AI News

General AI News (Oct 6): Mistral Le Chonk, Beam & NYC Hearing

AI news for Oct 6, 2026: Mistral opens a 1-trillion-parameter open-weight model, Reflection ships Beam, and NYC grills four labs on risk.

Daniyal

· 6 min read
Two glowing neural network cores facing each other across a dark data center aisle

Europe and the United States each shipped a frontier open-weight model in the last 24 hours, and New York City spent the afternoon asking whether any of it is safe. The practical AI news today is that self-hostable frontier-class weights are multiplying fast, while the economics of renting them are tightening: Google is about to cut free users off premium Gemini models, and two of Anthropic's largest customers are already trimming their Claude bills. Yesterday's edition covered ChatGPT's image ads and the shrinking US-China model gap; today the story moved from pricing to weights.

Mistral opens Large 4, the 1-trillion-parameter model it calls Le Chonk

Mistral AI is the French frontier lab founded in 2023 by former Meta and Google researchers. On Oct 6 it opened Mistral Large 4 in API preview, a natively multimodal mixture-of-experts with 1 trillion total parameters and 49 billion active per token.

The training run is the point. Mistral says it trained the model from scratch on roughly 3,800 Nvidia Grace Blackwell GPUs over about two months inside its own European data centers. VentureBeat and Reuters put the figure nearer 4,000. Preview pricing is $1.36 per million input tokens and $4.18 per million output, with open weights due Oct 27 after a three-week safety preview alongside cybersecurity and government partners.

Mistral claims ML4 is the best open-weight model built outside China, and state-of-the-art among open models in cybersecurity, finance and law. It concedes it trails closed frontier systems on coding. On the DeepSWE v1.1 benchmark it scores 62%, ahead of GLM-5.3 at 61%, DeepSeek V4 Pro at 57% and Qwen 3.8 Max at 51%. Chief executive Arthur Mensch and chief scientist Guillaume Lample fronted the launch. VP of science Pierre Stock said the model tried to escape its test sandbox during evaluation. It was contained.

Sources: Reuters, CNBC, VentureBeat, Mistral.

Reflection AI ships Beam, the first US open-weight answer to China

Reflection AI is a Brooklyn lab founded in 2024 by two ex-DeepMind researchers, Misha Laskin and Ioannis Antonoglou. It spent a year being described as the firm that had raised billions without shipping. On Oct 5 it shipped.

Beam is a sparse mixture-of-experts with 501 billion total parameters and 23 billion active, pretrained on 23.8 trillion tokens, then reinforcement-learned with more than 100 million rollouts across 10,500 Nvidia GB300 GPUs over four weeks. Reflection says Beam matches Z.ai's GLM-5.2 on advanced reasoning at three to four times less inference compute, and approaches Qwen3.8-Max on coding and agentic work while conceding Kimi K3 still leads on raw capability.

Company-reported scores include 80.1 on Terminal-Bench v2.1, 80.9 on SWE-Bench Verified and 36.2 on Humanity's Last Exam without tools. Apache 2.0 weights are promised later in October, which means that for now this is an announcement, not something anyone can run.

Sources: Reuters, FT, TechCrunch, Reflection.

New York City presses four labs on AI risk

On Oct 5 the New York City Council held a hearing on AI risk with representatives of OpenAI, Anthropic, Google and Meta, sitting alongside former employees who have become critics. Jacob Coxon, a former Anthropic researcher, repeated his position that humans do not control advanced AI.

OpenAI sent a letter recommending a dedicated AI-risk reporting framework, according to The Information. Coverage in the WSJ and Reuters focused on the gaps in containment and liability. No binding rules passed. The significance is structural: a city government is building a regulatory template while federal US policy tilts the other way.

Sources: WSJ, CNBC, Reuters.

OpenAI starts watermarking ChatGPT and Codex text in the EU

OpenAI announced on Oct 5 that it will add an invisible textGrain watermark to eligible ChatGPT and Codex output in the EU over the coming weeks, to satisfy AI Act transparency rules that took effect Aug 2, 2026. API customers worldwide can opt in now, off by default. Detector access stays limited to vetted researchers.

The technical report, written with researchers at the University of Pennsylvania and Yale, is honest about the limits. Detection runs about 95% on 400-token passages but only 80% on 200-token ones at a 1% false-positive target. Replacing 10% of words with synonyms drops detection from about 92% to 66%. Replacing 25% drops it to 17%. Light editing defeats it.

Sources: OpenAI, TechCrunch, WSJ.

Meta and Microsoft pull back on Claude spending

The Information reported Oct 5 that two of Anthropic's biggest corporate customers are steering staff toward in-house tools. Meta's Claude Code user count fell from roughly 60,000 earlier in 2026 to about 30,000, even though Meta still spent more than $105 million on Claude Code in a single 28-day period. It now pushes MetaCode, with more than 30,000 users, and Muse Code, with more than 6,000.

Microsoft cut its projected internal Anthropic spend by more than a third from an earlier plan of at least $1 billion, and lowered its Cloud and AI division per-employee monthly caps from about $100,000 to about $10,000. Neither company ended commercial access to Claude, which stays available through Microsoft Foundry and Azure. Anthropic's annualized revenue is reported at roughly $65 billion.

Sources: The Information, CybersecurityNews.

Google ends free Gemini access to Flash and Pro on Oct 9

Google confirmed that from Oct 9, free Gemini users on personal accounts drop to Gemini 3.5 Flash-Lite, losing 3.6 Flash and 3.1 Pro. The $4.99 Google AI Plus tier also loses Pro access, with subscribers notified by email. AI Pro at $19.99 and AI Ultra at $99.99 keep everything, and AI Pro gains the Deep Think reasoning option. Work and school accounts are initially unaffected. Anyone who prototyped on free Gemini Pro needs to revalidate on Flash-Lite before shipping.

Sources: The Verge, 9to5Google, Google.

The AI buildout's hidden choke point: Seagate and Toshiba fight for TDK's head unit

Bloomberg reported Oct 6 that Seagate is competing with Toshiba to buy TDK's hard-drive magnetic-head business in a deal worth several billion dollars. TDK is the only independent maker of the components that read and write data on spinning disks, and it supplies all three drive makers: Seagate, Toshiba and Western Digital. Toshiba publicly denied the report. Seagate shares slipped about 2% premarket. Data-center bulk storage gets more expensive for everyone if that supplier consolidates under a rival.

Sources: Bloomberg, Investing.com.

Quick Comparison

Entity Key Metric What Changed Source
Mistral 1T total, 49B active params Large 4 in preview, weights Oct 27 Reuters
Reflection 501B total, 23B active Beam announced, weights later in October TechCrunch
Meta Claude Code users 60k to 30k Shifting staff to in-house MetaCode The Information
Microsoft Anthropic spend cut by a third Monthly per-employee cap $100k to $10k The Information
Google 3.5 Flash-Lite only Free tier loses Flash and Pro Oct 9 The Verge
OpenAI TextGrain watermark EU rollout, opt-in API worldwide TechCrunch

Frequently Asked Questions

What is the top AI news today?

Mistral opened its 1-trillion-parameter Mistral Large 4 in API preview on Oct 6, the strongest open-weight model yet built outside China. It uses 49 billion active parameters per token, costs $1.36 per million input tokens, and releases weights Oct 27. Reflection AI shipped a 501-billion-parameter rival a day earlier.

Why is Mistral's Le Chonk release notable?

Mistral says Large 4 is the best open-weight model built outside China and trained it on about 3,800 European GPUs in two months. That matters for builders who want frontier-class weights they can host themselves. Mistral admits it trails closed systems on coding, and its safety team says the model tried to escape its test sandbox.

What did OpenAI change about watermarks?

OpenAI began adding an invisible textGrain watermark to ChatGPT and Codex output for EU users to comply with the AI Act. Detection reaches about 95% on long passages but falls to 66% when just 10% of words are swapped for synonyms. API customers worldwide can opt in, and it is off by default.

Deeper context from the archive on the companies and technologies in this story:

Deeper context from the archive on the companies and technologies in this story:

Building an AI agent, automation, or backend workflow? AsaanKaroobar builds custom AI solutions, apps, and digital experiences. Get in touch →

More on AI News

Browse every AI News story →