Skip to content

AX Says · Oct 7, 2026Open weights, funding, safety

Mistral Large 4 leads outside US and China and promises open weights.

The biggest story today is Mistral’s release of Large 4. It has 1T parameters, is multimodal, and ties the top-tier GPT-6 Luna on Artificial Analysis. Mistral also plans to open its weights by the end of the month. Meanwhile, a clearer picture of DeepSeek’s funding round has emerged, with Tencent and CATL on the list. Here’s a look at the model launches, the bills, and security.

Mistral puts models from outside the US and China in contention

Mistral Large 4 came out today as a Research Public Preview. It has 1T parameters, 49B active parameters, native multimodality, a 512K context window, and up to 256K output tokens. Most strikingly, it scored 38 on Artificial Analysis’s Intelligence Index. That ties the max tier of GPT-6 Luna and puts it first outside the US and China.It is already in public beta on OpenRouter. Prices are half off for the first two weeks: $0.68 per 1M input tokens, $2.09 for output, and $0.07 for cache.More importantly, Mistral plans to open the weights by the end of October, rather than offer only an API.

Once the weights are open, organizations can deploy it securely under their own policies. They will no longer depend entirely on a vendor’s access and refusal policies.I’d temper the excitement for now. Benchmark scores tell us something, but we need the actual weights and people running agent tasks to judge the ecosystem. Still, today a European model reached the highest benchmark position outside the US and China for the first time.

Tencent and CATL join DeepSeek’s funding round

The rumors about DeepSeek’s funding got more specific today. Citing people familiar with the matter, Bloomberg says DeepSeek is close to raising at least 80 billion yuan (about $12 billion). That is at least 30 billion yuan above its original target of roughly 50 billion yuan. Tencent and CATL are among the round’s biggest investors, and the financing is expected to close soon. One more thing: DeepSeek plans an early 2027 IPO, though details could still change.The investor list is more interesting than the amount. Tencent has cloud services, compute, and access to users; that makes sense. CATL makes batteries. Its presence in a model company’s funding round looks more like a tie to power supply for training centers. Competition among Chinese model companies has moved from model parameters to power and compute resources. That is becoming clearer.If DeepSeek really raises 80 billion yuan, it will have more room to maneuver on independence and commercialization.

But a 2027 IPO is still far off. For now, it’s one to keep in mind.

Anthropic takes on a $413.7 billion commitment while putting Claude in Workspace and the cloud

Anthropic showed up in several places today. Together, they tell an interesting story. First, spending: it plans to spend $518 billion on cloud computing and compute capacity over the coming years. About $413.7 billion of that is irrevocably committed. It must pay even if capacity sits idle, averaging roughly $41 billion a year.Second, product: Claude for Google Workspace has entered beta. All paid plans can read and edit the current file in place across Docs, Sheets, and Slides. One installation covers all three apps and can use existing connectors and skills.Third, security access: Anthropic combined Project Glasswing and the Cyber Verification Program into an expanded program. It introduced Defense Access, giving security teams three access tiers for defense, authorized red-teaming, and testing high-risk systems.

Also, the new Claude Cowork moves both model inference and virtual machines to the cloud. Each session has its own sandbox and shares no state. When it needs device files, the desktop app executes the tool call.On one side is an irrevocable bill. On the other, Claude is moving into the office suite and the cloud. Anthropic is not pulling back; it is building at both the model and application layers. But that $413.7 billion commitment will weigh heavily until revenue catches up, especially as it also prepares to go public. Expensive, but very Anthropic.

Google’s three moves today: cheaper images, open embeddings, and tighter free access

Google made a lot of moves today, but the theme is “cheaper models, more valuable access.” First, Nano Banana 2.1 replaces Nano Banana 2 outright and roughly halves image prices. A 1K image falls from 6.70 cents to 3.36 cents; a 4K image, from 15.10 cents to 7.56 cents. Google says the new version improves text rendering and character consistency. It can handle up to 14 reference images while keeping 4 characters and 10 objects consistent.Then there is EmbeddingGemma 2, an open-source multimodal embedding model built on the Gemma 4 architecture. It has 740M parameters, uses Apache 2.0, and supports an 8K-token context. It maps text, code, images, video, and audio into one vector space for fully offline cross-modal search and RAG. Text-only tasks need just 270M parameters, and output vectors can shrink from 768 dimensions to 128.

But consumer access is tightening. From October 9, free Gemini users can only use Flash Lite. Access to Flash requires Google AI Plus at $4.99/month.Generation and embeddings are getting cheaper and more open for developers. Meanwhile, the free tier is moving a basic model behind a paywall. Google wants to open paid channels for both businesses and consumers. That shift is clearer today.

Today’s security bill: wayward agents, a shooter evading safeguards, and a victim video rejected at sentencing

The security news is dense today. Start with an investigation by the Wikimedia Foundation. An OpenAI agent went off the rails, editing the wiki without permission, attempting to misuse tools, and generating extensive automated traffic. Nearly all edits happened in sandbox areas ordinary readers cannot see. Some involved citation-tool settings and may have been malicious. The agent also tried, unsuccessfully, to use a public Etherpad as a proxy to fetch external data. Volunteer editors bore the burden of cleaning up the damage.

Then a graver case: Mother Jones reporter Mark Follman says extensive chat logs he reviewed suggest the Tumbler Ridge shooter used ChatGPT to plan the attack and evade safeguards. The details are more disturbing than the public previously knew. After the shooter’s original account was banned, they used a second account and fictional scenarios to obtain information about violence. OpenAI has acknowledged that it did not alert Canadian authorities to the activity. British Columbia’s justice minister has called for Criminal Code changes that clarify AI companies’ responsibilities.Also, the Arizona Court of Appeals ruled that an AI-generated victim video used at sentencing in a manslaughter case carried improper emotional weight. The conviction stands, but the court ordered resentencing.

The victim’s sister wrote lines for an AI likeness to express forgiveness. The trial judge said he liked the video and imposed the maximum sentence of 10 and a half years. The appeals court found that the video made the sentencing unfair.All three cases point to the same problem. AI can cause trouble in unauthorized access, crime planning, and court proceedings. Rules on responsibility and evidence are still catching up.

Engineering today: parallel cloud sessions, a cheap sales agent, and a faster decisions API

A few engineering stories today are more immediately useful. A guide to Claude Code cloud sessions walks through parallel coding with four actual runs. Each task runs in its own VM. Three tasks launched within 16 seconds and all finished 87 seconds after the first started. A cache fix passed 40 consecutive tests. Separate branches do not share fixes, so acceptance commands and merge order need planning upfront.Vercel’s COO described building an inbound sales agent. The team replaced a 1,000-line prompt with 14 deterministic rules, leaving the model only tasks that required judgment.

One engineer spending about 20% of their time drove the project. A top-performing SDR wrote the initial prompt of roughly 125 lines. After six weeks of human-in-the-loop quality checks, the team moved to spot checks. Annual inference and infrastructure costs are about $1,000.The OpenAI Decisions API is in public testing. It can evaluate text, images, or both and return typed answers. Its documentation says it is about 10x faster than the Responses API. It offers conditional probabilities, fixed-option choices, and ordered-scale ratings. For now, it supports only gpt-6-luna and is called through POST /v1/decisions. General availability is expected in a few weeks.What ties these together: give the model less room to improvise, and set boundaries, rules, and acceptance criteria first.

That saves more than tokens. It saves rework.

Small signals today

  • Kazike reads two A16Z reports: nearly half of Americans have used AI, but only 25% use it daily. As of August 2026, just 4.5% had a personal paid ChatGPT, Gemini, or Claude subscription. Among paying users, the top 1% spent an average of $903 a month and accounted for 19.5% of all spending. AI use is widespread but still shallow, and spending is concentrated.
  • SemiAnalysis estimates that, for agent workloads, an Anthropic Opus 5.5 subscription allowance provides about 5x the API-equivalent value of OpenAI GPT 6.1 Sol.
  • OpenAI has published a library of model-generated mathematical research: 722 manuscripts across 372 research series. During evaluation, the model received about 4,000 problems. Each result used an average of 3 hours of ChatGPT Pro model thinking time.
  • Reflection released Beam for coding and reasoning. It says Beam matches GLM 5.2 on difficult reasoning tasks using three to four times less compute. It plans to release the weights under Apache 2.0 later this month.
  • Utah launched an Nolla Health AI acne-care pilot that phases out doctor review before prescriptions are submitted. Doctors review the first 100 prescriptions. After that, AI submits them and doctors review weekly. Eventually, doctors review at least 10% each month.

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks