project is not ready to answer.
Your take: OpenAI just made ChatGPT interactive for 1.2 billion people. Mistral put a trillion-parameter European model into public preview. And Manus turned a forced government breakup into a $4B independent company. Which of these changes the most about how AI works in the next 12 months? Drop it in the comments 👇
If you missed our previous updates, don’t worry, here they are: AMD Buys World Labs, OpenAI Ships Dots, Meta Goes Enterprise | Weekly Digest