Google announced Gemini 4 Argon on Wednesday, its new top AI model, and then told almost everyone they can’t have it yet. For now it’s going only to a group of trusted cyber defenders in Google’s Fairwind Program. Google says it’s taking part in the US government’s voluntary process for pre-release model access while it slowly opens things up, and it gave no date for a public release. When that happens, paid API customers and Google AI Ultra subscribers go first. Reuters says Argon is bigger than Google’s past Pro models, and that Google has dropped Gemini 3.5 Pro, which Sundar Pichai had promised for June.

Google’s headline number is 77.9% on DeepSWE v1.1, a test of long software engineering tasks, which Ars Technica notes is higher than rivals like OpenAI’s GPT-6 Astra and Anthropic’s Opus 5.5. Reuters points out that Argon still trailed on two of the four coding benchmarks Google included. The output limit jumps to 1 million tokens, up from 64,000 in earlier Gemini models, so it can write far more in one go. Pricing starts at an introductory $2 per million input tokens and $10 per million output tokens, and goes up to $4 and $20 once that period ends. Inside Google, teams of Argon agents freed up more than 300 TiB of memory across its data centers, and they’re moving C and C++ code to Rust, including more than 800,000 lines in the Fuchsia Zircon kernel.

Cyber defense is the main pitch for now. Google trained Argon to find and patch software flaws on its own, and the trusted defenders get a version without the usual cyber guardrails. Wiz has been using it through Scan for Good, its free program for protecting public infrastructure. Google says the model found a critical flaw in healthcare software used by hospitals around the world that exposed sensitive personal information, and that earlier frontier models missed it. Ars notes Google didn’t share specifics. Google also says it watches Argon’s chain of thought and actions and can stop it if it goes past what the user wanted, and it urged other labs to keep model reasoning visible too.

I think letting defenders go first makes sense, because a model that can find holes in hospital software is something I’d want pointed at that software before attackers get a turn. My problem is how little of this I can check right now. The Wiz find comes with no details, and the benchmarks are Google’s own. Even then, Reuters notes Argon trails on two of the four coding tests Google picked. The 1 million token output is the part I’m most curious about. I’d also like Google to say what the government review actually looked at before the wider release.

OpenAI said Wednesday that it found and shut down a coordinated campaign to pull hidden reasoning out of its models, and it tied a core group of the people involved to Moonshot AI, the Chinese company behind Kimi. OpenAI calls this adversarial distillation, which means using one model’s outputs or reasoning to train or improve another model. The hidden reasoning is the model’s internal record of how it works through a task. According to OpenAI, getting at it can reveal things left out of the final answer and help someone else copy what the model can do.

The activity started on July 1 at low volume. On July 24 and 25, OpenAI saw 16,000 requests using an extraction pattern from more than 4,000 users, and it later found related activity across more than 15,000 users. It says it fully shut the campaign down by July 28, and that those figures count attempts, which weren’t necessarily successful. The trick was to copy encrypted reasoning out of one conversation and ask the model in a different conversation to decrypt it and write it out. OpenAI says this wasn’t a breach, and nobody got direct access to stored user conversations. It has since closed the gap that let someone replay another user’s encrypted reasoning. It also banned or restricted the accounts involved and tightened its signup controls.

The attribution is the shakier part. OpenAI said it isn’t clear whether all the operators came from one actor. CyberScoop noted that the blog post offers no technical evidence for pinning the core cluster on Moonshot, and OpenAI told the outlet it wasn’t sharing more “for security reasons.” Moonshot didn’t immediately respond to CNBC. OpenAI says it shared its findings with other labs through the Frontier Model Forum and with government channels. CNBC points out this comes weeks after Anthropic accused Chinese developers, including Moonshot and Alibaba, of secretly using Claude to train their own models. It also lands the same week Moonshot said it was reviewing its Kimi models over a separate jailbreak report.

I believe OpenAI saw the traffic it describes, because the dates and request counts are specific and it says plainly that some of it may have failed. The Moonshot part is where I get stuck. Publicly naming a company is a big step, and CyberScoop has a point that the post gives readers nothing to weigh. I understand not wanting to tip off the next group by explaining how you caught this one, but then the claim rests on trusting OpenAI. I’m also surprised the decrypt-it-in-another-chat trick worked at all, given how simple it sounds.

Google Japan shows off a deliberately odd Gboard keyboard every October 1, and this is the 15th in the series. Android Authority says the date is a nod to the 101-key keyboard, since 10/1 can be read as “101.” Last year’s version swapped keys for dials. This year Google put the keys on conveyor belts.

To type, you wait for the key you want to roll past your finger and press it as it goes by. Each belt carries 29 keys and handles uppercase and lowercase letters. Google calls it a “Copernican revolution” in text input and says the point is to make one-handed and one-finger typing more efficient. Its Japanese blog post adds that you can type with one hand while drinking tea or using a mouse, and that a second unit would let you type with both hands.

You can’t buy one. The Gboard team has published the design as open source on GitHub, so you can build it yourself if you have the parts and some patience. The Verge’s Jess Weatherbed summed it up by saying you’ll learn patience waiting for your keys to come around.

I enjoy this yearly bit from Google Japan because the joke has a real question inside it. Making the keyboard move to one finger actually would help someone typing with one hand, right up until the letter they need has just gone past and they have to wait for it to loop back. I can’t picture typing a whole email on it. Publishing the design files is what I like most, since it turns a funny video into something people can build. I’d love to see how fast someone can get on it after a week of practice.

Silicon Spritz is a clear daily breakdown of tech and AI news for founders, builders, creators and curious minds, landing in your inbox Monday to Friday. The Long Pour is our occasional deep dive on one company. Subscribe to Silicon Spritz to get the next edition.