Google's most powerful model yet, and you can't touch it
Issue #049 · 4 min read
Google's most powerful model yet, and you can't touch it
Gemini 4 Argon ships, then gets locked behind a vetting process. Plus GPT-6.1 Sol undercuts its own flagship.
The big story
Google announced Gemini 4 Argon today and called it the most capable model it has ever built.
Then it did something stranger.
The company says the model is good enough at cybersecurity work that it's only handing it to "trusted cyber defenders" for now — not the general public, not most developers, according to The Verge.
Think about what that sentence actually requires. Google has to believe the model is strong enough at finding and exploiting security holes that wide release is a real risk. And it has to say so, in public, as a selling point.
Ars Technica's headline makes the gap plain: Google announced it, but you can't use it yet. TechCrunch frames it as a workhorse for coding and cybersecurity — which is true for the people who get access.
For everyone else, Gemini 3.5 Pro is still what's on the menu.
This isn't the first time this quarter a lab has drawn a line between "the model exists" and "the model is safe to hand out." It's becoming routine. If you're building a roadmap that assumes frontier model access on day one, stop assuming that. The gating is now a product decision, not a technical afterthought, and your vendor evaluation needs a line item for "who actually gets this, and when."
What shipped
OpenAI introduced GPT-6.1 Sol, pitched as near-Astra intelligence for coding, computer use, and professional work at one-fifth of Astra's standard API price, per OpenAI's announcement.
If that pricing holds up in practice, it changes the math on any feature you'd previously ruled out because the top-tier model was too expensive to run at volume. Worth a real cost test before you assume Astra is still the only option for quality work.
GitHub added opt-in dist-tag permissions for npm trusted publishing, letting short-lived OIDC credentials promote a package version to latest or move next and beta pointers, instead of relying on long-lived tokens — see the changelog entry. Narrow, but it closes a real credential-leak path for any team publishing packages, and it's the kind of security improvement nobody notices until the breach that didn't happen.
GitHub also opened self-serve trials of GitHub Advanced Security to GitHub Team customers, not just Enterprise, so you can test Code Security and Secret Protection from your org's billing page without a sales call — details here. If you've been told Advanced Security is an Enterprise-only conversation, that's no longer accurate, and it's worth five minutes to check what your plan actually includes now.
What I'd actually do this week
- If your org is on any Gemini waitlist or early-access program, ask your Google rep directly what "trusted cyber defender" vetting looks like and whether your team qualifies — don't assume you're in just because you're a customer.
- Run a real cost comparison between GPT-6.1 Sol and whatever flagship model you're currently budgeting for coding or computer-use tasks. A fifth of the price is the kind of number that should move a roadmap decision, not just a line in a changelog.
- If you publish npm packages, check whether your CI pipeline still uses long-lived tokens for tag promotion. The new OIDC option is opt-in, meaning nothing changes unless you turn it on.
Reply and tell me if you've actually gotten Gemini 4 Argon access — I want to know who's in the trusted circle.