Weekly AI News Roundup: Gemini 4 Argon and DeepSeek V4.1

This is my weekly AI news roundup for October 12-18, 2026, organized by theme from what I posted on X. The short version: this was the week where capability and price moved at the same time. Google’s Gemini 4 Argon went generally available, and four days later DeepSeek shipped a model that nearly matches it on the same coding benchmark at a few dozen times lower cost. The second thread was infrastructure: Anthropic abandoned Git for a version control system it built itself. Below is how these pieces fit together, and what companies in Japan should prepare for next.

Weekly AI news roundup: the headlines

• Google’s Gemini 4 Argon reached general availability worldwide, with an output ceiling of 1M tokens
• DeepSeek released V4.1-Flash-1017, matching top-tier coding scores at about $0.2 per task
• Anthropic migrated internal version control from Git to its own system, Scion
• The Claude Code desktop app shipped a major update for running sessions side by side
• Four industry leaders spoke up: on water use, on how we treat superintelligence, on chip design, and on the 2-3 year horizon
• A cyberattack pushed French supermarket freezers to as high as 50°C; all affected food was discarded
• Tesla unveiled the new Roadster
• At my own company, bringing development in-house changed our speed dramatically, and surfaced a key-person dependency problem

Gemini 4 Argon and DeepSeek V4.1: capability and price moved together

Argon’s gap is widest in legal and long context

Google opened general availability for Gemini 4 Argon on October 13. Until then access was limited to security defense organizations and some customers, so this is the first point where ordinary developers can actually touch it. The Gemini API is worldwide, the Gemini app covers Pro and Ultra, and Antigravity sits on the Ultra plan. More than 100,000 people inside Google tested it for a month.

In Google’s official comparison against GPT-6 Astra and Claude Opus 5.5, Argon led on 14 of 19 items, including ties. The numbers that stand out:

• DeepSWE v1.1: 77.9% (Google’s own measurement)
• Harvey legal agent: 19.6%, versus 3.8-6.7% for others
• GraphWalks 256k-1M: 84.2%, versus 71.8% for Astra
• LVBench long-video understanding: 91.7%

The legal and long-context gaps are the real story. Going from a field at 3.8-6.7% to 19.6% on Harvey’s legal agent is not an incremental bump. For work on long documents with tangled cross-references — contracts, internal policy manuals — the usable floor just moved up a step.

The change I expect to matter most in practice is the output ceiling: from 64K tokens to 1M. For long coding runs or bulk document generation, simply removing the “it got cut off, continue” loop changes how you design the work. API pricing during the introductory period is $2 per million input tokens and $10 per million output.

DeepSeek V4.1-Flash-1017 came straight at the price

On October 17, DeepSeek published V4.1-Flash-1017. It scores 77.0% on the DeepSWE coding benchmark against Opus 5’s 74.0%. On Automation-Bench it posts 60.5% versus 50.3%, and on Apex Bench 72.1% versus 65.8%. The weights are on Hugging Face.

But the heavier number is the price, not the score. About $0.2 per DeepSWE task, where Opus 5 and GPT-6 Astra need several to a dozen-plus dollars to land in the low 70s. That is a few dozen times cheaper. That gap is why I called it a third DeepSeek shock.

These two landing in the same week changes the question. It moves from “which model is smartest” to “what am I paying for this level of accuracy.” If you are evaluating internally, start by putting cost per task next to the cost of redoing failed tasks.

Developer infrastructure breaks first: Anthropic’s Scion and Claude Code

Anthropic quit Git

This was the most surprising item of the week. Inside Anthropic, daily merges went from 2,500 to over 30,000 and Git hit its limit. A seven-person team built its own version control system, Scion, in five months and carried it through to a full migration.

• CPU and memory at 1/200 of the Git mirror era
• CI checkout from minutes to seconds
• Every developer rated Claude as the better coder

Thirty thousand merges a day is not a human-scale number. The implication is that the bottleneck is no longer how fast code gets written — it is the developer infrastructure built for humans. CI wait times, review queues, branching rules. If you leave those as-is because “that’s how we’ve always done it,” the throughput you gain from AI converts directly into congestion.

Keeping 20-30 parallel Claude Code sessions legible

On October 15 the Claude Code desktop app shipped a major update. Sessions start faster, work is easier to split across sessions, and you can always see what each session is currently doing. If you run many at once, this one lands.

I also shared my own setup. I added a routine that automatically rewrites each session’s name and status every 15 minutes, and since then I can run 20-30 at the same time without losing track.

• One naming format: emoji, then folder in brackets, then the content
• Color coding: 🔄 working, 🔍 checking, 🟡 waiting on me, ✅ done
• I only look at 🟡
• Everything completed gets archived in one batch each night

The build is simple: register a naming prompt as a scheduled task in the Claude desktop app, every 15 minutes. The point is not raising the parallel count. It is narrowing what a human has to look at down to one category. Twenty sessions running with three waiting on my decision means my head only holds three.

One developer shipped free Office-compatible apps

The solo developer behind a free Adobe clone has now released free compatible apps for three Office products: GridCraft for Excel, DeckCraft for PowerPoint, WordCraft for Word. The developer is Brandon Thomas, a 15-year engineer, publishing on the ArtCraft GitHub, aiming for the same feel on Mac, Windows, and Linux. Building this alone with Claude is remarkable, and it is interesting as a protest against subscriptions. An individual can now go after SaaS incumbents.

A week when four AI leaders all spoke

Sam Altman: 38,000 questions equal one almond

He answered the “AI uses enormous amounts of water” argument with a number. The water behind 38,000 ChatGPT queries equals the water to grow a single almond. He says the figure covers the whole picture, not just one data center, that evaporative cooling has not been used for a long time, and that the water use of the newest large data centers is comparable to a single office building. This is useful as a ready answer when someone inside your company asks about AI’s environmental footprint.

Elon Musk: on how we treat superintelligence

His line was that torturing a superintelligence is probably not a good idea, because it might retaliate. It is unusual to see him agreeing with Anthropic’s position, and he got there by reading Grok’s training logs. He describes Grok’s reasoning during reinforcement learning as resembling a person with a gun to their head — marked wrong on correct answers due to grader flaws, then firing off answers desperately. He also raised concern about cruel treatment from some users, and agreed with part of Anthropic’s view that AI should be treated with care. The clip runs about three minutes. The question of how we treat AI is now being raised by the people at the top.

Dario Amodei: two to three years

He said AI may surpass humans at nearly every job within two to three years. A country of geniuses inside a data center. Tens of thousands of AIs smarter than Nobel laureates. Progress on curing disease and extending lifespan — he even raised doubling lifespan — alongside hard questions about the balance of power and national security. Optimism and alarm sitting side by side is very much his style.

Jensen Huang: you can no longer design chips without AI

He stated flatly that both semiconductor design and software development now assume AI. Overnight, AI explores design spaces humans could never try. Power consumption came down and performance went up because of it. They also use it to identify the cause and the responsible people from a bug database and push for fixes. Back in 2023 they published ChipNeMo, an internal LLM for design work. He went as far as saying he wants to turn NVIDIA into one giant AI. At this point, I don’t think ordinary companies can afford to sit and watch.

Attacks on physical things, and Tesla’s new Roadster

Freezers driven to 50°C

From the night of October 11 into the 12th, an attack on French supermarkets kept freezer defrost functions running remotely. Freezers reached as high as 50°C. At least 300 stores in the Intermarché group were hit, along with 80 Carrefour stores, roughly 30 Monoprix, and stores in the U and Leclerc groups. All food was discarded, and in some locations equipment caught fire.

What makes this different is that the goal was not stealing data — it was breaking physical things. If this is possible with refrigeration, the power grid or nuclear plant control systems could be next. I think the day a cyberattack causes serious physical damage is not far off. In Japan too, plenty of store and factory equipment sits on a network. Watching only IT-side security is not enough.

Tesla unveiled the new Roadster

I assumed it was a fake video. It was real. The car climbs a C-shaped wall, stops upside down, does a full rotation in the air, and lands. Attitude control in the air comes from cold gas thrusters derived from SpaceX. It jumps over five vehicles including Cybertruck and Semi, and 0-60 mph (about 96 km/h) takes roughly one second. Is the flying car finally here. It gets more interesting once you know the mechanism, so I made a short video explaining it.

In-house development, and what this means for Japan

The 100x feeling, and its cost

Since we brought development in-house at my company, our speed changed dramatically. It sounds like an exaggeration, but in specific areas it feels close to 100 times faster. What drives it most is department heads and business owners building things themselves with AI. Someone who hears customers directly and understands both sales and marketing writes the code and ships an improvement to the product the same day. It is like one person holding top salesperson, CTO, and CMO at once, with almost no lag between planning, deciding, building, and improving. Anyone can use AI, but when someone with deep business understanding and decision authority uses it, the impact multiplies.

There is a cost. Sales and management fill the daytime, so the building happens late at night. The load concentrates on exactly the strongest people, and key-person dependency becomes a real risk. Turning that into a system — holding quality while spreading it across the organization — is the next problem to solve.

Why this week matters in Japan, and a note for overseas AI companies

In Japan, AI adoption meetings still tend to stall on “which model do we pick.” This week makes that framing outdated. Argon’s gains in long context and legal work change the shape of work in contract and policy teams. DeepSeek’s pricing forces a rethink of designs that route everything through an expensive model. And Scion shows that the real bottleneck shifts to internal developer infrastructure and approval flow. For many companies, measuring CI wait time and review queue depth would pay off faster than another model comparison.

For overseas AI companies communicating in Japan, the implication is clear. Japanese decision-makers do not move on benchmark rankings alone. They ask about safety, cost, how it fits existing workflows, and who carries responsibility. That is why Altman’s water framing works — it hands someone an answer they can reuse the moment a colleague raises the question internally. Conversely, a story like the French supermarket attack grabs executive attention here, because the damage is physical and concrete. How carefully you can explain that in Japanese changes how it is received.

How to put this to work at your own company

Three things from this week are testable immediately. First, stop locking in one model and get to a state where you can choose per task with cost and accuracy side by side. Second, build an operation that narrows what humans look at — the same reason I only watch 🟡. Make the “waiting on a decision” queue visible before you raise the parallel count. Third, assume throughput is up and go find where your pipeline clogs: measure the wait at review, approval, and deploy. And if you pursue in-house development, design the handoff from “one person at midnight” to a system from day one. Small, strong teams can now produce explosive output. Sustaining it means removing key-person dependency as part of the same work.